Entropy-based dynamic ensemble classication algorithm for imbalanced data stream with concept drift Authors: JiaMing Gong, MingGang Dong, Suja A Alex, Hugh Cowley, Unil Yun, Unil Yun, Unil Yun Published: 2024 ABSTRACT Online imbalanced learning is an emerging topic that combines the challenges of class imbalance and concept drift. However, current works account for issues of class imbalance and concept drift. And only few works have considered these issues simultaneously. To this end, this paper proposes an entropy-based dynamic ensemble classification algorithm (EDAC) to consider data streams with class imbalance and concept drift simultaneously. First, to address the problem of imbalanced learning in training data chunks arriving at different times, EDAC adopts an entropy-based balanced strategy. It divides the data chunks into multiple balanced sample pairs based on the differences in the information entropy between classes in the sample data chunk. Additionally, we propose a density-based sampling method to improve the accuracy of classifying minority class samples into high quality samples and common samples via the density of similar samples. In this manner high quality and common samples are randomly selected for training the classifier. Finally, to solve the issue of concept drift, EDAC designs and implements an ensemble classifier that uses a self-feedback strategy to determine the initial weight of the classifier by adjusting the weight of the sub-classifier according to the performance on the arrived data chunks. The experimental results demonstrate that EDAC outperforms five state-of-the-art algorithms considering four synthetic and one real-world data streams. Introduction Online imbalanced learning has drawn considerable interests from both academia and industry. It aims to tackle data streams with skewed class distributions and concept drift. These problems commonly exist in real-world applications, such as sewage disposal, traffic control, spam filters, and electronic commerce [1]. Since such data streams are massive, dynamic, real-time, and high speed [2], they can be read-only once by an application [3]. This introduces the issue of classifying imbalanced data with concept drift [4] to online learning [5]. However, most standard algorithms consider the problem of balanced class distributions. Therefore, when the data stream becomes complex and imbalanced classification becomes inaccurate as majority classes are favored by classifiers [6]. To improve the accuracy of classifieries for imbalanced data, the sampling technique is one of the most effective approaches. Generally, the sampling technique can be divided into oversampling and undersampling [7, 8]. In oversampling, SMOTE is the representative approach for imbalanced data, first proposed in Ref. [9]. SMOTE leverages minority attributes to generate new samples, thereby balancing the amount of majority and minority attributes in samples. Several approaches based on SMOTE, have been proposed, such as borderline-SMOTE [10], ADASYN [11], REA [12], MWMOTE [13] and AMDO [14]. Additionally, Li et al., have proposed an entropy differences based oversampling approach for imbalanced learning [15]. The work in [16] presents a dynamic feature group method for imbalanced data distributions. It extracts group characteristics of samples and re-balanced distributions via the importance sampling technique. The authors of [17] employed a selectively recursive approach algorithm. By selectively placing the previously received positive samples into the current data chunk, the algorithm can significantly improve the accuracy at the expense of sample space. However, considering real-time and high speed data streams, these SMOTE-based techniques are not suitable in that they are time-consuming as they increase the amount of minority samples. Recognizing the critical demand for real-time analysis of data chunks in dynamic streaming environments, Monika et al. introduced a novel hybrid algorithm [18] that fuses the strengths of FFSMOTE and the bee algorithm. This eliminates the need for periodic analyses and model rebuilds, allowing for instantaneous updates as new data arrives. In the undersampling technique, the same amount of minority and majority samples are selected. The advantage of undersampling is that it significantly decreases the computational time by removing the majority samples. However, using undersampling technique alone substantially deteriorates in the accuracy due to the low information content. To address these issues and enhance the performance, undersampling is generally combined with ensemble classifiers that leverage multiple sub-classifiers to make decisions. Ditzler et al. have proposed two ensemble approaches, Learn++.CDS and Learn++.NIE [19], which share similar frameworks. Initially, these models introduce a certain amount of data into a data chunk from the data stream. Then, the sub-classifier can be created for each arrived data chunk, and the arrival time is recorded as the creation time. In this manner, a new data chunk is generated from the previous by the sub-classifiers. Each sub-classifier can be assigned a weight that linearly increases with the with creation time for each data chunk. Due to the use of sophisticated multiple sub-classifiers and weights, this framework is capable of realizing high accuracy and efficiency, and is applied widely. RUEO [20] utilizes an ensemble learning strategy, which combines multiple base classifiers into a single, more robust classifier. This approach harnesses the strengths of individual classifiers while mitigating their weaknesses, particularly in the face of complex and dynamic data streams. Based on a similar framework, Wang et al., employed a time decay function to capture the imbalance rate dynamically and proposed two learning algorithms, UOB and OOB [21], with the aim of adjusting the sample distribution to guarantee sample balance. UOB outperforms benchmarks in terms of the recognition accuracy for minority samples, and OOB is robust against dynamic changes in the class imbalance. Subsequently, these were improved further to develop and implement WEOB1 and WEOB2 [22]. In [23], the integrated learning framework has been leveraged to address the problem of imbalanced data streams. It was demonstrated that ensemble learning can solve the problem of cost-sensitive learning and online learning simultaneously. The above works have combined sampling techniques and ensemble learning to solve the problem of imbalanced data stream. However, these approaches have limited performance on imbalanced data streams with issue of concept drift [24]. MicFoal [25] efficiently classifies multiclass imbalanced network traffic with concept drift by selectively requesting labels for uncertain samples based on a variable threshold and adaptively adjusting the label budget. It addresses the imbalance and drift issues through a sample weighting formula and modular design for flexibility. To solve the concept drift problem, Du H et al., have designed and implemented a DWM algorithm [26], which is a general framework based on the weighted summation algorithm of the base classifier. In DWM, the rule for classification on the base classifier is maintained in an unchanged state. Moreover, the weight of the base classifier is dynamically trained and modified on each arrival data accordingly. However, this method causes classifier redundancy when modifying the classifier weights alone. To avoid increasing the classifier redundancy, the authors of [27] applied a dynamic weighted majority for incremental learning algorithm (DWMIL). This method divides data chunks into multiple sub-sample sets by undersampling. Then, it creates a classifier for each data chunk and continuously adjusts its weight according to the performance of the classifier in the subsequent data chunk while eliminating weights below a certain value. Thereby DWMIL yields better performance in improving the accuracy of the classifier, as well as increasing the execution efficiency of the algorithm. Using mathematical methods, Ghazikhani et al. established an online perception model [28] to resolve the problem of imbalanced data stream and concept drift. In their work, the class imbalance problem is solved by adjusting the weighted error of the base classifier, and the issue of concept drift in data streams is processed by a recursive least squares error model, respectively. Furthermore, considering concept drift, under the use of parallel ELM-based ensemble classifier in [29], UELM-MapReduce has been proposed to dynamically adjust the weight of each base classifier by removing the classifiers with low accuracy. Similarly, Learn++.NSE [30] trains a new classifier for each received data chunk and combines these classifiers with a dynamic weighted classifier. This allows identifying changes in the distribution of basic data and implementing actions to adapt to concept drift [31]. LEPID [32] adopts a selective and prioritized approach in identifying and emphasizing the significance of concept drift prototypes and minority class prototypes by assigning them augmented weights. This strategy directs attention towards regions that most accurately encapsulate the current conceptual landscape, enhancing the algorithm’s responsiveness to changes in the data stream. In summary, considering concept drift in imbalanced data streams, these algorithms have focused significantly on the amount of samples by neglecting the importance of the information and quality of the samples. This can lead to low recognition accuracy for minority samples. The classifier weights are initialized to fixed values, which can-not enhance the classifier performance to cope with the issue of concept drift [33]. To address these limitations, we propose an entropy-based dynamic ensemble classification EDAC algorithm for imbalanced data streams with concept drift. Compared to previous works, the main difference of EDAC is that it takes into account the entropy of the samples, which indicates the information carried by the samples and the quality of the samples. The general process of EDAC is as follows. Initially, according to the amount of entropy carried by the entire data chunk, EDAC divides the data chunk into multi-pairs of balanced samples. Next, to enhance the accuracy of the classifier, EDAC selects high quality samples to participate in the training of the classifier. Additionally, the weight on the classifier is initialized based on the performance on each classifier. And the weight is dynamically updated based on the performance on the new data chunk. We conduct simulations to verify the performance of EDAC on four synthetic and one real-world data stream [34]. The experimental results show that EDAC outperforms the-state-of-the-art algorithms [35]. In summary, the contributions of the proposed EDAC can be highlighted as follows: To balance the imbalanced data stream, EDAC presents an entropy-based balanced strategy, because the entropy can reflect the critical information of samples. The entropy is calculated based on probability theory. EDAC divides imbalanced data chunks into multiple balanced sample pairs. Each sample pair is trained with one classification rule, and multiple classification rules form the classifier of the current data chunk. In this manner EDAC yields better performance in improving the accuracy of classification. To balance the imbalanced data stream, EDAC presents an entropy-based balanced strategy, because the entropy can reflect the critical information of samples. The entropy is calculated based on probability theory. EDAC divides imbalanced data chunks into multiple balanced sample pairs. Each sample pair is trained with one classification rule, and multiple classification rules form the classifier of the current data chunk. In this manner EDAC yields better performance in improving the accuracy of classification. To enhance the classification of minority samples, EDAC introduces a density-based sampling method. Specifically, the strategy categorizes minority instances into high quality and common instances in terms of the density of similar samples. During this process, the strategy selects the amount of minority instances with the highest quality instances. Additinally, the common samples are randomly selected within a slight amount. In this manner, the amount of majority and minority instances can balanced while using the technique of undersampling. To enhance the classification of minority samples, EDAC introduces a density-based sampling method. Specifically, the strategy categorizes minority instances into high quality and common instances in terms of the density of similar samples. During this process, the strategy selects the amount of minority instances with the highest quality instances. Additinally, the common samples are randomly selected within a slight amount. In this manner, the amount of majority and minority instances can balanced while using the technique of undersampling. To address the issue of concept drift, EDAC also adopts the idea of ensemble classification to develop an approach for dynamically setting the weight on each classifier. Initially, the base classifier is created based on current data chunk, so the initial weight of the base classifier can be calculated through the current data chunk. Then, the initial weight and base classifier can be calculated an updated by using the subsequent data chunks. When the weight is less than the threshold, it indicates that the concept on that classifier is not suitable for the new data chunk, and thus, the associated classifier must be removed. Each data chunk has a base classifier with a weight that fits into the concept of the forthcoming data chunk. To address the issue of concept drift, EDAC also adopts the idea of ensemble classification to develop an approach for dynamically setting the weight on each classifier. Initially, the base classifier is created based on current data chunk, so the initial weight of the base classifier can be calculated through the current data chunk. Then, the initial weight and base classifier can be calculated an updated by using the subsequent data chunks. When the weight is less than the threshold, it indicates that the concept on that classifier is not suitable for the new data chunk, and thus, the associated classifier must be removed. Each data chunk has a base classifier with a weight that fits into the concept of the forthcoming data chunk. To verify the effectiveness and performance of EDAC, a simulation was conducted, and the performance was compared with those of five state-of-the-art algorithms, REA, ADAB, MWMOTEB, DWMIL, and DWM. IN the simulation, four synthetic data streams: SEA, Moving Gaussian, Hyper Plane, Checkerboard, and one real-world data stream, Electricity, were used The experimental results demonstrate the proposed algorithm outperforms the other algorithms in terms of the AUC, G-mean, and recall. To verify the effectiveness and performance of EDAC, a simulation was conducted, and the performance was compared with those of five state-of-the-art algorithms, REA, ADAB, MWMOTEB, DWMIL, and DWM. IN the simulation, four synthetic data streams: SEA, Moving Gaussian, Hyper Plane, Checkerboard, and one real-world data stream, Electricity, were used The experimental results demonstrate the proposed algorithm outperforms the other algorithms in terms of the AUC, G-mean, and recall. The proposed algorithm: EDAC The overview of the EDAC algorithm In this subsection, we provide the overview of EDAC. The framework of the proposed EDAC algorithm is shown in Fig 1. We assume that the data stream arrives in a chunk by chunk manner, and data chunk 1 is the first to arrive. We calculate the sample entropy value of each arriving data chunk to get the information content Ei of the sample population. The data chunk is divided into T equal subsets, where the sample information content of the positive samples in the subset is equal to that of the negative samples. Then the quality of each minority class instances of the data chunk is calculated. Next, minority samples are divided into high quality samples and common samples based on the the quality of samples. When selecting minority class instances, EDAC selects all high quality instances, and the common instances are randomly selected to make the total number of selected instances equal to the number of minority class instances in the overall data chunk. In the Fig 1, the red line {Line1, Line2, …, Linei} represents the threshold θ. The instances below the threshold are the sample with low quality, and the instances above the threshold are the sample with high quality. This minority class sample above the threshold carries more information and can better represent the minority sample groups in the data block; therefore, we try to select the minority class sample points above the threshold. We not only select these sample points with high information content, which will cause overgeneralization; but also randomly select the minority class sample points below the horizontal line to make the minority class sample more reasonable. Framework of EDAC. EDAC selects the same number of majority samples randomly as the number of minority samples to make the subset in equivalent to the original set. Then, according to the obtained equilibrium samples for {E1, E2, E3, …, Ei}, training classification rules {M1, M2, M3, …, Mi} are respectively formulated. The classifier Mi of the current data chunk comprises of {m1, m2, m3, …, mT}. The initial weight Wi of each classifier is determined by the performance of the classifier in the data chunk that arrives later. Finally, the prediction result F(i−1) is weighted by W(i−1) to obtain the final prediction result Fi on the current testing data chunk i. Entropy-based balanced strategy Normally, the imbalance ratio of a data chunk is calculated by dividing the large samples by minority samples. However, it can not describe the information content of different samples. EID [36] measures the information content of different samples and calculates the imbalance rate in terms of the difference between them. This manner of introducing the effective information content when calculating an imbalanced ratio can significantly improve the robustness of classifiers in imbalanced problems. Therefore, we propose entropy-based balanced strategy, that introduces the effective entropy of the samples to balance the data chunk. For each data chunk, we calculate the entropy of each class. Entropy represents the expected amount of information for the corresponding category. Then, we calculate the entropy of the current data chunk and divide the current data chunk into T subsets by the entropy of minority class. The problem of undersampling, which removes examples from the majority class, could cause missing concepts pertaining to the majority class of the classifier. Therefore, for each balanced sample pair divided by the entropy, we first fixed high quality minority class samples and select random common minority class samples, based on the number of minority class samples to improve the accuracy. To further strengthen our approach, we implement an undersampling technique to randomly select an equal number of majority class samples for each subset. These subsets are then used to train individual sub-classifiers, which collectively form the ensemble classifier for the given data chunk. By aggregating multiple classification rules, we mitigate the potential loss of majority class concepts, thereby enhancing the overall performance and robustness of our classification model. The specific process is as follows: ρ(xi)=1q·∑c=1qdist(xi,s(xi)c),0 θ then 5: HP ← m(i) 6: else 7: CP ← m(i) 8: end if 9: end for 10: Dn ← bootstrap Np negative instances from Dnc(t) 11: ins ← bootstrap Ncp instances from CP 12: Dp ← HP + ins 13: ht = BaseLearner(Dn, Dp) As shown in Fig 2, the circles denote the current data chunk minority class samples, and the triangles represent current data chunk majority class samples. We assume that the K value of the nearest neighbor is 3, and the threshold to determine sample quality θ is 0.6. S1 is an outlier, where its three nearest neighbors are another class and m(S1) = 0. In Fig 2(a), m(S2)=31+1+1=1>θ Fig 2(b), m(S3)=32+2+2=0.5<θ Fig 2(b), m(S4)=21.5+1=0.8>θ Fig 2(c), where S2, S4 are divided into high quality samples, and S3 is divided into common samples. Minority samples division. Algorithm 3 EDAC Input: Data chunk at timestamp t:D(t) = {xi ∈ X, yi ∈ Y}, i = 1, 2, …, N; The positive instances of D(t):Dpct,c=1,2,…,Np; The negative instances of D(t):Dnc(t),c=1,2,…,Nn; The threshold of classificaer weight δ; the number of dividing data chunk: T Output: prediction y′ 1: When the first data chunk arrives, where t = 1, each instance y’ in the data chunk D(1) is tested directly. 2: for t ← 2 to N do 3: y′=sign(∑l=1mwj(t-1)·Hj(t-1)(xt)) 4: end for 5: for j ← 1 to m do 6: the error αj(t) was selected by using 1 − Gmean; update weight of base classifiers: 7: wj(t)=(1-αj(t))·wj(t-1) 8: end for 9: Create a new base classifier 10: m ← m + 1 11: for i ← 1 to T do 12: H ← {h1, h2, … hT}, where {h1, h2, … hT} created by Algorithm 2 13: end for 14: The initialize weight of H is obtained by calculating its error rate ϵ1 in the samples of current data chunk: wm(t)←1-ϵ1 15: Persist the classifier which the weight higher than δ: 16: H(t)←{H(t)|wj(t)>δ} 17: m ← |H(t)| Dynamic weight adjustment strategy for the ensemble classifier Combining Algorithms 1 and 2, this subsection elaborates the EDAC and its integration with the with dynamic weight adjustment strategy. For each arriving data chunk Di, we construct a new classifier Hi, where T denotes the number of dividing data chunk obtained by Algorithm 1. The subclassification rules hi are obtained by Algorithm 2. The initial weight w1 is obtained by calculating its error rate ϵ1 in the samples of the current data chunk: w1 = 1 − ϵ1. We conduct a simple test of the newly created classifier to evaluate its classification performance. This self-feedback strategy can better reflect the situation when the classifier is generated. To reduce the influence of previous classifiers on new data stream samples and adapt to the newly arrived samples, the weight w1 of the generated classifier is gradually reduced with the arrival of the new data chunk Di+1. The error αi can be calculated by evaluation functions such as G-mean, F-measure, precision, or recall. The formula for weight adjustment is: wi=(1-αi)·wi-1,i∈2,3,…,n, (8) Complexity analysis of EDAC Through Algorithm 1, we can see the main complexity is the loop for calculating the entropy of samples. So, the complexity of Algorithm 1 can be bounded within O(k ⋅ n(n + 1)). Algorithm 2 mainly to calculate the density of each minority sample. The complexity for Algorithm 2 is O(k ⋅ n(n − 1)) + O(k ⋅ n); As we have analyzed above, we can find the complexity of EDAC largely depends on the step of Algorithm 1 O(n(n + 1)). Therefore, the complexity of algorithm 3 is O(n(n + 1)). Evaluation This section is to give the setting of the algorithm in detail and describes the imbalanced data stream with concept drift used in the experiment, including four artificial data streams, one real-world data stream. Furthermore, we compare our proposed approach with five other methods to demonstrate that our proposed approach can effectively process the imbalanced data streams with concept drift. Experimental setup and parameters setting To evaluate the effectiveness of the proposed method, we summarize and compare five methods of processing data stream samples. The specific parameters of the comparison algorithm are as follows: The number of nearest neighbors for REA is 10 and the post-balance ratio is 0.5. For EDAC, the error rate xi is the number of classification errors divided by the total number of samples. The threshold to determine sample quality xi is 0.1. We adopt UOB as the data preprocessing method of DWM and DWMIL. The ADAB and MWMOTEB are taking bagging as the basic framework, ADA, and MWMOTE are used as sampling methods separately. The number of partitions for each data chunk T of DWMIL, ADAB, MWMOTEB is 11. Update T value every ten data chunks for EDAC. The classification and regression tree (CART) as the learner of the base classifier for all methods, and for DWMIL and EDAC methods, the threshold for removing the dated classifier θ is set at 0.001. We detect concept drift instances by the CBCE algorithm. All experimentals are iterated 10 times to get the average result. All the other compared approaches have used the default parameter. We adopt a test-then-train strategy to evaluate the performance of the methods on each chunk, and use Recall and Area Under Curve (AUC) as the evaluation metrics to compare all methods. All these methods are implemented on an intel i5–3230m CPU, 4G RAM, and a single NVIDIA 610 GPU with MATLAB 2016a. In the experiments, we select four synthetic data streams: SEA, Moving Gaussian, Hyper Plane, Checkerboard, and one real-world data stream: Electricity(All data streams can be obtained in the attachment). The percentage of the minority class samples in each chunk is 5% of the chunk size. It can be known that only 5% of the examples inside each data chunk belong to the minority class. SEA has three attributes randomized in [0, 10]. Where the first two attributes determine the labels of the data, the third attribute is irrelevantly treated as noise. We divided the SEA with 100000 instances into 100 data chunks. Moving Gaussian has two attributes and contains two classes of Normal distribution where the coordinates of the two classes are [5, 0] and [7, 0] gradually move to [-5,0] and [-3,0]. We divided the Moving Gaussian with 50,000 instances into 50 data chunks. Hyper Planes concept of data changes gradually xi, where the attributed is 10, xi is used to decision hyperplane. And it was divided into 50 data chunks. Checkerboard is selected from the rotating chessboard, which has two attributes and 60000 instances and was divided into 200 data chunks. Electricity is the electricity price fluctuation up/down affected by demand and supply of the market from New South Wales, Australian. We divided the Electricity with 27,549 instances with seven attributes into 56 data chunks. Compared algorithms We have compared the proposed algorithm with five state-of-the-art algorithms: REA, DWMIL, DWM, ADAB and MWMOTEB. We hereby give the general idea of each algorithm is as follows: REA: The REA algorithm selects the previous minority class samples that have some correlations with the current minority class samples to make the data chunks achieve balance. Training a classifier for each balanced data chunk, and combined together as an ensemble classifier in a dynamically weighted manner. Time Complexity: O(n * m) for selecting relevant samples, where n is the number of previous minority samples and m is the number of current minority samples. Space Complexity: O(n) for storing the historical minority samples. Ensemble Complexity: REA trains multiple classifiers on balanced chunks and combines them into an ensemble. Time Complexity: O(c * (n + m)) for training c classifiers on balanced chunks. Space Complexity: O(c) for storing the ensemble. REA: The REA algorithm selects the previous minority class samples that have some correlations with the current minority class samples to make the data chunks achieve balance. Training a classifier for each balanced data chunk, and combined together as an ensemble classifier in a dynamically weighted manner. Time Complexity: O(n * m) for selecting relevant samples, where n is the number of previous minority samples and m is the number of current minority samples. Space Complexity: O(n) for storing the historical minority samples. Ensemble Complexity: REA trains multiple classifiers on balanced chunks and combines them into an ensemble. Time Complexity: O(c * (n + m)) for training c classifiers on balanced chunks. Space Complexity: O(c) for storing the ensemble. DWMIL: The DWMIL algorithm adopts UOB method to balance the data chunk, the initial weight is fixed to 1, and the weight of the classifier is adjusted dynamically according to the performance of the classifier. Finally, combined together as an ensemble classifier. Balancing Complexity: Uses the Unordered Oversampling Bagging (UOB) method to balance the data. UOB involves creating bags of samples from the original dataset, which is less complex than more sophisticated balancing techniques. Time Complexity: O(b * (n + m)), where b is the number of bags. Space Complexity: O(b * (n + m)). Ensemble Complexity: Adjusts the weights of the classifiers based on their performance, then combines them into an ensemble. Time Complexity: O(c * (n + m)) for training c classifiers. Space Complexity: O(c). DWMIL: The DWMIL algorithm adopts UOB method to balance the data chunk, the initial weight is fixed to 1, and the weight of the classifier is adjusted dynamically according to the performance of the classifier. Finally, combined together as an ensemble classifier. Balancing Complexity: Uses the Unordered Oversampling Bagging (UOB) method to balance the data. UOB involves creating bags of samples from the original dataset, which is less complex than more sophisticated balancing techniques. Time Complexity: O(b * (n + m)), where b is the number of bags. Space Complexity: O(b * (n + m)). Ensemble Complexity: Adjusts the weights of the classifiers based on their performance, then combines them into an ensemble. Time Complexity: O(c * (n + m)) for training c classifiers. Space Complexity: O(c). DWM: In DWM algorithm, the rule for classification on base classifier has been kept in a unchanged state. And the weight on base classifier is dynamically modified on each arrival data accordingly. Classification Complexity: Keeps the classification rule constant but dynamically modifies the weights of base classifiers. Time Complexity: O(n + m) for updating weights and making predictions. Space Complexity: O(c) for storing the classifiers and their weights. DWM: In DWM algorithm, the rule for classification on base classifier has been kept in a unchanged state. And the weight on base classifier is dynamically modified on each arrival data accordingly. Classification Complexity: Keeps the classification rule constant but dynamically modifies the weights of base classifiers. Time Complexity: O(n + m) for updating weights and making predictions. Space Complexity: O(c) for storing the classifiers and their weights. ADAB: The ADAB algorithm adopts ADASYN method to balance the data chunk. For each balanced data chunk, training a classifier and combined together as an ensemble classifier in a dynamically weighted manner. Balancing Complexity: Uses ADASYN (Adaptive Synthetic Sampling) to balance the data. ADASYN generates synthetic samples near the minority class samples, which can be computationally expensive.Time Complexity: O(n2) for generating synthetic samples, where n is the number of minority samples. Space Complexity: O(n) for storing the synthetic samples. Ensemble Complexity: Trains classifiers on balanced data and combines them into an ensemble. Time Complexity: O(c * (n + m)). Space Complexity: O(c). ADAB: The ADAB algorithm adopts ADASYN method to balance the data chunk. For each balanced data chunk, training a classifier and combined together as an ensemble classifier in a dynamically weighted manner. Balancing Complexity: Uses ADASYN (Adaptive Synthetic Sampling) to balance the data. ADASYN generates synthetic samples near the minority class samples, which can be computationally expensive.Time Complexity: O(n2) for generating synthetic samples, where n is the number of minority samples. Space Complexity: O(n) for storing the synthetic samples. Ensemble Complexity: Trains classifiers on balanced data and combines them into an ensemble. Time Complexity: O(c * (n + m)). Space Complexity: O(c). MWMOTEB: The MWMOTEB algorithm implements data chunk balance by using MWMOTE method. Then training a classifier for each data chunk. Finally, combined together as an ensemble classifier in a dynamically weighted manner. Balancing Complexity: Uses MWMOTE (Modified Wilson’s Minority Oversampling Technique) to balance the data. MWMOTE is similar to ADASYN but may involve different steps for generating synthetic samples. Time Complexity: O(n2). Space Complexity: O(n). Ensemble Complexity: Similar to ADAB. Time Complexity: O(c *(n + m)). Space Complexity: O(c). MWMOTEB: The MWMOTEB algorithm implements data chunk balance by using MWMOTE method. Then training a classifier for each data chunk. Finally, combined together as an ensemble classifier in a dynamically weighted manner. Balancing Complexity: Uses MWMOTE (Modified Wilson’s Minority Oversampling Technique) to balance the data. MWMOTE is similar to ADASYN but may involve different steps for generating synthetic samples. Time Complexity: O(n2). Space Complexity: O(n). Ensemble Complexity: Similar to ADAB. Time Complexity: O(c *(n + m)). Space Complexity: O(c). Comparison and analysis The AUC, G-mean, and Recall are used as evaluation indexes, and the strategy of testing before training is used to evaluate various algorithms on each data chunk. The average overall prediction accuracy of the comparative algorithms is shown in Tables 1–3. Under the metric of AUC by comparing with other algorithms, the EDAC algorithm has the better performance in Moving Gaussian, Hyperplane, Checkerboard, and Electricity. The best performance is in the Checkerboard data stream, which increases 28% compared with the DWM algorithm. On the SEA data stream, the values of the EDAC algorithm are basically the same as the other five algorithms. For the G-mean evaluation criteria, the best performance is on the Hyper Plane data stream, 28% higher than that of the REA algorithm. Because we increase the recognition accuracy of samples by introducing effective information content. The Recall value of the EDAC algorithm is obviously superior to other algorithms. In which, the best performance is in Moving Gaussian data stream, which increases 55% compared with REA algorithm. 10.1371/journal.pone.0311133.t001Table 1The average overall prediction accuracy of different algorithms in AUC measurement. Data READWMILDWMADABMWMOTEBEDAC Moving Gaussian 0.7740.8550.8010.8400.7810.867 SEA 0.9780.9760.9230.9890.9830.981 Hyper plane 0.6380.6800.5650.6070.5490.698 Checkerboard 0.8390.8860.6090.8820.8730.893 Electricity 0.7730.8300.6190.8130.7850.830 The average overall prediction accuracy of different algorithms in AUC measurement. 10.1371/journal.pone.0311133.t002Table 2The average overall prediction accuracy of different algorithms in gm measurement. Data READWMILDWMADABMWMOTEBEDAC Moving Gaussian 0.2190.7600.7830.8120.5430.800 SEA 0.8930.9200.9050.9350.9220.921 Hyper plane 0.3360.5890.5500.4040.3940.616 Checkerboard 0.7570.8110.5960.8150.7890.823 Electricity 0.6210.7070.5870.6220.5240.718 The average overall prediction accuracy of different algorithms in gm measurement. 10.1371/journal.pone.0311133.t003Table 3The average overall prediction accuracy of different algorithms in recall measurement. Data READWMILDWMADABMWMOTEBEDAC Moving Gaussian 0.1110.6500.6370.5150.3330.661 SEA 0.8680.9360.8960.9340.9080.956 Hyper plane 0.1480.4780.5610.2070.2120.628 Checkerboard 0.7090.7710.6740.7260.6760.816 Electricity 0.5570.6790.6500.4620.3580.729 The average overall prediction accuracy of different algorithms in recall measurement. EDAC algorithm dynamically updates the weight of the classifier, through the evaluation index value, and constantly eliminates the classifier with a lower weight. Compared with the DWM algorithm and DWMIL algorithm that only uses a single evaluation index to update the classifier by Single method that reduces the weight of the classifier through the increase of time. The number of negative samples identified in all negative samples is larger than other algorithms, and because the recall value is better than other algorithms, the g-mean value is better than other algorithms; AUC value is the area enclosed by the ROC curve and coordinate axis. EDAC improves the recognition accuracy of positive samples and negative samples through various mechanisms, making the model more robust, and the area under ROC offline is larger than other algorithms, finally making AUC value better than other algorithms. In general, the EDAC algorithm has better performance than other algorithms in dealing with imbalanced data streams with conceptual drift. The AUC and Recall performance on each chunk are show in Figs 3–12. Since the REA, ADAB, MWMOTEB, DWMIL, EDAC use the same base classifier, they all have roughly the same initial value. For the metric of AUC, the EDAC algorithm shows excellent performance in various data streams. ADAB is almost as good as MWMOTEB, since they adopt a similar sampling mechanism and ensemble algorithm. On average, the DWM algorithm is the worst of all algorithms, DWMIL algorithm is slightly better than the REA algorithm. EDAC and DWMIL both have a good ability to resist concept drift. But the performance of EDAC is higher than that of DWMIL. They adopt a similar classifier weight update strategy, but EDAC increases the recognition accuracy of minority instances in the sampling process. Illustration for AUC performance of each chunk on data streams Checkerboard. Illustration for AUC performance of each chunk on data streams Electricity. Illustration for AUC performance of each chunk on data streams Hyper Plane. Illustration for AUC performance of each chunk on data streams Moving Gaussian. Illustration for AUC performance of each chunk on data streams SEA. Illustration for Recall performance of each chunk on data SEA. Illustration for Recall performance of each chunk on data streams Checkerboard. Illustration for Recall performance of each chunk on data streams Electricity. Illustration for Recall performance of each chunk on data streams Hyper Plane. Illustration for Recall performance of each chunk on data streams Moving Gaussian. For Recall, the CBCE algorithm keeps updating itself to adapt to the continuous concept drift of Moving Gaussian data stream, so the identification of minority instances is low. The results about Moving Gaussian show that CBCE predicts all instances as negative and therefore receives 0 in Recall. REA algorithm stores the minority instances in the previous data chunk. But due to concept drift, majority instances would be considered as minority instances in some data chunks. It could lead to the performance of algorithm REA which is possible to introduce noise instances into classifier training getting worse and worse. Therefore, it gradually drops to 0. It is obviously shown in the SEA data stream that the recovery of algorithm REA is the slowest when the concept changes greatly. Generally speaking, EDAC recovers its better performance quickly compared with other methods. DWMIL also responds quickly, but its recognition accuracy of minority samples is far less than that of the EDAC algorithm. DWM needs to spend about 10 chunks to recover the performance. ADAB and MWMOTE use an oversampling algorithm to balance data chunk, but the classifier they trained in the current data chunk was used to evaluate the data chunk arrived later. In this case, the performance of the sampling method is poor, even worse than the DWM algorithm which does not process anything. These three indicators evaluate the performance of classification models from different perspectives. AUC provides information about the overall discriminative ability of the classifier, G-Means helps balance accuracy and recall, especially in cases of imbalanced categories, while Recall focuses particularly on the completeness of positive example recognition. Our algorithm EDAC may have lower GM values than other algorithms in certain specific situations. Overall, our algorithm EDAC has achieved excellent performance in all three indicators. In the later stage, we will increase the number of sample points and sacrifice a small amount of time to increase the accuracy of the algorithm. Results of time complexity Table 4 shows the running time of the six algorithms in the five data streams. It can be known that the cost of EDAC algorithm is less than the other methods on four data streams. The reason is that EDAC adopts dynamically bagging with twice lower than fixed T decision trees Compared with DWMIL. And in the process of sampling, we fixed selection a part of high quality minority class samples, which saves time in each basic classifier training compared with the UOB algorithm. REA maintains a single decision tree for each chunk, and keep all the decision tree trained in each data chunk. This method is suitable for small-scale samples. Thus, REA is faster for the Electricity data stream which has a smaller number of instances and chunks. However, it uses the KNN algorithm to select similar samples from all the previous minority class samples to the current class. As the number of chunks increases, the number of calculations increases. On the Checkerboard which contains 200 chunks, the cost of EDAC is obviously lower than that of REA. The running cost of algorithm DWM is the largest for all compared algorithm, which updates the ensemble classifiers on every incoming instance where other algorithms are based on data chunks, and the processing time will be faster than this one-to-one learning method. ADAB, MWMOTEB and DWMIL use the same data chunk processing method, but the sampling method is different, where ADAB adopts ADASYN sampling method, MWMOTEB applies MWMOTE sampling method, DWMIL exploits UOB sampling method. Therefore, the time spent by ADAB and MWMOTEB is not much different, while the time spent by DWMIL is relatively less. 10.1371/journal.pone.0311133.t004Table 4List of the time used by each algorithm. Data READWMILDWMADABMWMOTEBEDAC Moving Gaussian 11.910.6133.615.925.08.0 SEA 45.632.8224.236.659.524.3 Hyper plane 14.510.9383.319.243.710.4 Checkerboard 282.754.6272.378.7176.950.8 Electricity 4.812.895.716.317.27.9 List of the time used by each algorithm. Conclusions In this paper, we investigate the problem of imbalanced data stream with concept drift. The EDAC algorithm has been proposed to tackle such issue. The main idea of this EDAC is to leverage the entropy value of data samples to calculate the effective information content of samples. Based on the gained effective information above, EDAC divides the imbalanced data chunk into multiple balanced sample pairs. Next, with aim of developing the capability of decision boundary, the base classifier is built upon balanced sample pairs by introducing the effective information content of samples into classifier creation. Also, the weight on each based classifier is calculated and updated based on forthcoming data chunks. Last, we have conducted the simulation via synthetic data streams but also in real-world data streams. Compared to the previous work, the results have shown that the proposed algorithm has the best performance for improvement of the classification accuracy towards the imbalanced data stream with concept drift. Supporting information (XLSX) (XLSX) (XLSX) (XLSX) (XLSX) (DOCX) (ZIP) 18 Jan 2024 PONE-D-23-37090Entropy-based Dynamic Ensemble Classification Algorithm for Imbalanced Data Stream with Concept DriftPLOS ONE Dear Dr. Gong, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Mar 03 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Suja A Alex, Ph.D. Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. Please note that PLOS ONE has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, all author-generated code must be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Partly ********** 2. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: No ********** 3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: No ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes ********** 5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: There are several queries after reading the submitted manuscript. My comments are as follows: 1. The utility of the author's proposed methodology is unclear to the scientific readers. Further, it is not clear that the proposed methodology is applicable to which kind of dataset. 2. What is the need to consider synthetic dataset when several open datasets are available. 3. Authors have not explored the popular ML algorithms like SVM, ANN, Random Forest etc. Rather, the least explore classification algorithms were discussed. When a new concept is presented, it should be applied on open datasets as well as established algorithms. 4. The AUC, Recall and GM values are comparatively less.In such situation,what is the utility of proposed methodology. 5. Data imbalance have been addressed successfully with GAN which is reported in recently published literatures.What is the novelty or significant findings as compared to published literature. ********** 6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 12 Mar 2024 Dear editor, Thank you very much for your guidance, requirements, and valuable suggestions. I have made the necessary revisions and provided the data materials as requested. On the submission page, I have uploaded a revised manuscript showing the changes, a clean copy of the revised manuscript, a Respond to Reviewers document, and a zip file containing the data files. Best regards, Jiaming Gong Submitted filename: Response to Reviewers.docx 3 Jun 2024 PONE-D-23-37090R1Entropy-based Dynamic Ensemble Classification Algorithm for Imbalanced Data Stream with Concept DriftPLOS ONE Dear Dr. Gong, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Your manuscript has been evaluated by two new reviewers, and their comments are appended below. While Reviewer 2 is satisfied with your revisions, Reviewer 3 has raised concerns regarding the clarity of the advantages of your approach, comparisons made to other methods, and reporting of the performance of your method. Please note that while novelty in the sense of 'impact' is not a requirement for publication in PLOS ONE, submissions to the journal must present a contribution to academic knowledge, and we understand the reviewer to be requesting that you clarify what contribution your work presents over other work in the field. Please ensure you address each of the reviewer's comments when revising your mansucript. Please submit your revised manuscript by Jul 15 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Hugh Cowley Staff Editor PLOS ONE [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #2: (No Response) Reviewer #3: All comments have been addressed ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #2: Yes Reviewer #3: Partly ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #2: Yes Reviewer #3: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #2: Yes Reviewer #3: No ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #2: Yes Reviewer #3: No ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #2: The authors developed a new approach to tackle the problem of imbalanced data stream with concept drift. The strategy was extensively explained through texts, algorithms, and images. It was validated with five data streams (four synthetic and one real) and compared with five state-of-the-art algorithms. All results were presented in tables and graphs and clearly discussed. All data streams were made available as an attachment. Reviewer #3: 1. Entropy-based dynamic ensemble classification algorithm (EDAC) are not clearly articulated. You need to provide more compelling evidence and explanations to demonstrate the advantages of your approach over existing methods and clarify the specific types of datasets it is suited for. 2. In your related work section and discussion, you should provide a more thorough comparison of EDAC with these GAN-based methods, focusing on the novelty and specific advantages of your approach. 3. You need to provide a more comprehensive analysis of your method's performance, discuss the implications of these results, and outline strategies for improvement. Optionally, Consider exploring alternative evaluation metrics that may better highlight the strengths of your approach. ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #2: No Reviewer #3: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 29 Jun 2024 Dear reviewer, Response to Reviewers I am very grateful for the valuable feedback you and Reviewer have provided on my manuscript. I have carefully read the feedback from commentator and conducted in-depth reflection on it. I will explain and respond to the questions raised by commentator one by one. 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. The dataset mentioned in the document is in Attachment 《dataset》, the data points behind the mean, median, and variance measures in the experiment are shown in Appendix 《Experimental Results Chart》. And we provided explanations for obtaining the dataset in the manuscript I deeply apologize for the issue of whether all the data in the manuscript has been fully provided. I understand the requirements of the PLOS data policy and will do my best to provide the basic data for all findings described in the manuscript. In the revised manuscript, I will provide relevant data as part of the manuscript or its supporting information in accordance with policy requirements. In the experiments, we select four synthetic data streams: SEA, Moving Gaussian, Hyper Plane, Checkerboard, and one real-world data stream: Electricity. The percentage of the minority class samples in each chunk is 5% of the chunk size. It can be known that only 5% of the examples inside each data chunk belong to the minority class. SEA has three attributes randomized in [0,10]. Where the first two attributes determine the labels of the data, the third attribute is irrelevantly treated as noise. We divided the SEA with 100000 instances into 100 data chunks. Moving Gaussian has two attributes and contains two classes of Normal distribution where the coordinates of the two classes are [5,0] and [7,0] gradually move to [-5,0] and [-3,0]. We divided the Moving Gaussian with 50,000 instances into 50 data chunks. Hyper Planes concept of data changes gradually xi, where the attributed is 10, xi is used to decision hyperplane. And it was divided into 50 data chunks. Checkerboard is selected from the rotating chessboard, which has two attributes and 60000 instances and was divided into 200 data chunks. Electricity is the electricity price fluctuation up/down affected by demand and supply of the market from New South Wales, Australian. We divided the Electricity with 27,549 instances with seven attributes into 56 data chunks. The attributes of the public dataset used in this article are described above. We will upload the relevant datasets and experimental test results in the form of attachments. 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Thank you for raising the question about the manuscript's presentation and language. We have taken great care to ensure that our manuscript is presented in an intelligible fashion and written in standard English. Our manuscript was completely independently written by ourselves, fully meeting the writing requirements of our journal. Our manuscript has been polished by the Editage and revised based on the feedback provided. 6. Review Comments to the Author Reviewer #3: 1. Entropy-based dynamic ensemble classification algorithm (EDAC) are not clearly articulated. You need to provide more compelling evidence and explanations to demonstrate the advantages of your approach over existing methods and clarify the specific types of datasets it is suited for. Firstly, thank you very much for your careful review and valuable feedback on our article. We attach great importance to your question about the utility of the method and its applicability to the dataset. Here are our specific responses to these questions: We understand that the clarity of the utility of methods may cause confusion for scientific readers. To address this issue, we will add more explanations and examples on the effectiveness of the method in the revised draft. We will demonstrate the effectiveness and advantages of the method through specific experimental results, data analysis, and charts, so that readers can have a clearer understanding of the value and significance of the method in practical applications. As for the issue of the method being applicable to datasets, we realize that the article does not explicitly state which type of dataset the method is applicable to. In the revised draft, we will add relevant content to clearly explain the types and scope of datasets that this method is applicable to. We will list some suitable dataset features and requirements for using this method based on existing research and practical experience, so that readers can better judge whether the method is suitable for their research or project. 2. In your related work section and discussion, you should provide a more thorough comparison of EDAC with these GAN-based methods, focusing on the novelty and specific advantages of your approach. Firstly, we note that there have been some recent literature reports on using GANs (Generative Adversarial Networks) to address data imbalance issues, demonstrating the potential of GANs in generating synthetic data to balance the number of samples in different categories. But GANs mainly solves the problem of static imbalanced data classification. As it is a neural network algorithm, it takes a long time to process high-speed data streams and is not suitable for solving dynamic data stream classification problems. Compared with published literature, the novelty and important findings of our study are mainly reflected in the following aspects: The combination of integrated classification: For the first time, we have combined the integrated classification method to solve the problem of data imbalance. By integrating multiple classifiers, our method can fully utilize the advantages of different classifiers, improve overall classification performance, and better handle the challenges brought by data imbalance. Improving the quality of generated samples: we introduce a feedback mechanism for ensemble classifiers to guide the training of the generator. In this way, the generator can generate more representative and diverse samples, thereby improving the performance of the classifier. This mechanism helps the generator better capture the details of data distribution and generate more challenging samples to enhance the classifier's generalization ability. Improving classifier performance: Our method achieves more comprehensive and detailed classification of data by integrating multiple classifiers. Each classifier can focus on different subsets or features of data, thereby improving overall classification performance. In addition, we have introduced a dynamic integration strategy to dynamically adjust the weights of each classifier based on its performance, in order to further improve classification accuracy. Comprehensive evaluation and optimization: We not only comprehensively evaluated the quality of generated samples and classifier performance, but also proposed an optimization strategy to balance the training between the generator and classifier. This helps to ensure that the samples generated by the generator are both representative and diverse, while also ensuring that the classifier can fully utilize these samples for training. In summary, our study provides a new solution for addressing data imbalance by combining integrated classification. This combination not only improves the quality of generated samples, but also enhances the performance of the classifier, providing strong support for achieving more accurate classification. 3. You need to provide a more comprehensive analysis of your method's performance, discuss the implications of these results, and outline strategies for improvement. Optionally, Consider exploring alternative evaluation metrics that may better highlight the strengths of your approach. Thank you very much for carefully reviewing our paper and providing valuable feedback. We fully agree with your viewpoint on the need for a more comprehensive analysis of method performance, discussion of the impact of results, and suggestions for improvement. We have added an explanation of the meaning of the results and outlined improvement strategies in 2.3 Comparison and Analysis:These three indicators evaluate the performance of classification models from different perspectives. AUC provides information about the overall discriminative ability of the classifier, G-Means helps balance accuracy and recall, especially in cases of imbalanced categories, while Recall focuses particularly on the completeness of positive example recognition. Our algorithm EDAC may have lower GM values than other algorithms in certain specific situations. Overall, our algorithm EDAC has achieved excellent performance in all three indicators.In the later stage, we will increase the number of sample points and sacrifice a small amount of time to increase the accuracy of the algorithm. Thank you again for your valuable feedback and attention. We look forward to receiving your further guidance and support. Best regards, Jiaming Gong Submitted filename: Response to Reviewers.docx 23 Jul 2024 PONE-D-23-37090R2Entropy-based Dynamic Ensemble Classification Algorithm for Imbalanced Data Stream with Concept DriftPLOS ONE Dear Dr. Gong, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Sep 06 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Unil Yun, Ph.D. Academic Editor PLOS ONE Additional Editor Comments: Additional revision is required. Please note that I have acted as a reviewer for this manuscript, and you will find my comments below, under Reviewer 4. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #2: All comments have been addressed Reviewer #4: All comments have been addressed ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #2: (No Response) Reviewer #4: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #2: (No Response) Reviewer #4: No ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #2: (No Response) Reviewer #4: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #2: (No Response) Reviewer #4: Yes ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #2: (No Response) Reviewer #4: Authors suggested an approach of Entropy-based Dynamic Ensemble Classification Algorithm for Imbalanced Data Stream with Concept Drift. 1. The suggested algorithms need to be updated in detail. 2. Theoretical analysis is neexed. The theoretical complexity analysis and comparision with the state of the srt approach is needed. 3. Recent related references published on 2024, 2023, 2022 need to be added and analyzed. 4. There are still inconsistent sentences. Tight proofreading is needed. ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #2: No Reviewer #4: Yes: Unil Yun ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 15 Aug 2024 Dear reviewer, Response to Reviewers I am very grateful for the valuable feedback you and Reviewer have provided on my manuscript. I have carefully read the feedback from commentator and conducted in-depth reflection on it. I will explain and respond to the questions raised by commentator one by one. 3. Has the statistical analysis been performed appropriately and rigorously? Descriptive Statistical Analysis and Performance Metrics: Indeed, we have employed descriptive statistical analysis methods, specifically by calculating the Area Under the Curve (AUC), Recall, and G-Means, to comprehensively evaluate the performance of our proposed model, EDAC. These metrics were chosen as they provide a multi-faceted view of the model's classification capabilities: AUC quantifies the model's overall discriminative power, giving a measure of how well the model can distinguish between classes across different decision thresholds. Recall emphasizes the ability of the model to identify all positive instances, crucial in applications where false negatives are costly. G-Means helps assess the trade-off between precision and recall, addressing the challenge of imbalanced datasets by striving for a balance between these two aspects of performance. Rigorousness and Limitations: While these metrics offer a robust foundation for evaluating our model, we acknowledge that no single metric can fully encapsulate all aspects of model performance. Therefore, we have taken a multi-metric approach to ensure a comprehensive assessment. 6. Review Comments to the Author (1)The suggested algorithms need to be updated in detail. We acknowledge that a more detailed description of the proposed algorithms is essential for readers to fully understand and replicate our work. In the revised manuscript, we have expanded Section 1 The proposed algorithm: EDAC , to provide a step-by-step explanation of the Entropy-based Dynamic Ensemble Classification Algorithm. This includes a detailed algorithm flowchart, pseudo-code, and explanations of each component's functionality and purpose. We have also included additional figures and examples to illustrate the algorithm's working principles and decision-making processes. (2) Theoretical analysis is neexed. The theoretical complexity analysis and comparision with the state of the srt approach is needed. We fully agree that a thorough theoretical analysis and comparison with the state-of-the-art approaches are crucial to demonstrate the merits and positioning of our work. In response to your suggestion, we have strengthened Section 2.2 "Theoretical Analysis and Algorithm Comparison" by including a detailed examination of the time and space complexity of our algorithm, as well as a comprehensive comparison with the most recent and relevant algorithms in the field. Conducted a Rigorous Complexity Analysis: We have carefully analyzed the time and space complexity of our algorithm, taking into account the key factors that influence its performance, such as the input data size, problem dimensionality, and any inherent parallelism or optimization strategies employed. This analysis provides a clear picture of how our algorithm scales with respect to these parameters, enabling readers to assess its suitability for various applications and datasets. (3)Recent related references published on 2024, 2023, 2022 need to be added and analyzed. We appreciate your suggestion to include recent references from 2024, 2023, and 2022. In the revised manuscript, we have thoroughly reviewed the latest literature in the field and added relevant references to Introduction, where we discuss related work. We have analyzed these recent publications, discussing how they relate to and differ from our proposed approach. This update ensures that our work is situated within the most current research landscape. (4)There are still inconsistent sentences. Tight proofreading is needed. We apologize for any inconsistencies or unclear sentences in the original manuscript. We have undertaken a rigorous proofreading process, carefully reviewing each sentence for clarity, coherence, and grammatical correctness. We have corrected any errors identified and ensured that the revised manuscript is consistent in its terminology, notation, and explanations. We believe that these changes have significantly improved the readability and accessibility of our work. In conclusion, we are grateful for your insightful comments and suggestions, which have helped us improve the quality and impact of our manuscript. We hope that the revisions we have made address your concerns and that the revised version meets your expectations. We look forward to your further feedback and the opportunity to clarify any remaining questions you may have. Thank you again for your valuable feedback and attention. We look forward to receiving your further guidance and support. Best regards, Jiaming Gong Submitted filename: Response to Reviewers.docx 12 Sep 2024 Entropy-based Dynamic Ensemble Classification Algorithm for Imbalanced Data Stream with Concept Drift PONE-D-23-37090R3 Dear Dr. Gong, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. If you have any questions relating to publication charges, please contact our Author Billing department directly at authorbilling@plos.org. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Unil Yun, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): The authors reviewed carefully the manuscript based on the reviewers' comments. Now it is acceptable. Please note that I have acted as a reviewer for this manuscript, and you will find my comments below, under Reviewer 4. Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #4: All comments have been addressed ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #4: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #4: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #4: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #4: Yes ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #4: Authors revised carefully this manuscript. Now it is acceptable. The related references are added and theoretical snalysis are also improved. ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #4: No ********** 23 Sep 2024 PONE-D-23-37090R3 PLOS ONE Dear Dr. Gong, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset If revisions are needed, the production department will contact you directly to resolve them. If no revisions are needed, you will receive an email when the publication date has been set. At this time, we do not offer pre-publication proofs to authors during production of the accepted work. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few weeks to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Unil Yun Academic Editor PLOS ONE