Enhancing Repeat-Aware Recommendation from a Temporal-Sequential Perspective Conference: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management; October 21–25, 2023; Birmingham, United Kingdom. Proceedings of the 32nd ACM International Conference on Information
Shigang Quan, Shui Liu, Zhenzhe Zheng, Fan Wu · 2025
Read Enhancing Repeat-Aware Recommendation from a Temporal-Sequential Perspective Conference: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management; October 21–25, 2023; Birmingham, United Kingdom. Proceedings of the 32nd ACM International Conference on Information by Shigang Quan, Shui Liu, Zhenzhe Zheng, Fan Wu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-08531.txt.
Introduction
Shigang Quan and three co-authors present a 2025 model for guessing what a user will buy or play again. Repeat use, such as rebuying goods or replaying songs, is a daily pattern. Past models read the list of past acts but slight the timing between them. The team argues that gaps in time and the order of acts both hint at repeat intent.
Their model, called TSRec, reads time gaps from both the user's side and the item's side. It also weighs how the latest repeat run matches the one before. Tests on three public data sets beat prior top methods. Code is shared in an open link given in the paper. The item-side timing helps when user records run thin.
Catalog record
| CREATOR | Shigang Quan, Shui Liu, Zhenzhe Zheng, Fan Wu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-08531
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Enhancing Repeat-Aware Recommendation from a Temporal-Sequential Perspective Conference: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management; October 21–25, 2023; Birmingham, United Kingdom. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM ’23), October 21–25, 2023, Birmingham, United Kingdom Price: 15.00 ISBN: 979-8-4007-0124-5/23/10 DOI: 10.1145/XXXXXX.XXXXXX 1457 CCS: Information systems Recommender systems Shigang Quan, Shui Liu, Zhenzhe Zheng, Fan Wu Abstract . Repeat consumption, such as repurchasing items and relistening songs, is a common scenario in daily life. To model repeat consumption, the repeat-aware recommendation has been proposed to predict which item will be re-interacted based on the user-item interactions. In this paper, we investigate various inherent characteristics to enhance the repeat-aware recommendation. Specifically, we explore these characteristics from two aspects: one is from the temporal aspect where we consider the time interval relationship in the user behavior sequence; the other is from the sequential aspect where we consider the sequential-level relationship in the user behavior sequence. And our intuition is that both the temporal pattern and sequential pattern will reflect users’ intentions of repeat consumption. By utilizing these two patterns, a novel model called T emporal and S equential repeat-aware Rec ommendation( TSRec for short) is proposed to enhance repeat-aware recommendation. TSRec has three main components: 1) User-specific Temporal Representation Module (UTRM), which encodes and extracts user historical repeat temporal information. 2)Item-specific Temporal Representation Module (ITRM), which incorporates item time interval information as side information to alleviate the data sparsity problem of user repeat behavior sequence. 3) Sequential Repeat-Aware Module (SRAM), which represents the similarity between the user’s current and the last repeat sequences. Extensive experimental results on three public benchmarks demonstrate the superiority of TSRec over state-of-the-art methods. The implementation code is available 1 1 1 https://anonymous.4open.science/r/TSRec-2306/ . 1. Introduction Figure 1 . Data analyses to illustrate the inherent characteristics related to repeat consumption from the temporal aspect on the LastFM dataset. The x-axis is the repeat time interval of an item and the y-axis is the repeat frequency for different time intervals. The recommendation system has achieved great success due to its ability to capture dynamic user interests across various online platforms, such as E-commerce ( Rendle et al. 2010 ) , social media ( Yin et al. 2015 ; Zhou et al. 2019 ) , music recommendation ( Kapoor et al. 2015 ; Reiter-Haas et al. 2021 ) , etc. The fundamental objective of a recommendation system is to predict the user’s next item by estimating the similarity between the target item and the user’s previously interacted items. Recently, rather than just modeling user behavior, the repeat-aware recommendation task has been proposed to explore repeat consumption, a phenomenon where users repeatedly purchase or click on certain items. Modeling users’ repeat behavior can improve the accuracy of next-item prediction. To this end, some researchers attempt to model users’ temporal behavior using distribution models, such as the Negative Binomial Distribution model ( Bhagat et al. 2018 ) and Point Process Model ( Wang et al. 2019 ) . However, these methods only focus on modeling user repeat behavior and neglect user dynamic preferences for new items, thus limiting the applicability and generalizability to scenarios with relatively low repeat ratios. Conversely, other methods, as mentioned in ( Ren et al. 2019 ; Anderson et al. 2014 ; Ariannezhad et al. 2022 ) , aim to improve upon these limitations by incorporating an additive neural component to predict the intention of user repeat consumption, or by introducing a supervised loss, with a ground truth repeat label, to regulate model training. These methods are characterized by the absence of modeling intentions behind repeat consumption, which impedes their effectiveness in striking a balance between recommending new and repeat items. In this paper, we investigate various inherent characteristics to enhance the repeat-aware recommendation. Specifically, we explore these characteristics from two aspects. One is from the temporal aspect: we consider the time interval relationship in the user behavior sequence; the other is from the sequential aspect: we consider the sequential-level relationship in the user behavior sequence. And our intuition from these two aspects is as follows: • Intuition 1: for repeatedly consumed products, users typically exhibit a fixed consumption cycle for these products. For example, after users consume milk, they are likely to repurchase it during a certain period, based on their practical demand. • Intuition 2: we may infer the likelihood of users making repeat consumption based on the inherent correlation within a subset of user behavior sequences. For example, when people listen to music, they tend to repeat a series of songs rather than a single track. The above intuitions motivate us to conduct preliminary data analysis on public datasets. After preliminary data analysis, we identified patterns in the data that align with the intuitions mentioned above from two aspects: for intuition 1, we calculated the distribution of item repeat frequency with fixed repeat time intervals on the LastFM dataset in Figure 1 (a) and the proportion of these items is 13.15%. The temporal pattern is that repeat consumption typically exhibits repeat behavior within a defined time interval; for intuition 2, we calculated the Jaccard similarity of item sets between the user behavior sequence preceding the item and the last repeat behavior sequence of the target item across three datasets and found a relatively high degree of item overlap (45%-60%), as shown in Table 1 . The sequential pattern is that repeat consumption typically manifests as a high correlation across the entire sequence. Table 1 . The calculated Jaccard similarity on three different datasets. It is derived by calculating the union and intersection between the user’s current interacted item sequence sets and the former repeat interacted item sequence sets. Dataset RetailRocket LastFM Diginetica Jacc.(%) 49.69 58.10 45.23 In this paper, we propose a novel model called T emporal and S equential repeat-aware Rec ommendation ( TSRec for short) to automatically model repeat consumption from temporal and sequential aspects. The variants of TSRec include User-specific Temporal Representation Module (UTRM), Item-specific Temporal Representation Module (ITRM), and Sequential Repeat-Aware Module (SRAM). Specifically, to capture the temporal pattern , we design a UTRM based on the target attention layer to encode and extract user historical repeat temporal information. But the above module suffers from data sparsity issues, because changes in the user’s preferences may lead to a low frequency of repeat consumption of specific items. As depicted in Figure 1 (b), even users exhibiting high-frequency repeat consumption behavior (e.g., User ID=59) have a relatively low overall frequency of repeat consumption for a target item (Item ID=41), with a frequency of nearly 10. Therefore we propose UTRM, which incorporates item time interval information from other users as side information. Our preliminary experiment in Figure 1 has shown that the item repeat time intervals for different users were relatively uniform. Based on this finding, we define the Repeat Time Interval Matrix based on the time intervals of the target item and design an external ITRM to encode the item-specific temporal information. To represent the user multi-scale temporal information, we refer to the idea of "Network In Network" (NIN) ( Lin et al. 2014 ) and design ITRM, comprising CNN and pooling layers,