TrafficKAN-GCN: Graph Convolutional-based Kolmogorov-Arnold Network for Traffic Flow Optimization

Jiayi Zhang, Yiming Zhang, Yuan Zheng, Yuchen Wang, Jinjiang You, Yuchen Xu, Wenxing Jiang, Soumyabrata Dev · 2025

Read TrafficKAN-GCN: Graph Convolutional-based Kolmogorov-Arnold Network for Traffic Flow Optimization by Jiayi Zhang, Yiming Zhang, Yuan Zheng, Yuchen Wang, Jinjiang You, Yuchen Xu, Wenxing Jiang, Soumyabrata Dev (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-2503-03276.txt.

Introduction

Jiayi Zhang and seven co-authors present a hybrid neural model for city traffic in a paper from 2025. Classic routing algorithms work on static maps but struggle with the shifting, tangled flow of real streets. Their model joins graph convolutional networks, which learn the street network's shape, with Kolmogorov-Arnold networks, which flex to fit curved patterns in the flow data.

Tests used real traffic data from the Baltimore area. The model matched rival methods on accuracy and handled noisy, irregular data better. It also rerouted flow in ways that eased congestion and adapted to a shock event, the collapse of a major bridge.

Catalog record

CREATORJiayi Zhang, Yiming Zhang, Yuan Zheng, Yuchen Wang, Jinjiang You, Yuchen Xu, Wenxing Jiang, Soumyabrata Dev
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-03276

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TrafficKAN-GCN: Graph Convolutional-based Kolmogorov-Arnold Network for Traffic Flow Optimization Jiayi Zhang, Yiming Zhang, Yuan Zheng, Yuchen Wang, Jinjiang You, Yuchen Xu, Wenxing Jiang, Soumyabrata Dev Abstract Urban traffic optimization is crucial for improving transportation efficiency and alleviating congestion. Traditional algorithms like Dijkstra’s and Floyd’s, along with Genetic Algorithms (GA), are widely used for route planning but struggle with large-scale, dynamic traffic networks due to computational complexity and limited ability to capture nonlinear spatial-temporal dependencies. To address these challenges, we propose a hybrid framework combining Kolmogorov-Arnold Networks (KAN) with Graph Convolutional Networks (GCN) for urban traffic flow optimization. KAN’s ability to approximate nonlinear functions, along with GCN’s capacity to model spatial dependencies in graph-structured data, provides a more effective representation of traffic dynamics. We evaluate KAN-GCN on real-world datasets and compare it to traditional models like Multi-Layer Perceptrons (MLP) and decision trees. While MLP-GCN shows higher accuracy, KAN-GCN excels in handling noisy and complex traffic patterns. Results highlight challenges in applying KAN to large-scale networks and suggest potential improvements. Additionally, we explore extending the framework to Transformer models for scalable and dynamic traffic forecasting. Our findings indicate that combining KAN with GCN offers promising potential for intelligent traffic management, with further optimization possible for large-scale applications. 1 Introduction On March 26, 2024, the Francis Scott Key Bridge in Baltimore collapsed after being struck by the container ship Dali (Figure 1 ). This catastrophic event led to the tragic deaths of six construction workers and caused severe disruptions to both local traffic and the broader supply chain [ 1 ] , as the bridge served as a critical artery for transportation and commerce in the region. The collapse led to significant congestion on alternative routes, including the I-95 and I-895 highways, exacerbating traffic bottlenecks and increasing travel times for both commuters and freight transport [ 2 ] . The collapse of such a vital infrastructure component underscores the pressing need for effective traffic network optimization strategies, particularly in urban environments where transportation resilience and efficiency are paramount. In response to such disruptions, robust and adaptive traffic management solutions are required to mitigate congestion and ensure the smooth operation of urban transportation systems [ 3 ] . Figure 1: Baltimore’s transportation network, showing major roads and highways (Left). The Francis Scott Key Bridge collapse on March 26, 2024, which disrupted traffic and economic activities (Right). Traditional graph-based mathematical models [ 4 ] , such as Dijkstra’s algorithm [ 5 ] , Floyd’s algorithm [ 6 ] , and Genetic Algorithms (GA) [ 7 ] , have long been utilized for route planning and traffic optimization [ 8 ] . While these models are effective in static environments [ 9 ] , they often struggle with dynamic, large-scale traffic systems due to computational constraints and their inability to adapt to real-time fluctuations in road conditions [ 10 ] . As urban road networks continue to grow in complexity, there is a pressing need for data-driven models capable of learning and predicting traffic patterns dynamically [ 11 ] . Graph learning has a wide range of applications, from recommendation systems [ 12 , 13 ] and music influence analysis [ 14 ] to path planning [ 15 , 16 ] , and our work is closely related to path planning and urban traffic optimization. Recent advancements in deep learning and graph-based machine learning have introduced novel methodologies to address these challenges [ 17 ] . In particular, Graph Convolutional Networks (GCNs) have shown strong potential in modeling traffic networks, leveraging their ability to learn spatial dependencies from graph-structured data [ 18 ] . However, conventional GCN-based approaches still face challenges in capturing nonlinear relationships and handling the highly dynamic nature of traffic conditions [ 19 ] . To ameliorate the aforementioned issues, we propose a hybrid optimization framework that integrates KAN with GCN to optimize urban traffic flow [ 20 , 21 , 22 ] . KAN [ 22 , 23 ] , a recent advancement in neural network architectures, offers superior expressiveness and function approximation capabilities, making it well-suited for capturing complex, nonlinear traffic dynamics. By combining the structural learning capabilities of GCN with the adaptive approximation power of KAN, our proposed model aims to enhance the robustness and flexibility of traffic optimization strategies. The main contributions of our method are threefold: 1. In our comparison with conventional models, we conducted a comprehensive evaluation of the KAN-GCN model against traditional machine learning approaches, such as MLP [ 24 ] and decision trees, using real-world traffic datasets. While MLP-GCN demonstrated superior prediction accuracy, KAN-GCN showed significant advantages in its robustness to noisy data and its ability to adapt to complex traffic patterns [ 21 , 25 ] . 2. Our study provides valuable insights into the challenges and opportunities of integrating KAN into urban traffic networks, highlighting potential avenues for future research to enhance traffic optimization models and strategies. 3. We propose extending the KAN-GCN framework by incorporating Transformer-based architectures for time-series traffic forecasting. This extension underscores the scalability of the model and its ability to address the dynamic and large-scale challenges inherent in traffic forecasting. This paper is organized as follows: Section 2 presents a comprehensive review of related work, starting with an analysis of graph-based traffic network models and proceeding to discuss the integration of KAN with these models. Section 3 details the dataset and the assumptions made for the study. In Section 4, we outline the data preprocessing process, including feature engineering and selection techniques. Section 5 introduces the methodology, covering traffic network modeling, the KAN-GCN architecture, loss function, and training strategy, as well as stakeholder preferences and their incorporation into the model. Section 6 provides an impact analysis on different stakeholders, and Section 7 presents the experimental evaluation, including the evaluation metrics and quantitative experiments. Finally, Section 8 concludes the paper and offers insights for future research directions. 2 Related Work 2.1 Graph-Based Traffic Network Optimization Traditional traffic network optimization has heavily relied on graph-based mathematical models, such as Dijkstra’s Algorithm and the Floyd-Warshall Algorithm, which, although effective in static scenarios, face significant challenges in dynamic urban environments. Dijkstra’s Algorithm, with O(n\log n) complexity, is efficient for small-scale networks but becomes computationally expensive in large-scale applications [ 5 ] . The Floyd-Warshall Algorithm, which computes shortest paths for all node pairs, suffers from O(n^{3}) complexity, rendering it impractical for real-time traffic optimization in large-scale networks [ 6 ] . Genetic Algorithms (GA) have also been used to optimize traffic flow by simulating evolutionary selection processes [ 4 ] , but their high computational cost and slow convergence make them less suitable for real-time applications. Recent studies have explored the use of graph-based models in other fields, such as Xu et al.’s work on AlignGroup, which leverages graph-based techniques for group recommendation [ 12 , 13 ] , and the MFCSNet model, which applies graph-based social network analysis to measure musical influence [ 14 ] . These examples demonstrate the potential of graph-based models, when combined with machine learning, to optimize traffic systems by aligning stakeholder preferences and leveraging dynamic data, offering a promising solution to overcome the limitations of traditional optimization methods in traffic flow management. With the rise of deep

Read the complete text