Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ODMixer: Fine-grained Spatial-temporal MLP for Metro Origin-Destination Prediction

Yang Liu, Binglin Chen, Yongsen Zheng, Lechao Cheng, Guanbin Li, Liang Lin

TL;DR

This work targets metro Origin–Destination prediction under incomplete current data by adopting a fine-grained OD-pair perspective. It introduces ODMixer, a dual-branch spatial–temporal MLP with an OD Matrix Processing module, an Embedding Layer, and the ODIM consisting of a Channel Mixer and a Multi-view Mixer, augmented by a Bidirectional Trend Learner to capture long-term patterns. Empirical results on HZMOD and SHMOD show ODMixer achieving state-of-the-art wMAPE (about 5%–7% improvement) and strong robustness, efficiency, and cross-city transfer capabilities. The approach offers practical impact for real-time metro operations, including improved scheduling and emergency response, with a clear path toward broader deployment and future enhancements such as richer city-context features and generalized transfer learning.

Abstract

Metro Origin-Destination (OD) prediction is a crucial yet challenging spatial-temporal prediction task in urban computing, which aims to accurately forecast cross-station ridership for optimizing metro scheduling and enhancing overall transport efficiency. Analyzing fine-grained and comprehensive relations among stations effectively is imperative for metro OD prediction. However, existing metro OD models either mix information from multiple OD pairs from the station's perspective or exclusively focus on a subset of OD pairs. These approaches may overlook fine-grained relations among OD pairs, leading to difficulties in predicting potential anomalous conditions. To address these challenges, we learn traffic evolution from the perspective of all OD pairs and propose a fine-grained spatial-temporal MLP architecture for metro OD prediction, namely ODMixer. Specifically, our ODMixer has double-branch structure and involves the Channel Mixer, the Multi-view Mixer, and the Bidirectional Trend Learner. The Channel Mixer aims to capture short-term temporal relations among OD pairs, the Multi-view Mixer concentrates on capturing spatial relations from both origin and destination perspectives. To model long-term temporal relations, we introduce the Bidirectional Trend Learner. Extensive experiments on two large-scale metro OD prediction datasets HZMOD and SHMO demonstrate the advantages of our ODMixer. Our code is available at https://github.com/KLatitude/ODMixer.

ODMixer: Fine-grained Spatial-temporal MLP for Metro Origin-Destination Prediction

TL;DR

This work targets metro Origin–Destination prediction under incomplete current data by adopting a fine-grained OD-pair perspective. It introduces ODMixer, a dual-branch spatial–temporal MLP with an OD Matrix Processing module, an Embedding Layer, and the ODIM consisting of a Channel Mixer and a Multi-view Mixer, augmented by a Bidirectional Trend Learner to capture long-term patterns. Empirical results on HZMOD and SHMOD show ODMixer achieving state-of-the-art wMAPE (about 5%–7% improvement) and strong robustness, efficiency, and cross-city transfer capabilities. The approach offers practical impact for real-time metro operations, including improved scheduling and emergency response, with a clear path toward broader deployment and future enhancements such as richer city-context features and generalized transfer learning.

Abstract

Metro Origin-Destination (OD) prediction is a crucial yet challenging spatial-temporal prediction task in urban computing, which aims to accurately forecast cross-station ridership for optimizing metro scheduling and enhancing overall transport efficiency. Analyzing fine-grained and comprehensive relations among stations effectively is imperative for metro OD prediction. However, existing metro OD models either mix information from multiple OD pairs from the station's perspective or exclusively focus on a subset of OD pairs. These approaches may overlook fine-grained relations among OD pairs, leading to difficulties in predicting potential anomalous conditions. To address these challenges, we learn traffic evolution from the perspective of all OD pairs and propose a fine-grained spatial-temporal MLP architecture for metro OD prediction, namely ODMixer. Specifically, our ODMixer has double-branch structure and involves the Channel Mixer, the Multi-view Mixer, and the Bidirectional Trend Learner. The Channel Mixer aims to capture short-term temporal relations among OD pairs, the Multi-view Mixer concentrates on capturing spatial relations from both origin and destination perspectives. To model long-term temporal relations, we introduce the Bidirectional Trend Learner. Extensive experiments on two large-scale metro OD prediction datasets HZMOD and SHMO demonstrate the advantages of our ODMixer. Our code is available at https://github.com/KLatitude/ODMixer.
Paper Structure (34 sections, 14 equations, 7 figures, 11 tables)

This paper contains 34 sections, 14 equations, 7 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traffic Flow Prediction
  4. Origin-Destination Prediction
  5. MLP-based Model
  6. Preliminaries
  7. Methodology
  8. Architecture Overview
  9. Dual-branch Structure
  10. OD Matrix Processing
  11. Unfinished Orders Processing
  12. Embedding Layer
  13. Origin-Destination Interact Module (ODIM)
  14. Channel Mixer
  15. Multi-view Mixer
  16. ...and 19 more sections

Figures (7)

  • Figure 1: Comparison of different views and models for metro OD prediction. (a) illustrates the difference in encoding the OD Matrix from the station view and the OD pair view. N is the number of stations, T is the number of time intervals. (b) shows various models for processing the $N^{2}$ tokens.
  • Figure 2: Comparison of flow changes between OD pairs with the same origin. The x-axis represents July 4 and 5, 2019, the y-axis represents the flow value between OD pairs. The OD: (190,196) in the legend means the origin is station 190 and the destination is station 196. Figure (a) shows that although the OD pairs with the same origin have similar change trends, the flow values are very different. Figure (b) illustrates that the flow rate trends of two OD pairs with the same origin exhibit significant differences over time.
  • Figure 3: The overall framework of the ODMixer and its essential modules. (a) illustrates the architecture of double-branch ODMixer, both branches learn features using ODIM. Subsequently, the two branches interact with each other using the BTL, contributing to the final output. (b) depicts the BTL module, which aims to jointly model the features from both branches, thereby enhancing the information exchange. (c) represents the ODIM module, which models the temporal attributes of the OD pair features using the Channel Mixer, while the comprehensive relations among OD pairs are learned using the Multi-view Mixer. (d) shows the Channel Mixer, Origin Mixer, Des Mixer.
  • Figure 4: Impact of Hyper-parameters. The two figure illustrate the impact of different parameters on prediction results across two datasets. Figure (a) displays the effect of varying the number of layers on performance, while Figure (b) shows the influence of different feature dimensions on performance.
  • Figure 5: Case for SHMOD and HZMOD Dataset. (a) shows the prediction results of the DGSL and ODMixer models from September 10 to 13, with the 12th being a Monday. (b) shows the prediction results of the HIAM and ODMixer models from January 21 to 24, with the 21th being a Monday. It can be observed that when traffic flow changes rapidly, the performance of HIAM and DGSL is significantly worse than that of ODMixer.
  • ...and 2 more figures