BiST-IF: Enhancing origin–destination flow prediction via bi-directional spatio-temporal inference and interconnected feature evolution

Piao Yu, Xu Zhang, Yongshun Gong, Jian Zhang, Haoliang Sun, Junjie Zhang, Xinxin Zhang, Yilong Yin

Expert Systems with Applications, vol. 264, article 125679 (2025)

Also indexed in: Semantic Scholar · OpenAlex · dblp · Publisher page (ScienceDirect) · XJTLU Scholar record

Quick facts

Main results

Quoted text is verbatim from the paper.

Key points

• OD Delay Correction (ODDC) completes delayed recent OD matrices using period delay probability and real-time flow features. Without it, HZMetro RMSE rises from 1.799/1.797/1.797 to 1.804/1.809/1.827 (10/30/60 min).
• OD flow Bi-directional Attention (ODBA) combines BiLSTM trend extraction, attention between origin and destination features, and a spatio-temporal feature extraction layer, applied to weekly, daily and recent OD inputs (ratio 2:1:2).
• Mutual Information Flow Evolution (MFE) uses weekly Out-OD features as the key in 4-head attention over OD features; a new gated feature update module then stabilises the periodic pattern.
• On HZMetro and NYC-TOD2018, BiST-IF beats nine baselines on every metric and horizon, with average MAE/RMSE reductions of 7.55%/7.31% and 5.77%/2.33% against the best baseline.
• In ablations, removing F-update, MI-Mech, BiLSTM, BiAttn or STExtraction hurts accuracy, yet each variant still beats the best baseline. Dropping BiLSTM raises HZMetro 10-min MAE from 0.907 to 0.937.

Abstract

Origin–destination (OD) flow prediction is crucial for predicting inter-station passenger flows in intelligent transport systems. However, previous OD prediction methods have ignored the delay of OD flows and failed to focus on the supplementary effect of the arrival OD flow (Out-OD) flow data on OD flows. We innovatively propose an OD flow prediction method based on Bidirectional Attention and Interconnected Feature Evolution (BiST-IF) to address these challenges. Firstly, we propose a correction method based on period delay probability and real-time flow features for OD flow information. Furthermore, for the flow information of different periods, we design a bi-directional attention module to achieve the preliminary prediction by deconstructing and analyzing the temporal and flow features of OD flow and fusing the temporal and spatial features. After that, we design the mutual information module of Out-OD and OD flow, combining it with the new gating algorithm to enhance the periodic features. Finally, the prediction results from the periodic pattern and the temporary fluctuation of OD flow are fused to obtain the OD flow prediction results. Extensive experiments on two real-world datasets show that the prediction performance of BiST-IF significantly outperforms other state-of-the-art models. Specifically, BiST-IF improves the MAE and RMSE by an average of 7.55% and 7.31% on the HZMetro dataset, and 5.77% and 2.33% on the NYC-TOD2018 dataset, respectively, compared to the best baseline approach.

Abstract shown up to the publicly indexed portion; see the publisher page for the full text.

Frequently asked questions

What does BiST-IF stand for?

The paper title and body expand BiST-IF as Bi-directional Spatio-Temporal Inference and Interconnected Feature Evolution. The abstract expands the same acronym as Bidirectional Attention and Interconnected Feature Evolution.

What is the OD delay problem, and how does BiST-IF correct it?

OD data are collected at regular intervals, so some passengers are still in transit at the latest collection, and ignoring these en route flows biases prediction. The OD Delay Correction (ODDC) module uses period delay probability and real-time flow features to estimate the delayed flows and complete each recent OD matrix. Without it, HZMetro RMSE rises from 1.799, 1.797 and 1.797 to 1.804, 1.809 and 1.827 at 10, 30 and 60 minutes.

What is Out-OD flow and why does BiST-IF use it?

Out-OD is the arrival OD flow: it records passengers arriving at a destination station from each origin at the time of recording, without their departure times, whereas plain OD data records departures and intended destinations but not arrival times. BiST-IF's Mutual Information Flow Evolution (MFE) module uses weekly Out-OD history to supplement this missing arrival information, with Out-OD features serving as the attention key for OD features and gated feature updates stabilising the periodic pattern.

How much does BiST-IF improve over the baselines?

BiST-IF is compared with nine baselines: HA, GRU and LSTM; the OD flow methods GEML, MPGCN and HIAM; and the urban flow methods STGCN, DGCN and STTN. Relative to the best baseline, it lowers MAE and RMSE by an average of 7.55% and 7.31% on HZMetro and by 5.77% and 2.33% on NYC-TOD2018. The paper names HIAM as the top-performing baseline.

How large is BiST-IF and how costly is it to train?

On NYC-TOD2018, BiST-IF has 3,986,921 parameters and trains at 89.69 s per epoch, against 13,851,238 parameters and 102.98 s per epoch for HIAM. The paper describes this as a moderate computational cost, and in its NYC-TOD2018 convergence comparison with STGCN, DGCN, MPGCN and HIAM, BiST-IF converges best.

Who are the first authors of BiST-IF?

Piao Yu and Xu Zhang contributed equally as co-first authors, and Yongshun Gong is the corresponding author.

How do I cite BiST-IF?

Cite it as: Yu, P., Zhang, X., Gong, Y., Zhang, J., Sun, H., Zhang, J., Zhang, X., & Yin, Y. (2025). Enhancing origin–destination flow prediction via bi-directional spatio-temporal inference and interconnected feature evolution. Expert Systems with Applications, 264, Article 125679. https://doi.org/10.1016/j.eswa.2024.125679 BibTeX is available at https://codezx6.github.io/papers/bist-if.html#cite and in https://codezx6.github.io/publications.bib.

Is the code for BiST-IF available?

Yes, at https://github.com/CodeZx6/BiST-IF.

Keywords

Origin–destination flow prediction Spatio-temporal data Bi-directional attention mechanism Traffic prediction Intelligent transport systems OD delay correction Out-OD flow Mutual information mechanism

Also referred to as: Bi-directional Spatio-Temporal Inference and Interconnected Feature Evolution; Bidirectional Attention and Interconnected Feature Evolution; OD Delay Correction (ODDC); OD flow Bi-directional Attention (ODBA); Mutual Information Flow Evolution (MFE).

Research area map and related search terms

Field path (broad to narrow): artificial intelligence › machine learning › data mining › spatio-temporal data mining › time series forecasting › intelligent transportation systems › smart cities › urban computing › traffic prediction › passenger flow prediction › origin-destination (OD) prediction › OD prediction with delayed data › BiST-IF

Task

origin-destination (OD) flow prediction OD matrix forecasting metro passenger flow prediction inter-station passenger flow forecasting taxi OD demand prediction incomplete OD data

Method

OD delay correction arrival flow (Out-OD) bi-directional attention BiLSTM mutual information attention gated feature update

Datasets

HZMetro (Hangzhou metro) NYC-TOD2018 (New York yellow taxi)

Compared with

HIAM GEML MPGCN STGCN DGCN STTN

中文

OD 流量预测 OD 矩阵预测 地铁客流预测 站间客流预测 出租车 OD 需求预测 OD 延迟校正

Related areas

travel demand modeling public transit analytics metro systems rail transit smart card data (AFC) ride-hailing demand spatio-temporal graph neural networks attention mechanisms recurrent neural networks missing data imputation data completion mutual information gating mechanisms

Applications

metro operations train scheduling crowd control in stations taxi dispatch transport planning

Related benchmarks (not used in this paper)

SHMetro NYC Citi Bike OD Chicago taxi OD Beijing subway

中文 (扩展)

客流预测 OD 预测 起讫点矩阵 出行需求建模 公共交通分析 轨道交通 地铁 刷卡数据 网约车需求 数据补全 缺失数据填补 注意力机制 循环神经网络 互信息 门控机制 地铁运营 交通规划

Terms are grouped by role. "Related areas", "Applications" and "Related benchmarks (not used in this paper)" describe the surrounding field, not results of this paper. See the site-wide research area map.

Related papers

OD demand prediction

• DSTCN: Exploiting dynamic spatio-temporal correlations for origin-destination demand prediction (Expert Systems with Applications, vol. 299, article 130095, 2026)

Other topics

• DUET: Unified Dual-Space Emotion Control for Diffusion and Flow-Matching Driven Text-to-Speech (2026)
• PhysioSER: Learning Physiology-Informed Vocal Spectrotemporal Representations for Speech Emotion Recognition (2026)
• URMDet-SimFire: Trustworthy multimodal detection with uncertainty and reliability modeling for fire and smoke analytics (2026)
• S2CMEN: A Mutually Enhancement Network for Superpixel Segmentation and Classification of Hyperspectral Image (2025)
• MR-UFP: Enhancing urban flow prediction via mutual reinforcement with multi-scale regional information (2025)
• Automatic visual recognition for leaf disease based on enhanced attention mechanism (2024)
• S3CFSL: Spatial-Spectral–Semantic Cross-Domain Few-Shot Learning for Hyperspectral Image Classification (2024)
• ST-FCL: Spatio-temporal fusion and contrastive learning for urban flow prediction (2023)
• MC-STL: Mask- and Contrast-Enhanced Spatio-Temporal Learning for Urban Flow Prediction (2023)

Cite this paper

@article{yu2025bistif,
  title        = {Enhancing origin–destination flow prediction via bi-directional spatio-temporal inference and interconnected feature evolution},
  author       = {Yu, Piao and Zhang, Xu and Gong, Yongshun and Zhang, Jian and Sun, Haoliang and Zhang, Junjie and Zhang, Xinxin and Yin, Yilong},
  journal      = {Expert Systems with Applications},
  year         = {2025},
  volume       = {264},
  pages        = {125679},
  doi          = {10.1016/j.eswa.2024.125679},
  issn         = {0957-4174},
  url          = {https://doi.org/10.1016/j.eswa.2024.125679}
}

Yu, P., Zhang, X., Gong, Y., Zhang, J., Sun, H., Zhang, J., Zhang, X., & Yin, Y. (2025). Enhancing origin–destination flow prediction via bi-directional spatio-temporal inference and interconnected feature evolution. Expert Systems with Applications, 264, Article 125679. https://doi.org/10.1016/j.eswa.2024.125679