MC-STL: Mask- and Contrast-Enhanced Spatio-Temporal Learning for Urban Flow Prediction

Xu Zhang, Yongshun Gong, Xinxin Zhang, Xiaoming Wu, Chengqi Zhang, Xiangjun Dong

CIKM 2023, pp. 3298–3307 · Free to read in the ACM Digital Library

Also indexed in: Semantic Scholar · OpenAlex · dblp · Wikidata · ACM Digital Library

Quick facts

Main results

Quoted text is verbatim from the paper.

Key points

• Spatio-temporal random masking hides different regions at different timestamps. On TaxiBJ it gives RMSE 14.53 / MAE 8.70, against 14.91 / 8.86 with conventional image random masking.
• A graph processing module (GPM) sums the self-attention weights of the mask-pre-trained ViT encoder over heads and batch and uses them as the adjacency matrix of a two-layer GCN.
• The contrastive pretext task treats each temporally ordered batch as a unit: a temporally shuffled anchor is the negative, a unit from another period is the positive, and triplet loss trains the global-local cross-attention encoder.
• Against nine baselines, MC-STL has the lowest RMSE and MAE on all six evaluation sets (TaxiBJ P1-P4, full TaxiBJ, BikeNYC); on full TaxiBJ it reaches RMSE 14.53 versus 14.72 for ST-GSP.
• On TaxiBJ every ablated variant is worse than the full model. Removing external factors raises RMSE by 2.48 to 17.01; removing all pre-training raises it only to 14.65.

Abstract

As a critical mission of intelligent transportation systems, urban flow prediction (UFP) benefits in many city services including trip planning, congestion control, and public safety. Despite the achievements of previous studies, limited efforts have been observed on simultaneous investigation of the heterogeneity in both space and time aspects. That is, regional correlations would be variable at different timestamps. In this paper, we propose a spatio-temporal learning framework with mask and contrast enhancements to capture spatio-temporal variabilities among city regions. We devise a mask-enhanced pre-training task to learn latent correlations across the spatial and temporal dimensions, and then a graph-based method is developed to extract the significance of regions by using the inter-regional attention weights. To further acquire contrastive correlations of regions, we elaborate a pre-trained contrastive learning task with the global-local cross-attention mechanism. Thereafter, two well-trained encoders have strong capability to capture latent spatio-temporal representations for the flow forecasting with time-varying. Extensive experiments conducted on real-world urban flow datasets demonstrate that our method compares favorably with other state-of-the-art models.

Frequently asked questions

Is MCSTL the same as MC-STL?

Yes. The CIKM 2023 paper names the model MC-STL (mask- and contrast-enhanced spatio-temporal learning). Its official code repository is github.com/CodeZx6/MCSTL.

What problem does MC-STL address?

The paper notes that most existing urban flow models do not consider spatial and temporal heterogeneity simultaneously, although correlations between regions vary across timestamps. MC-STL uses two pre-training tasks, mask-enhanced and contrast-enhanced, to learn these time-varying spatio-temporal correlations.

How does MC-STL's masking differ from image-style random masking?

Image random masking hides a region at every time step, which the paper argues captures spatial but not temporal relationships between regions. MC-STL masks different regions at different timestamps and reconstructs them from the remaining regions. On TaxiBJ this gives RMSE 14.53 / MAE 8.70, against 14.91 / 8.86 with image random masking.

How does MC-STL build its graph?

A graph processing module (GPM) sums the multi-head self-attention weights of the mask-pre-trained ViT encoder over heads and over the batch, then thresholds and normalises them. The result is the adjacency matrix of a two-layer GCN that extracts important regions.

Which datasets, metrics and baselines does MC-STL use?

MC-STL is evaluated on TaxiBJ (periods P1-P4 and the full set) and BikeNYC, six evaluation sets in total, with RMSE and MAE. The nine baselines are HA, RNN, ARIMA, ST-ResNet, AGCRN, Multi-STGCnet, ATFM, ST-SSL and ST-GSP; ST-SSL is reported only on BikeNYC and TaxiBJ (P3).

How well does MC-STL perform?

In Table 2, MC-STL has the lowest RMSE and MAE on all six evaluation sets. On full TaxiBJ it reaches RMSE 14.53 / MAE 8.70 against 14.72 / 8.79 for ST-GSP, and on BikeNYC 4.81 / 2.38 against 5.52 / 2.47 for Multi-STGCnet.

How do I cite MC-STL?

Cite it as: Zhang, X., Gong, Y., Zhang, X., Wu, X., Zhang, C., & Dong, X. (2023). Mask- and Contrast-Enhanced Spatio-Temporal Learning for Urban Flow Prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (pp. 3298–3307). ACM. https://doi.org/10.1145/3583780.3614958 BibTeX is available at https://codezx6.github.io/papers/mcstl.html#cite and in https://codezx6.github.io/publications.bib.

Is the code for MC-STL available?

Yes, at https://github.com/CodeZx6/MCSTL.

Keywords

urban flow prediction spatio-temporal predictive modeling spatio-temporal pre-training mask-enhanced learning contrastive learning traffic prediction graph convolutional network (GCN)

Also referred to as: MCSTL; Mask- and Contrast-Enhanced Spatio-Temporal Learning; mask-enhanced spatio-temporal learning; spatio-temporal (ST) random masking; graph processing module (GPM); global-local spatio-temporal cross-attention.

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 › crowd flow prediction › urban flow prediction › self-supervised spatio-temporal pre-training › MC-STL

Task

urban flow prediction crowd flow prediction citywide crowd flow forecasting inflow and outflow prediction grid-based traffic flow forecasting spatio-temporal forecasting

Method

masked spatio-temporal pre-training self-supervised pre-training for traffic prediction spatio-temporal contrastive learning Vision Transformer (ViT) encoder attention-derived graph GCN global-local cross-attention

Datasets

TaxiBJ BikeNYC

Compared with

ST-ResNet AGCRN ATFM ST-SSL ST-GSP Multi-STGCnet

中文

城市流量预测 人群流量预测 交通流量预测 流入流出预测 时空预测 掩码预训练 时空对比学习 自监督预训练

Related areas

traffic forecasting spatio-temporal graph neural networks graph neural networks Transformers for time series masked autoencoders self-supervised learning contrastive learning representation learning mobility modeling human mobility spatio-temporal heterogeneity citywide prediction grid-based prediction deep spatio-temporal residual networks

Applications

traffic management congestion control trip planning public safety ride-hailing dispatch bike-sharing rebalancing

Related benchmarks (not used in this paper)

TaxiNYC BikeDC BikeChicago PEMS-BAY METR-LA PeMS04 PeMS08

中文 (扩展)

交通预测 时空数据挖掘 时间序列预测 智能交通 智慧城市 城市计算 时空图神经网络 图神经网络 掩码自编码器 自监督学习 表征学习 人类移动性 网格流量预测 交通管理 共享单车调度

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

Urban flow prediction

• MR-UFP: Enhancing urban flow prediction via mutual reinforcement with multi-scale regional information (Neural Networks, vol. 182, article 106900, 2025)
• ST-FCL: Spatio-temporal fusion and contrastive learning for urban flow prediction (Knowledge-Based Systems, vol. 282, article 111104, 2023)

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)
• DSTCN: Exploiting dynamic spatio-temporal correlations for origin-destination demand prediction (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)
• BiST-IF: Enhancing origin–destination flow prediction via bi-directional spatio-temporal inference and interconnected feature evolution (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)

Cite this paper

@inproceedings{zhang2023mcstl,
  title        = {Mask- and Contrast-Enhanced Spatio-Temporal Learning for Urban Flow Prediction},
  author       = {Zhang, Xu and Gong, Yongshun and Zhang, Xinxin and Wu, Xiaoming and Zhang, Chengqi and Dong, Xiangjun},
  booktitle    = {Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
  year         = {2023},
  series       = {CIKM '23},
  pages        = {3298--3307},
  publisher    = {ACM},
  doi          = {10.1145/3583780.3614958},
  url          = {https://doi.org/10.1145/3583780.3614958}
}

Zhang, X., Gong, Y., Zhang, X., Wu, X., Zhang, C., & Dong, X. (2023). Mask- and Contrast-Enhanced Spatio-Temporal Learning for Urban Flow Prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (pp. 3298–3307). ACM. https://doi.org/10.1145/3583780.3614958