TRECK: Long-Term Traffic Forecasting With Contrastive Representation Learning

Recent research mainly applies deep learning (DL) methods to short-term traffic forecasting. However, there is a growing interest in long-term forecasting, which allows action optimization at more steps in the future. Motivated by the encouraging success of contrastive representation learning, we propose a powerful and light framework, namely, Traffic Representation Extraction with Contrastive learning frameworK (TRECK), to improve traffic forecasting performance, especially for longer prediction terms. TRECK i) learns disentangled seasonal representations with contrastive learning, ii) enhances the learning of event data with entity embedding and iii) improves generalization and encourages obtaining more effective representations for the forecasting task through multi-task learning. TRECK can be directly applied to typical sequence-to-sequence DL prediction models. We evaluate TRECK when integrated with vanilla base models (RNN and BiLSTM) on large-size and real-world datasets. Experimental results show that TRECK can considerably boost the performance of base models and offer them the capability of handling increasing forecasting horizons. With TRECK, even a naive model like RNN can outperform state-of-the-art Transformer-based and GNN-based methods. Moreover, while avoiding any laborious feature design, the representations extracted by TRECK are more desirable than hand-crafted time features, yielding an 18.69% lower average MAE. Further analysis highlights its efficacy in diverse traffic conditions and in generating prediction intervals.