Causal Cross-embedded Spatio-temporal LSTM for Web Traffic Prediction

Web service traffic forecasting is vital for dynamic resource scaling, load balancing, and anomaly detection, but remains challenging due to frequent large-scale fluctuations caused by heterogeneous user behaviors. Traditional time-series models and recent deep neural networks have made progress by capturing temporal patterns, yet they largely overlook latent causal relationships between services that can significantly influence traffic dynamics. In this paper, we propose a novel causal cross-embedded spatio-temporal LSTM (CEST-LSTM) architecture that integrates spatio-temporal modelling with a causal inference mechanism to improve web traffic prediction. The model consists of a spatio-temporal LSTM branch for capturing temporal dependencies across services and a causal branch that leverages convergent cross mapping-based cross-embedding to uncover and incorporate latent inter-service causal influences. A cross-embedding fusion mechanism seamlessly combines these causal features with spatio-temporal representations. On real-world datasets (e.g., Microsoft Azure and Alibaba Cloud), CEST-LSTM achieves a variance-explained prediction accuracy of approximately 93%, surpassing state-of-the-art baselines such as temporal graph convolutional networks (T-GCN) and spatio-temporal attention GCNs (STA-GCN). Comparative experiments and ablation studies confirm that the causal branch consistently improves forecasting accuracy – for example, removing the causal module reduces accuracy by several percentage points. These results demonstrate that integrating latent causal relationship modelling into spatio-temporal neural networks yields substantial improvements in web traffic prediction, offering a promising direction for robust and interpretable forecasting in complex web systems.