Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting

Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, Yongjun Xu

Introduction

Multivariate Time Series (MTS) forecasting is a critical task in modern infrastructure management, ranging from predicting traffic flow across urban sensor networks to monitoring electricity consumption in power grids. In recent years, the field has been dominated by increasingly complex Spatial-Temporal Graph Neural Networks (STGNNs). These models utilize sophisticated graph convolutions to capture the intricate dependencies between different variables. However, this complexity comes at a high computational cost. This paper introduces STID (Spatial-Temporal IDentity), a simple yet highly effective MLP-based baseline that challenges the necessity of complex graph operations.

The Indistinguishability Problem

The authors identify a fundamental challenge in MTS forecasting termed indistinguishability. This occurs in two primary forms:

  1. Spatial Indistinguishability: Two distinct sensors may record nearly identical historical patterns but exhibit vastly different future outcomes due to their specific locations or surrounding environments.
  2. Temporal Indistinguishability: A single sensor might record similar data at different times (e.g., a Monday morning vs. a Sunday morning), yet the underlying dynamics of the system lead to different future trajectories.

While STGNNs attempt to solve these issues through complex message-passing and attention mechanisms, the authors hypothesize that the performance gains are often due to the models implicitly learning context rather than the complexity of the graph operations themselves.

The STID Framework

STID proposes a streamlined architecture based on Multi-Layer Perceptrons (MLPs) that explicitly incorporates context through learned identity embeddings. The model follows a four-stage process:

  1. Feature Embedding: Raw historical data is projected into a hidden dimension via a fully connected layer.
  2. Identity Attachment: This is the core innovation of the model. To resolve indistinguishability, the model concatenates the feature embeddings with three trainable matrices:
  3. Spatial Identity ($E$): A unique, learnable vector for each sensor or variable.
  4. Time-of-Day Identity ($T^{TiD}$): Embeddings representing specific time slots (e.g., 5-minute intervals).
  5. Day-of-Week Identity ($T^{DiW}$): Embeddings that distinguish between different days of the week.
  6. MLP Encoder: The combined representation (features plus identities) is processed through a series of MLP layers with residual connections.
  7. Regression: A final linear layer generates the multi-step forecast.

By explicitly providing the model with "where" and "when" information, STID eliminates the need for the model to infer this context from raw data alone.

Experimental Evaluation

STID was evaluated against several state-of-the-art (SOTA) models, including Graph WaveNet, MTGNN, and GMAN, across five standard benchmarks:

  • Traffic Flow: PEMS04, PEMS07, and PEMS08.
  • Traffic Speed: PEMS-BAY.
  • Energy Consumption: Electricity (336 clients).

Notably, the Electricity dataset lacks a pre-defined graph structure. While many STGNNs require complex graph-learning modules to function on such data, STID handles it natively because it does not rely on graph convolutions.

Key Results and Insights

Superior Accuracy and Efficiency

STID achieved the lowest Mean Absolute Error (MAE) across nearly all forecasting horizons (3, 6, and 12 steps ahead) on all five datasets. Beyond accuracy, the efficiency gains were substantial. On the PEMS07 dataset, STID required only 14 seconds per epoch, compared to 827 seconds for the GMAN model—a speedup of approximately 60x. This efficiency is achieved with a remarkably small footprint: just 3 MLP layers and a hidden dimension of 32.

The Importance of Spatial Identity

Ablation studies revealed that the Spatial Identity embedding is the most critical component. Removing it led to the most significant degradation in performance, confirming that knowing the specific identity of a sensor is more important than modeling complex inter-sensor relationships for many standard benchmarks.

Learned Periodicity

Visualizations of the temporal embeddings showed that STID naturally learns the cyclical nature of time. The model grouped weekdays together and separated them from weekends without any hardcoded calendar rules, demonstrating that simple embeddings can capture complex temporal patterns.

Conclusion

STID serves as a powerful reminder that in architecture design, "less is more" if the data representation is sufficiently informative. By addressing the indistinguishability problem through simple identity embeddings, STID outperforms complex STGNNs while offering dramatic improvements in training speed. Future work may explore how to adapt these static identities for dynamic systems where relationships between variables change over time, or for open systems where the number of sensors is not fixed.

Summary updated

BibTeX
@inproceedings{shao2022spatial,
  author = {Zezhi Shao and Zhao Zhang and Fei Wang and Wei Wei and Yongjun Xu},
  title = {{Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2022},
  doi = {10.1145/3511808.3557702},
}
Figure