Pre-training Enhanced Spatial-temporal Graph Neural Network for Multivariate Time Series Forecasting (2206.09113v2)

Abstract: Multivariate Time Series (MTS) forecasting plays a vital role in a wide range of applications. Recently, Spatial-Temporal Graph Neural Networks (STGNNs) have become increasingly popular MTS forecasting methods. STGNNs jointly model the spatial and temporal patterns of MTS through graph neural networks and sequential models, significantly improving the prediction accuracy. But limited by model complexity, most STGNNs only consider short-term historical MTS data, such as data over the past one hour. However, the patterns of time series and the dependencies between them (i.e., the temporal and spatial patterns) need to be analyzed based on long-term historical MTS data. To address this issue, we propose a novel framework, in which STGNN is Enhanced by a scalable time series Pre-training model (STEP). Specifically, we design a pre-training model to efficiently learn temporal patterns from very long-term history time series (e.g., the past two weeks) and generate segment-level representations. These representations provide contextual information for short-term time series input to STGNNs and facilitate modeling dependencies between time series. Experiments on three public real-world datasets demonstrate that our framework is capable of significantly enhancing downstream STGNNs, and our pre-training model aptly captures temporal patterns.

• The paper proposes STEP, a framework that enhances Spatial-temporal Graph Neural Networks for multivariate time series forecasting by effectively incorporating long-term historical data through pre-training.
• The core of STEP is TSFormer, a novel pre-training model using masked autoencoding and a lightweight Transformer to efficiently capture rich representations from extended time series sequences.
• Empirical validation shows STEP significantly outperforms state-of-the-art baselines on benchmark datasets like METR-LA and PEMS-BAY, demonstrating improved accuracy in various forecasting horizons.

An Evaluation of the Pre-training Enhanced Spatial-temporal Graph Neural Network for Multivariate Time Series Forecasting

The paper "Pre-training Enhanced Spatial-temporal Graph Neural Network for Multivariate Time Series Forecasting," addresses significant limitations in current Spatial-Temporal Graph Neural Networks (STGNNs) by proposing a framework named STEP, which stands for a multiscale pre-training enhanced graph neural approach. This research primarily focuses on improving multivariate time series forecasting performance by employing a novel framework that combines long-term pre-training with spatial-temporal modeling.

Multivariate time series (MTS) prediction is crucial across various domains, from transportation networks to energy consumption analysis. Despite the impressive performance of traditional STGNNs, which leverage both spatial and temporal patterns through graph neural networks and sequential models, existing models predominantly emphasize short-term historical data due to computational constraints. This restriction limits the capacity of such models to capture long-term dependencies and comprehensive temporal and spatial patterns, which are crucial for accurate MTS forecasting.

To address this limitation, the authors introduce a pre-training model that exploits long-term MTS data, proposing a method called STEP. STEP enhances existing STGNN models by capturing extensive historical information through a scalable time series pre-training technique. This pre-training model, named TSFormer, uses a masked autoencoding strategy with Transformer blocks, focusing on extending learning from very long-term histories like the past two weeks. The aim is to develop rich segment-level representations that incorporate context for short-term predictions.

The paper makes several substantial contributions to improving MTS forecasting:

1. Framework Development: The STEP framework acts as an adjunct to every STGNN by providing extensive historical context. By enhancing STGNNs with learned pre-training weights from TSFormer, the framework successfully improves model performance beyond the capability of short-term input alone.
2. Innovative Pre-training Model: TSFormer, the pre-training component, is particularly noteworthy. It utilizes a high masking ratio strategy in line with recent advances in vision models and is built on lightweight Transformer architecture. These design decisions enable efficient learning from extended-length sequences, while the masked autoencoding encourages capturing higher-level semantic representations necessary for accurate forecasting.
3. Graph Structure Learning: A notable aspect is the adaptive graph learner integrated within STEP, which corrects or generates dependency graphs lacking in predefined scenarios. This component leverages the segment-level representations to guide accurate dependency modeling.
4. Empirical Validation: The authors present strong empirical results demonstrating STEP's efficacy. On datasets such as METR-LA, PEMS-BAY, and PEMS04, STEP consistently outperforms advanced baselines, showcasing significant improvements in accuracy metrics like MAE, RMSE, and MAPE across various forecasting horizons.
5. Design Insights: The authors have also explored detailed introspections on positional embeddings and representation similarities. Such insights underscore how TSFormer internalizes periodic patterns and long-term dependencies from the multivariate time series, facilitating enhanced downstream ML tasks.

In conclusion, the paper establishes STEP as a prominent advancement in the domain of MTS forecasting by overcoming the limitations of traditional STGNNs. This work not only achieves superior empirical results but also opens avenues for further research in extensive sequence learning for graph-based temporal models. Future work may explore extending the TSFormer modeling capabilities or integrate similar pre-training enhancements in other domains outside MTS forecasting.