Graph-Guided Latent Variable Target Inference for Mitigating Concept Drift in Time Series Forecasting

With the proliferation of the Internet of Things (IoT), there is an abundance of data available to humans. However, the monitoring environments for data collection are becoming increasingly diverse, leading to the occurrence of concept drift in the collected data. Concept drift refers to the phenomenon where the distribution of data changes over time, making it challenging for prediction models trained on historical data to adapt to the changing distribution. Previous research has primarily focused on predicting or compensating for distributions with fixed durations in Euclidean space to mitigate non-stationarity. However, we have observed that concept drift often occurs at different time scales, and detecting them using fixed scales has inherent limitations. Based on this observation, we propose a Graph-Guided Latent Variable Target Inference network that maps current data and variable duration query targets onto a graph neural network in latent space. We apply self-attention transformations to the representations and correlations on the graph in the dimensions of time, features, and query targets. The model updates its parameters based on these non-Euclidean correlation patterns, enabling the graph to evolve towards the direction of the query targets and obtain an evolved latent distribution. Finally, the decoder generates a prediction data stream regarding the query targets based on the evolved latent distribution. The experiments were conducted on five datasets, where our proposed method was compared against the five most advanced baselines. The findings demonstrated a substantial advantage in prediction performance provided by our approach.