Dissecting Advanced Time Series Forecasting Models with AIChronoLens

Abstract

Mobile traffic forecasting is instrumental in efficiently managing network resources. In this poster paper, we dissect the behavior of advanced time series forecasting techniques, namely DLinear and PatchTST, when applied to the problems of predicting future mobile traffic volumes. Being black-box models hard to interpret, we ground our analysis on EXplainable Artificial Intelligence (XAI) by using AIChronoLens, a new tool that links legacy XAI explanations with the temporal properties of the input sequences. We find that the DLinear significantly improves the prediction accuracy over PatchTST and state-of-the-art techniques like Long-Short Term Memory (LSTM). The analysis with AIChronoLens shows that, unlike PatchTST, DLinear is capable of focusing its prediction decisions on a few key samples of the input sequences, which makes it possible for DLinear to match the ground truth closely.