PLAE: Time-Series Prediction Improvement by Adaptive Decomposition

Univariate time-series forecasting is a kind of commonly encountered yet tough problem. Most of the forecast algorithms' performance is constrained by the limited information due to the single input dimension. No matter how capable a forecast algorithm is, an accurate output cannot be rendered on an unpredictable time-series. This paper presents PLAE (Predictability Leveraging Auto-Encoder), a Seq2Seq model for univariate time-series data aiming to enhance the accuracy of the given algorithm without dimensional adaptation. The main idea is decomposing the original input data into a group of more predictable microscopic time-series on which the forecast algorithm can deliver a more accurate output. And the final prediction is rendered by aggregating those components back to the original one-dimension. Experiments on three public data sets and one real-world data set show that PLAE can improve the forecast accuracy for 23.38% in terms of MAPE and 19.76% in terms of RMSE. Besides, experimental evidence shows that PLAE's adaptive non-linear decomposition mechanism outperforms the pre-defined additive decomposition w.r.t. both forecasting performance and components' interpretability.