Mlinear: Rethink the Linear Model for Time-series Forecasting

• DOI: 10.48550/arxiv.2305.04800
• Publication Date: 2023-01-01
• Authors: Jianing Chen; Chuhao Chen; Xiangxu Meng
• Keywords: Independence; Linear prediction

Recently, significant advancements have been made in time-series forecasting research, with an increasing focus on analyzing the nature of data, e.g, channel-independence (CI) and channel-dependence (CD), rather than solely focusing designing sophisticated models. However, current research has primarily focused either CI or CD isolation, challenge effectively combining these two opposing properties to achieve a synergistic effect remains unresolved issue. In this paper, we carefully examine CD, raise practical question that not answered, e.g.,"How mix time series better predictive performance?" To answer question, propose Mlinear (MIX-Linear), simple yet effective method based mainly linear layers. The design philosophy includes aspects:(1) dynamically tuning semantics different input series, (2) providing deep supervision adjust individual performance "CI predictor" "CD predictor". addition, empirically, introduce new loss function significantly outperforms widely used mean squared error (MSE) multiple datasets. Experiments datasets covering fields demonstrated superiority our over PatchTST which is lateset Transformer-based terms MSE MAE metrics 7 identical sequence inputs (336 512). Specifically, ratio 21:3 at 336 length 29:10 512 input. Additionally, approach 10 $\times$ efficiency advantage unit level, taking into account both training inference times.