Xiaojun Cheng
Research Subjects
Xiaojun Cheng has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Bo Peng
Department of Economics and Management, Bozhou University, Bozhou, China
1 shared publication - Dong Jiang
Zhejiang Gongshang University
1 shared publication - Kai Yuan
Bozhou University, Bozhou, Anhui, China
1 shared publication - Qinyi Long
Zhejiang Gongshang University
1 shared publication - Zhengsong Xu
Department of Economics and Management, Bozhou University, Bozhou, China
1 shared publication
Latest Publications
- Open Access
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
MAPatchTST: Local-Structure-Aware Patch Encoding and Graph-Clustering-Guided Hybrid Channel Modeling for Multivariate Time Series Forecasting
Patch-based Transformers provide an effective framework for long-horizon multivariate time series forecasting by reducing the effective sequence length and preserving local continuity. However, conventional patch tokenization usually projects each segment as a static block, which may weaken trend transitions and short-term deviations within the patch. Channel-independent forecasting reduces cross-variable interference but discards useful dependencies, whereas fully channel-dependent modeling can propagate weak correlations and noise across variables. To address these limitations, this paper proposes MAPatchTST, which jointly improves fixed-scale patch representation and channel interaction. A mean-adaptive patch encoder first decomposes each patch into a running local baseline and a deviation component. Learnable temporal decay, deviation modulation, and intra-patch attention are then applied to preserve local changes before global temporal encoding. Based on the enhanced patch representations, a sparse association graph is constructed to generate task-oriented channel embeddings. The channels are subsequently clustered according to their propagated representations. Dependency modeling is performed within each cluster, while cross-cluster information exchange is restricted to reduce noise propagation. A residual prediction objective with regularization is further adopted to stabilize model training under non-stationary conditions. Experiments on ETTh1, ETTh2, ETTm1, ETTm2, Electricity, and Traffic show that MAPatchTST consistently improves the direct PatchTST baseline, reducing average MSE and MAE by approximately 3.55% and 4.05%, respectively. Ablation and sensitivity analyses further demonstrate the distinct contributions of local patch enhancement, sparse channel organization, and cluster-wise interaction.
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMAPatchTST: Local-Structure-Aware Patch Encoding and Graph-Clustering-Guided Hybrid Channel Modeling for Multivariate Time Series Forecasting
Patch-based Transformers provide an effective framework for long-horizon multivariate time series forecasting by reducing the effective sequence length and preserving local continuity. However, conventional patch tokenization usually projects each segment as a static block, which may weaken trend transitions and short-term deviations within the patch. Channel-independent forecasting reduces cross-variable interference but discards useful dependencies, whereas fully channel-dependent modeling can propagate weak correlations and noise across variables. To address these limitations, this paper proposes MAPatchTST, which jointly improves fixed-scale patch representation and channel interaction. A mean-adaptive patch encoder first decomposes each patch into a running local baseline and a deviation component. Learnable temporal decay, deviation modulation, and intra-patch attention are then applied to preserve local changes before global temporal encoding. Based on the enhanced patch representations, a sparse association graph is constructed to generate task-oriented channel embeddings. The channels are subsequently clustered according to their propagated representations. Dependency modeling is performed within each cluster, while cross-cluster information exchange is restricted to reduce noise propagation. A residual prediction objective with regularization is further adopted to stabilize model training under non-stationary conditions. Experiments on ETTh1, ETTh2, ETTm1, ETTm2, Electricity, and Traffic show that MAPatchTST consistently improves the direct PatchTST baseline, reducing average MSE and MAE by approximately 3.55% and 4.05%, respectively. Ablation and sensitivity analyses further demonstrate the distinct contributions of local patch enhancement, sparse channel organization, and cluster-wise interaction.