Multi-timescale representation with adaptive routing for deep tabular learning under temporal shift.

Abstract

In real-world applications, tabular datasets often evolve over time, leading to temporal shift that degrades the long-range neural network performance. Most existing temporal encoding or adaptation solutions treat time cues as fixed auxiliary variables at a single scale. Motivated by the multi-horizon nature of temporal shifts with heterogeneous temporal dynamics, this paper presents TARS (Temporal Abstraction with Routed Scales), a novel plug-and-play method for robust tabular learning under temporal shift, applicable to various deep learning model backbones. First, an explicit temporal encoder decomposes timestamps into short-term recency, mid-term periodicity, and long-term contextual embeddings with structured memory. Next, an implicit drift encoder tracks higher-order distributional statistics at the same aligned timescales, producing drift signals that reflect ongoing temporal dynamics. These signals drive a drift-aware routing mechanism that adaptively weights the explicit temporal pathways, emphasizing the most relevant timescales under current conditions. Finally, a feature-temporal fusion layer integrates the routed temporal representation with original features, injecting context-aware bias. Extensive experiments on eight real-world datasets from the TabReD benchmark show that TARS consistently outperforms the competitive compared methods across various backbone models, achieving up to +2.38% average relative improvement on MLP, +4.08% on DCNv2, etc. Ablation studies verify the complementary contributions of all four modules. These results highlight the effectiveness of TARS for improving the temporal robustness of existing deep tabular models.