Evaluating Tabular Deep Learning under Temporal Shift and Extensive Feature Engineering

Evaluating Tabular Deep Learning under Temporal Shift and Extensive Feature Engineering

Rubachev I.V. (NRU HSE, Moscow, Russia)

Abstract

Progress in machine learning research is valuable when it transfers from academic benchmarks to practical applications. For tabular deep learning, this requires understanding how academic evaluation settings reflect real-world applications. In this paper, we analyze existing tabular benchmarks and identify two industrially common but underrepresented properties: temporal distribution drift and feature spaces produced by extensive data acquisition and feature engineering pipelines. The first calls for time-based train/validation/test splits, while the second changes the number and structure of predictive, uninformative, and correlated features compared with typical academic datasets. To study how recent advances in tabular deep learning transfer to these conditions, we introduce TabReD, a collection of eight industry-grade tabular datasets with time-based splits and high-dimensional engineered feature spaces. We reassess a broad range of methods, including GBDTs, MLP-like neural networks, numerical feature embeddings, attention-based and retrieval-based models, advanced training recipes, and ensembles. The results show that time-based evaluation changes method rankings and relative performance compared with random splits; on TabReD, simple MLP-like architectures with numerical embeddings and GBDTs are the most robust, while several more complex recent methods transfer less reliably.

Keywords

tabular data; benchmark; temporal shift; feature engineering; deep learning.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 5, 2026, pp. 305-324

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(5)-17

For citation

Rubachev I.V. Evaluating Tabular Deep Learning under Temporal Shift and Extensive Feature Engineering. Proceedings of the Institute for System Programming, vol. 38, issue 5, 2026, pp. 305-324 DOI: 10.15514/ISPRAS-2026-38(5)-17.

Full text of the paper in pdf (in Russian) Back to the contents of the volume