Incremental Evaluation and Training in Relational Deep Learning

2026-08-13

AI Summary

This work addresses the limitations of existing relational deep learning models, which are typically evaluated on static data snapshots and thus fail to capture the dynamic evolution of databases over time or model performance under continuous data influx. To overcome this, the authors propose an incremental, multi-stage evaluation and training paradigm tailored for temporal dynamics. They model multi-table databases as temporal heterogeneous graphs and introduce a novel evaluation metric that prioritizes recent prediction accuracy. By integrating incremental fine-tuning, transfer learning, and multi-stage training, their approach enables dynamic training and assessment on large-scale real-world databases. Experimental results demonstrate that the proposed method significantly outperforms baseline models trained from scratch, effectively mitigates temporal concept drift, and substantially reduces training costs, thereby validating the feasibility and efficiency of transfer learning in relational deep learning.

Abstract

Relational Deep Learning (RDL) models multi-tabular databases as temporal heterogeneous graphs to enable end-to-end representation learning. However, prevailing RDL evaluation practices rely on static, single-episode dataset snapshots, overlooking the continuous, time-evolving nature of real-world databases. Consequently, current RDL benchmarks fail to capture how model performance changes as new data accumulates over time. To address this limitation, we introduce an incremental, multi-episode evaluation and training paradigm to assess and improve the temporal robustness and adaptability of state-of-the-art RDL models. Using established large-scale datasets, we examine data evolution and model training dynamics, demonstrating that temporal concept drifts occur in the majority of predictive tasks. We present multiple incremental training regimes for fine-tuning the models and demonstrate that transfer learning is both feasible and highly effective in the RDL setting. Alongside a new temporal evaluation metric that prioritizes near-future accuracy, we show that our incrementally fine-tuned models consistently outperform the standard, expensive, from-scratch trained baselines.

Problem

• Relational Deep Learning
• incremental evaluation
• temporal robustness
• concept drift
• multi-episode training

Innovation

• Incremental Learning
• Relational Deep Learning
• Temporal Concept Drift
• Transfer Learning
• Multi-episode Evaluation