Incremental Training Boosts Relational Deep Learning Model Adaptability

Relational Deep Learning (RDL) models are designed to process multi-tabular databases by representing them as temporal heterogeneous graphs, enabling end-to-end representation learning. However, current evaluation methods for RDL typically rely on static snapshots of datasets, failing to account for the dynamic, time-evolving nature of real-world databases. This oversight means existing benchmarks do not accurately reflect how model performance changes as new data accumulates over time. To address this critical gap, researchers propose a new incremental, multi-episode evaluation and training approach. This paradigm is designed to assess and enhance the temporal robustness and adaptability of state-of-the-art RDL models. Through experiments on large-scale datasets, the study reveals that temporal concept drifts are prevalent in most predictive tasks. The research introduces several incremental training regimes for fine-tuning models and demonstrates that transfer learning is highly effective within the RDL context. By incorporating a new temporal evaluation metric focused on near-future accuracy, the study shows that incrementally fine-tuned models consistently outperform expensive, from-scratch trained baselines.