Overcoming catastrophic forgetting in tabular data classification: A pseudorehearsal-based approach
Pablo García-Santaclara, Bruno Fernández-Castro, Rebeca P. Dı́az Redondo · Engineering Applications of Artificial Intelligence · 2025
Continual learning (CL) poses the important challenge of adapting to evolving data distributions without forgetting previously acquired knowledge while consolidating new knowledge. In this paper, we introduce a new methodology, coined the Tabular-data Rehearsal-based Incremental Lifelong Learning framework (TRIL3), designed to address the phenomenon of catastrophic forgetting in an online, task-free setting. TRIL3 utilizes an Incremental Learning Vector Quantization (ILVQ) algorithm as an efficient prototype-based incremental generative model to store and generate synthetic data to preserve knowledge over time, and the Deep Neural Decision Forest (DNDF) algorithm, which was modified to run incrementally to learn supervised classification tasks for tabular data. Based on tests conducted to determine the optimal percentage of synthetic data and comparisons with other available task-free CL proposals, we conclude that TRIL3 outperforms other methods in the literature using only 50% of synthetic data.