Lifelong-Learning Embeddings: Incremental and Continual Representation Learning for Dynamic E-Commerce Trends

ICLR 2026 Conference Submission17547 Authors

Keywords: Embeddings, Representation learning, Incremental Learning, Continual Learning, E-Commerce

TL;DR: Lifelong-Learning Embeddings (LLE) dynamically expand vocabularies, adapt embedding dimensionality, and apply continual learning to preserve past knowledge while integrating new data.

Abstract:

E-commerce is a highly dynamic domain where products and consumer behaviors evolve rapidly. Embedding-based representations are central to deep learning–based personalization systems, yet conventional embeddings are static and therefore, they cannot easily incorporate new tokens (e.g., new products) without retraining, which is costly and often infeasible due to privacy or data retention constraints. To address this, we propose Lifelong-Learning Embeddings, a framework that (1) incrementally extends embeddings to integrate new tokens, (2) adapts embedding dimensionality to balance expressiveness and efficiency, and
(3) employs continual learning to mitigate catastrophic forgetting. Experiments on a real-world dataset and two benchmark datasets show that our approach consistently outperforms static embeddings in accuracy while incurring only modest training-time overhead, demonstrating its effectiveness and adaptability in dynamic e-commerce environments.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 17547