Reproduce “Task Relation-aware Continual User Representation Learning” on OpenResearch

About this paper

Sein Kim, Namkyeong Lee, Donghyun Kim, Minchul Yang, Chanyoung Park

KAIST and NAVER Corporation researchers developed TERACON, a continual learning framework for universal user representation that employs task embedding-guided soft masks and a relation-aware knowledge retention module. It efficiently learns from a sequential stream of tasks, preventing catastrophic forgetting while allowing all model parameters to update, and demonstrably improves performance over baselines across diverse tasks including recommendations and user profiling.

Key insights

• Soft masking of layer outputs, guided by task embeddings, allows for continuous parameter updates across tasks, maintaining the model's learning capacity indefinitely and enabling positive backward knowledge transfer.
• Explicitly modeling task relationships within the mask generation process, including 'opposite relatedness,' helps the model differentiate between beneficial knowledge transfer and detrimental interference between tasks.
• An adaptive, relation-aware user sampling approach for knowledge retention significantly improves training efficiency, making continual learning more practical for large-scale industrial applications with extensive user data.

Summary by alphaXiv.