Continual Learning Based on Task Masking for Multi-domain Recommendation

Tran-Ngoc-Linh Nguyen, Chi-Dung Vu, Hoang-Ngan Le, Anh-Dung Hoang, Xuan-Hieu Phan, Quang-Thuy Ha, Hoang-Quynh Le, Mai-Vu Tran

Abstract: In recent years, Multi-Domain Recommendation has become a big challenge of recommendation systems, due to the large number of overlapped users and items between multiple domains of many platforms. Multi-Domain Recommendation focuses on capturing informative domain specific features from all domains to improve the corresponding accuracy. In this paper, we proposed a light model, inspired by technique used in Continual Learning selecting the important parameters of users preference for each domain that highly reduces the bias. Our model could also be applied on top of any existing latent model effectively making them usable in multi-domain recommendation settings. We call our architecture as CL4Rec.