Continual Transfer Learning for Cross-Domain Click-Through Rate Prediction at Taobao

Lixin Liu, Yanling Wang, Tianming Wang, Dong Guan, Jiawei Wu, Jingxu Chen, Rong Xiao, Wenxiang Zhu, Fei Fang

Abstract: We study the problem of cross-domain click-through rate (CTR) prediction for recommendation at Taobao. Cross-domain CTR prediction has been widely studied in recent years, while most attempts ignore the continual learning setting in industrial recommender systems. In light of this, we present a necessary but less-studied problem named Continual Transfer Learning (CTL), which transfers knowledge from a time-evolving source domain to a time-evolving target domain. We propose an effective and efficient model called CTNet to perform CTR prediction under the CTL setting. The core idea behind CTNet is to treat source domain representations as external knowledge for target domain CTR prediction, such that the continually well-trained source and target domain parameters can be preserved and reused during knowledge transfer. Extensive offline experiments and online A/B testing at Taobao demonstrate the efficiency and effectiveness of CTNet. CTNet is now fully deployed online at Taobao bringing significant improvements.