TotalRecall: A Bidirectional Candidates Generation Framework for Large Scale Recommender \& Advertising Systems

Qifang Zhao, Yu Jiang, Yuqing Liu, Meng Du, Qinghui Sun, Chao Xu, huan xu, Zhongyao Wang

Keywords: Recommender System, Advertising System, Collaborative filtering, Matrix Factorization, Contrastive Learning, Candidates Generation, Embeddings

Abstract: Recommender (RS) and Advertising/Marketing Systems (AS) play the key roles in E-commerce companies like Amazaon and Alibaba. RS needs to generate thousands of item candidates for each user ($u2i$), while AS needs to identify thousands or even millions of high-potential users for given items so that the merchant can advertise these items efficiently with limited budget ($i2u$). This paper proposes an elegant bidirectional candidates generation framework that can serve both purposes all together. Besides, our framework is also superior in these aspects: $i).$ Our framework can easily incorporate many DNN-architectures of RS ($u2i$), and increase the HitRate and Recall by a large margin. $ii).$ We archive much better results in $i2u$ candidates generation compare to strong baselines. $iii).$ We empirically show that our framework can diversify the generated candidates, and ensure fast convergence to better results.