Multi-supervisor association network cold start recommendation based on meta-learning

Xiaoyang Liu, Ziyang Zhang, Xiaoqin Zhang, Pasquale De Meo, Hocine Cherifi

Abstract: We present a novel meta-learning based recommender system, called FO-MSAN (First-Order Meta-Learning and Multi-Supervisor Association Network), to generate high quality recommendations in the cold-start scenario. FO-MSAN advances the state of the art in two major aspects: a) it uses a gradient update strategy that replaces the calculation of second derivatives with the calculation of first derivatives, which has clear computational advantages, and b) it uses a multi-supervisor association network that improves its ability to represent the characteristics of users and how users interact with the products offered by the platform.We tested FO-MSAN on three real-world datasets (MovieLens-100K, Bookcrossing, and Yelp2018), and it showed significant improvements in both score prediction and ranking tasks compared to popular baselines. Specifically, FO-MSAN achieved at least a 2.47 % improvement in MAE for score prediction and over 1 % improvement in NDCG@5 and Precision@5 for ranking tasks.