Matching novelty while training: Novel recommendation based on personalized pairwise loss weighting

Most works of recommender system seek to provide highly accurate item prediction while having potentially great bias to popular items. Both users and items' providers will suffer if their system has strong preference for monotonous popular items. A better system should consider also item novelty. Previous works of novel recommendation focus mainly on re-ranking a top-N list generated by an accuracy-focused base model. As a result, these frameworks are 2-stage and essentially limited to the base model. In addition, when training the base model, the common BRP loss function treats all pairs in the same manner, consistently suppresses interesting negative items which should have been recommended. In this work, we propose a personalized pairwise novelty weighting for BPR loss function, which covers the limitations of BPR and effectively improves novelty with marginal loss in accuracy. Base model will be guided by the loss weights to learn user preference and to generate novel suggestion list in 1 stage. Comprehensive experiments on 3 public datasets show that our approach effectively promotes novelty with almost no decrease in accuracy.