Most works of recommender systems focus on providing users with highly accurate item predictions based on the assumption that accurate suggestions can best satisfy users. However, accuracy-focused models also create great system bias towards popular items and, as a result, unpopular items rarely get recommended and will stay as “cold items” forever. Both users and item providers will suffer in such scenario. To promote item novelty, which plays a crucial role in system robustness and diversity, previous studies focus mainly on re-ranking a top-N list generated by an accuracy-focused base model. The re-ranking algorithm is thus completely independent of the base model. Eventually, these frameworks are essentially limited by the base model and the separated 2 stages cause greater complication and inefficiency in providing novel suggestions. In this work, we propose a personalized pairwise novelty weighting framework for BPR loss function, which covers the limitations of BPR and effectively improves novelty with negligible decrease in accuracy. Base model will be guided by the novelty-aware loss weights to learn user preference and to generate novel top-N list in only 1 stage. Comprehensive experiments on 3 public datasets show that our approach effectively promotes novelty with almost no decrease in accuracy.

推荐系统的大多数工作都基于准确的建议可以最好地满足用户的假设，为用户提供高度准确的项目预测。但是，以准确性为中心的模型也会对热门商品产生很大的系统偏见，结果，不受欢迎的商品很少会被推荐，并且将永远保留为“冷商品”。在这种情况下，用户和项目提供者都会受苦。为了促进在系统健壮性和多样性中起关键作用的项目新颖性，以前的研究主要集中在对由以准确性为中心的基本模型生成的前N个列表进行重新排序。因此，重新排序算法完全独立于基本模型。最终，这些框架本质上受到基本模型的限制，而分离的2个阶段导致提供新建议的过程更加复杂，效率低下。在这项工作中，我们提出了针对BPR损失函数的个性化成对新颖性加权框架，该框架涵盖了BPR的局限性，并有效地提高了新颖性，而准确性降低幅度可忽略不计。基本模型将以新颖性感知损失权重为指导，以仅在1个阶段中学习用户偏好并生成新颖的top-N列表。对3个公共数据集的综合实验表明，我们的方法有效地促进了新颖性，而准确性几乎没有降低。