A Proposal for Recommender System Based on Neural Collaborative Ranking Model with Multiple Implicit Feedback and Ordering Relationships

In current electronic commerce, it is desired to utilize implicit feedback, which is indirect evaluation data of user behavior. Neural Collaborative Ranking (NCR) is a type of recommendation model that estimates the most likely items to be observed next as implicit feedback in a given list. The model evaluates item preferences by assuming that items with implicit feedback by a user are preferred over items without such feedback. However, there are actual cases that multiple levels of implicit feedback can be observed such as browsing and purchasing histories. The NCR model cannot distinguish and learn these multiple levels of implicit feedback, and cannot fully utilize the information of multiple levels of implicit feedback. In this study, we propose an NCR model that considers multiple levels of implicit feedback by introducing the method proposed by Ding et al. Specifically, when training a model using multiple types of feedback data, we define the inputs to consider their ordinal relationships with the pairwise inputs of the NCR. This approach enables effective incorporation of different levels of implicit feedback, such as browsing versus purchasing, and demonstrates improved accuracy through actual data analysis.