Kazushige Hiroi

High-accuracy time-variant recommender system with Number-Score curve

Recommender systems are becoming more effective as information becomes extensive. Conventional recommender systems show a list of recommended items, but the number of recommended items shown must be limited in some systems. For example, the recommender system for TV programs in a hard disk recorder has only a limited user interface and hardware resources, such as a remote controller, low display resolution, and limited hard disk capacity. Moreover, recommended TV programs must be changed in accordance with the time and day in this system. Thus, for these “time-variant” recommender systems, items must be recommended as accurately as possible within the limited number of recommended items. In this paper, we propose a high-accuracy time-variant recommender system with a “Number-Score curve”, which indicates the relationship between the recommendation score and the number of recommended items. We examine a recommender system for TV programs in the hard disk recorder that can achieve 82% precision and 78% recall for the recommended results with the Number-Score curve.