Dynamic sequential clustering of time-variant data using cluster merging and splitting based on data movement and variance for IPTV program recommendation

Abstract

Item recommendation based on collaborative filtering is often based on similar user clustering. A user belonging to a similar preference group is recommended with the items that are also preferably consumed by the other users in the group. For this, similar user clustering is a key issue for reliable recommendation. In this thesis, we study a dynamic sequential clustering method for time-varying data, which is more often applicable for automatic recommendation of TV programs to active TV viewers in IPTV environments. The traditional clustering methods have often focused on time-invariant data. In TV watching environments, a user preference vector is defined as the representative of each TV viewer, which is time-variant because the user preference on TV programs is changing over time. To cope with this, we study a dynamic clustering method that handles such time-variant sequential data. To group similar users for recommendation, our proposed dynamic clustering method roughly consists of four parts: (1) clustering time-variant user’s preference vectors; (2) merging the clusters which are closely located; (3) splitting a cluster which can be divided into two or more smaller user groups ; (4) updating the user membership. The proposed dynamic clustering method is tested with user’s TV watching history data by Nelson Korea but in order to see the effectiveness of the our proposed method, we also test it with artificially generated data sets as well.