Link prediction based hybrid recommendation system using user-page preference graphs

The distribution of the amount of preference information across customers is not same in every domain of recommendation problems. It is necessary to treat each user differently based on their available preference information. On the other hand, graph structure can provide better representation of user-item preference information. By exploiting graph structure, recommendation systems could be made more reliable and effective. In this paper, several graph structure based features are used to predict likelihood of a page or item being preferred by a user, or future connection probability of an unconnected user-page node pair in the user-page preference graph. A novel, user-specific, parametric method to integrate page-page content similarity and co-occurrence similarity in the graph context is introduced. Two graphs were created; one using page-page content similarity; other one using page-page co-occurrence similarity. An approach that utilizes features derived from these two graphs to make web page recommendation is introduced. For each user-page pair, one combined feature component is first obtained by making a weighted summation of the eight features extracted from each graph. Use of supervised learning for deriving relative weights of the two eight-feature sets to obtain a combined value of the feature components yielded highly promising results. Finally, the two feature components, from the two graphs are combined in user-specific way to train a model and to make recommendations. Experimental results on Yahoo Front Page Today Module Click Log dataset show better results compared to other approaches in the literature. To the best of our knowledge, ours is the first such effort in the context of graph-based recommendation systems.