A link prediction approach for item recommendation with complex number

Recommendation can be reduced to a sub-problem of link prediction, with specific nodes (users and items) and links (similar relations among users/items, and interactions between users and items). However, previous link prediction approaches must be modified to suit recommendation instances because they neglect to distinguish the fundamental relations similar vs. dissimilar and like vs. dislike. Here, we propose a novel and unified way to cope with this deficiency, modeling the relational dualities using complex numbers. Previous works can still be used in this representation. In experiments with the MovieLens dataset and the Android software website AppChina.com, the proposed Complex Representation-based Link Prediction method (CORLP) achieves significant performance in accuracy and coverage compared with state-of-the-art methods. In addition, the results reveal several new findings. First, performance is improved, when the user and item degrees are taken into account. Second, the item degree plays a more important role than the user degree in the final recommendation. Given its notable performance, we are preparing to use the method in a commercial setting, AppChina.com, for application recommendation.