Pairwise One Class Recommendation Algorithm

Abstract

We address the problem of one class recommendation for a special implicit feedback scenario, where training data only contain binary relevance data that indicate user’ selection or non-selection. A typical example is the followship in social network. In this context, the extreme sparseness raised by sparse positive examples and the ambiguity caused by the lack of negative examples are two main challenges to be tackled with. We dedicate to propose a new model which is tailored to cope with this two challenges and achieve a better topN performance. Our approach is a pairwise rank-oriented model, which is derived on basis of a rank-biased measure Mean Average Precision raised in Information Retrieval. First, we consider rank differences between item pairs and construct a measure function. Second, we integrate the function with a condition formula which is deduced via taking user-biased and item-biased factors into consideration. The two factors are determined by the number of items a user selected and the number of users an item is selected by respectively. Finally, to be tractable for larger dataset, we propose a fast leaning method based on a sampling schema. At the end, we demonstrate the efficiency of our approach by experiments performed on two public available databases of social network, and the topN performance turns out to outperform baselines significantly.