A Three-Way Recommender System For Popularity-Based Costs

Recommender systems help e-commerce corporations to make profit among a large amount of customers. Three-way recommender systems handle this issue through considering misclassification and promotion costs. The setting of costs in existing approaches is the same for items with different popularity. However, success recommendation of unpopular items is more profitable. In this paper, we define a new cost-sensitive recommendation problem. The new problem is more general than existing ones in that the cost function is variable w.r.t the popularity of the item. First, we adopt a three-way approach with three kinds of actions: recommending, not recommending and promoting. For any item, a threshold pair is calculated from its cost matrix. Second, we employ the M-distance to obtain the probability which measures how much a user likes an item. Consequently, the action to any item for any user is determined. Experiments are undertaken on the well-known Movielens dataset. Compared with the existing three-way recommendation algorithm, our algorithm results in less average cost through recommending more unpopular items.