Recommender systems have been widely used by online stores to suggest items of interest to users. These systems often identify a subset of items from a much larger set that best matches the user's interest. A key concern with existing approaches is overspecialization, which results in returning items that are too similar to each other. Unlike existing solutions that rely on diversity metrics to reduce similarity among recommended items, we propose using choice probability to measure the overall quality of a recommendation list, which unifies the desire to achieve both relevancy and diversity in recommendation. We first define the recommendation problem from the discrete choice perspective. We then model the problem under the multilevel nested logit model, which is capable of handling similarities between alternatives along multiple dimensions. We formulate the problem as a nonlinear binary integer programming problem and develop an efficient dynamic programming algorithm that solves the problem to optimum in O(nKSR(2)) time, where n is the number of levels and K is the maximum number of children nests a nest can have in the multilevel nested logit model, S is the total number of items in the item pool, and R is the number of items wanted in recommendation.
展开▼
