Combo-recommendation based on potential relevance of items

Combo recommendation expects to recommend a collection of products to users in a Groupon way. The representative application is combo recommendation in the travel industry, which is also called package recommendation and may include different landscapes according to the inherent features. Compared with traditional recommendation scenario, combo recommendation has the following characteristics: (1) sparsity: information for combos is much less than that for individual items; (2) collectivity: every combo is composed of multiple individual products with different features; (3) diversity: products composed of combos may have different features; (4) relevance: products inside combos have some kind of potential relevant. Traditional recommendation algorithms may perform poor for they consider nothing about these four characteristics in the models. Aiming at improving performance of combo recommendation, our work proposes a novel combo recommendation algorithm called RBM-CR based on the Restricted Boltzmann Machine. RBM-CR algorithm takes advantage of users’ consumption histories to derive the correlations among products by mapping from visible features to hidden features, and to profile users and combos by those hidden features. Finally, experiments on real dataset verify effectiveness and accuracy of our algorithm.