Pairwise interaction tensor factorization based tag recommendation

The tag recommendation system is a series of tags that are most likely to be used to tag a resource for the target user. Currently, the Tucker decomposition model has better prediction quality than the traditional FolkRank algorithm, but it has high time complexity and is difficult to apply to large and medium-sized data sets. Although the time complexity of the regular decomposition model is linear, its prediction quality is not high. To solve these problems, firstly， the paired interaction tensor decomposition model PITD on the basis of improving the Tucker decomposition model is proposed. The model considers only some of the two-to-two interactions between the three characteristics of users, resources, and tags, reducing the impact of irrelevant information on model performance and efficiency. Then, the PITD model is deduced by Bayesian personalization method, and the corresponding optimization algorithm is designed. Finally, extensive experiments on real data sets show that the PITD model has better recommendation performance than the comparison algorithm.