Context-aware Recommendation Based On User Behavior Mining

The rapidly evolving Internet has revolutionized our traditional habits in an unprecedented manner. Nowadays, making restaurant reservation and other recreations go hand in hand with the Internet. However, the roaring Internet also brings about the "information overload" issue, which overwhelmed us in the age of big data. Consequently, the recommender system comes into being in that background. The traditional personalized recommender system focuses on information such as ratings, labels, texts, but some websites lack such data sources. Therefore, in some scenarios, navigation could serve as the implicit user feedbacks, and be exploited further in order to predict users’ preferences.Firstly, we extract the user behavior information from navigation logs and quantify the users’ interest by mining the behavior semantics based on these implicit feedbacks. After that, we construct a high-dimensional context-aware model and establish correlation between users and items. Then, by virtue of semantic opposites of certain behavior characteristics, we compress the discrete data and reduce the sparsity of the model.Secondly, the classical matrix factorization algorithms tend to ignore the semantic ordering of each pair of items, for it is merely designed to solve rating prediction problem. In order to compensate for the existing methods, we propose a Bayesian personalized ranking algorithm based on the user behavior semantic information, which keeps the correlation of behavior semantics in the latent space through Laplace regularization. Finally, we formulate the ordering of each pair of items, and optimize the proposed model by BRP-Opt algorithm.At the end of this paper, we compared our model with several classical recommendation algorithms in the Entree dataset. The experimental results indicate that the proposed algorithm outperforms the others in several evaluation metrics.