User behavior analysis method and device are disclosed. The method comprises: obtaining a behavior record of users regarding commodities; acquiring, by determining a commodity feature vector in a predetermined commodity feature space of each of the commodities, a commodity feature matrix consisting of the commodity feature vectors of the commodities; for each of the users, calculating, based on a sub behavior record of the corresponding user regarding each of the commodities, a preference score of the corresponding user regarding each of the commodities, so as to get a score matrix composed of the preference scores of the users regarding the commodities; and for each of the users, determining, based on a regularized least squares based solution of a difference function of the score matrix and a prediction matrix, a user feature vector of the corresponding user in its related predetermined commodity feature space.