Collaborative Filtering Efficiently Using Purchase Orders

We propose a new collaborative filtering method that can predict the next purchase item by efficiently using the sequential information in purchase histories for recommendations. Markov models and maximum entropy models are both widely used techniques for such recommendations. In Markov models, parameters can be estimated and updated fast and efficiently, but predictions may not be accurate. On the other hand, the accuracy of maximum entropy models is generally high, however parameter estimation incurs a high computational cost. We achieve both fast parameter estimation and high predictive accuracy by combining multiple simple Markov models based on the maximum entropy principle. Experiments using real log data sets of music, movie and cartoon distribution services show that the proposed method outperforms other conventional methods found in the literature.