A Unified Recommendation Framework Based on Probabilistic Relational Models

Abstract

Recommender systems are being increasingly adopted in various e-commerce applications. A wide range of recommendation approaches have been developed to analyze past consumer-product interactions, consumer attributes, and product attributes to predict customer preferences and future transactions. In this paper we propose a unified recommendation framework based on a relational learning formalism, probabilistic relational models (PRMs). This framework includes most existing recommendation approaches, such as collaborative filtering, content-based, demographic filtering, and hybrid approaches, as special cases. Relational learning attempts to capture the relational data patterns within a database containing multiple interlinked data tables. As a main statistical model for relational learning, PRMs extend from standard models such as Bayesian networks by describing probabilistic dependencies among attributes of the same object as well as attributes of related objects in a relational domain. The inputs for recommendation are typically stored in a relational database and the general recommendation problem can be naturally formulated as a relational learning problem. We extended the original PRMs in order to capture relational data patterns that are important for recommendation modeling. We also customized the algorithm for learning PRMs for both dependency model construction and parameter estimation to exploit the special characteristics of the recommendation problem. Through an experimental study, we demonstrate that the proposed framework not only conceptually unifies existing recommendation approaches but also allows the exploitation of a wider range of relational data patterns in an integrated manner, leading to improved recommendation performance.