A Personalized Recommender System from Probabilistic Relational Model and Users’ Preferences

Abstract

Recommender systems are applications to retrieve useful information from large amount of online data to assist users in discovering interesting items/products in the system. Collaborative filtering, content-based filtering, demographics-based filtering and hybrid approach are main approaches to realize recommendation systems. Most of the existing algorithms use a single approach to deal with recommendation problems. Besides, traditional recommendation approaches mainly deal with single dyadic relationships between users and items whereas data in real world are generally conceptualized in terms of objects and relations between them. Recommender systems based on Probabilistic Relational Model (PRM)1,2, a framework for learning probabilistic models from relational data, have tried to address this issue. However, existing PRM-based recommendation algorithms do not fit into our context where we are struggling with the contradictory situation of a real-world application that requires building a personalized recommender when no user profile exists. Therefore, we propose a novel approach to build a personalized PRM-based recommendation model with the help of users’ preferences on decision making criteria. Using our approach, content-based, collaborative filtering as well as hybrid models can be achieved from the same PRM. Applying the model on a real-world data from a cold system, we show that our model is actually capable of personalizing recommendations in coldstart situation.