Complete this Puzzle A Connectionist Approach to Accurate Web Recommendations based on a.pdf

Complete this Puzzle: A Connectionist Approach to Accurate Web Recommendations based on a Committee of Predictors Olfa Nasraoui1 and Mrudula Pavuluri2 1 Dept. of Computer Science and Engineering Speed Scientific School, University of Louisville Louisville, KY 40292 2 Dept. of Electrical and Computer Engineering The University of Memphis Memphis, TN 38152-3180 Abstract. We present a Context Ultra-Sensitive Approach based on two-step Recommender systems (CUSA-2- step-Rec). Our approach relies on a committee of profile-specific neural networks. This approach provides recommendations that are accurate and fast to train because only the URLs relevant to a specific profile are used to define the architecture of each network. Similar to the task of completing the missing pieces of a puzzle, each neural network is trained to predict the missing URLs of several complete ground-truth sessions from a given profile, given as input several incomplete subsessions. We compare the proposed approach with collaborative filtering showing that our approach achieves higher coverage and precision while being faster, and requiring lower main memory at recommendation time. While most recommenders are inherently context sensitive, our approach is context ultra-sensitive because a different recommendation model is designed for each profile separately. Keywords: personalization, recommender systems, collaborative filtering, web usage mining, neural networks 1 Introduction The Web information age has brought a dramatic increase in the sheer amount of information (content), the accessibility to this information (usage), as well as the intricate complexities governing the relationships within this information (structure). Hence, not surprisingly, information overload, when searching and browsing the WWW, has become the plague du jour. One of the most promising and potent remedies against this plague comes in the form of personalization. Personalizat