Session: Algorithms (Learning)
Date: Saturday, October 20, 16:00-17:45
- Addressing uncertainty in implicit preferences
by Sandra Clara Gadanho, Nicolas Lhuillier
The increasing amount of content available via digital television has made TV program recommenders valuable tools. In order to provide personalized recommendations, recommender systems need to collect information about user preferences. Since users are reluctant to invest much time in explicitly expressing their interests, preferences often need to be implicitly inferred through data gathered by monitoring user behavior. Which is, alas, less reliable.
This article addresses the problem of learning TV preferences based on tracking the programs users have watched, whilst dealing with the varying degrees of reliability in such information. Three approaches to the problem are discussed: use all information equally; weight information by its reliability or simply discard the most unreliable information.
Experimental results for these three approaches are presented and compared using a content-based filtering recommender built on a Naïve Bayes classifier.
- Robustness of collaborative recommendation based on association rule mining
by J. J. Sandvig, Bamshad Mobasher, Robin Burke
Standard memory-based collaborative filtering algorithms, such as k-nearest neighbor, are quite vulnerable to profile injection attacks. Previous work has shown that some model-based techniques are more robust than k-nn. Model abstraction can inhibit certain aspects of an attack, providing an algorithmic approach to minimizing attack effectiveness. In this paper, we examine the robustness of a recommendation algorithm based on the data mining technique of association rule mining. Our results show that the Apriori algorithm offers large improvement in stability and robustness compared to k-nearest neighbor and other model-based techniques we have studied. Furthermore, our results show that Apriori can achieve comparable recommendation accuracy to k-nn.
- Usage-based web recommendations: a reinforcement learning approach
by Nima Taghipour, Ahmad Kardan, Saeed Shiry Ghidary
Information overload is no longer news; the explosive growth of the Internet has made this issue increasingly serious for Web users. Users are very often overwhelmed by the huge amount of information and are faced with a big challenge to find the most relevant information in the right time. Recommender systems aim at pruning this information space and directing users toward the items that best meet their needs and interests. Web Recommendation has been an active application area in Web Mining and Machine Learning research. In this paper we propose a novel machine learning perspective toward the problem, based on reinforcement learning. Unlike other recommender systems, our system does not use the static patterns discovered from web usage data, instead it learns to make recommendations as the actions it performs in each situation. We model the problem as Q-Learning while employing concepts and techniques commonly applied in the web usage mining domain. We propose that the reinforcement learning paradigm provides an appropriate model for the recommendation problem, as well as a framework in which the system constantly interacts with the user and learns from her behavior. Our experimental evaluations support our claims and demonstrate how this approach can improve the quality of web recommendations.
- Improving new user recommendations with rule-based induction on cold user data
by An-Te Nguyen, Nathalie Denos, Catherine Berrut
With recommender systems, users receive items recommended on the basis of their profile. New users experience the cold start problem: as their profile is very poor, the system performs very poorly. In this paper, classical new user cold start techniques are improved by exploiting the cold user data, i.e. the user data that is readily available (e.g. age, occupation, location, etc.), in order to automatically associate the new user with a better first profile. Relying on the existing α-community spaces model, a rule-based induction process is used and a recommendation process based on the “level of agreement” principle is defined. The experiments show that the quality of recommendations compares to that obtained after a classical new user technique, while the new user effort is smaller as no initial ratings are asked.
RecSys 2007 (Minnesota)
Sponsors and Benefactors
 |
 |
 |
 |