Archive for the ‘ Recommender Systems ’ Category

Efficient Adaptive-Support Association Rule Mining for Recommender Systems – Lin, Alvarez, Ruiz Kluwer 2001

This paper deals with online collaborative recommendation ASARM by focusing on only those association rules pertaining to a particular user or article at a time. The algorithm also provides enhancements by introducing a heuristic to adapt minimum support for association rules to be generated and instead of using a confidence threshold, uses a range of rule size instead.

Among several other competing techniques, it scores better in terms of both quality of results and efficiency of rule generation thus making it a convenient approach for online collaborative recommendations. Among the techniques, Shardanand and Maes (1995) and Resnick et al (1994) . propose variants of linear (spearman) correlation has been used as well addressing the same problem but it generates only linear relationships between rules and thus misses out on apparent but non-linear association rules. Breese et al have used Bayesian Networks but the problems with this approach is that a prior conditional probability for each rule has to be calculated which is an expensive operation. Secondly, the quality of induced rules cannot be measured. Billsus and Pazzani have used Neural Networks based on feature reduction schemes including Singular Value Decompostion (SVD) and Information Gain. The space is reduced to a lower dimensionality and then neural networks are used to create the recommendation model.

For user associations, all rules which are in the rule size range and hav a minimum support in the data are fired. For article associations though, besides the minimum support, a score is also associated with each article and for a rule to be fired containing a particular article for head, this score should be above a minimum threshold value. This score is the sum of the scores of all those rules for which this article is fired. The score of a rule is the sum of the product of support of the rule and its confidence.

After empirical tests on a commerical online movie recommendation data, the authors come to the following conclusions:

The method is slightly better in terms to the neural networks based approaches proposed by Billsus and Pazzani. For collaborative recomendations, if a user’s target minimum support is too low, it takes a long time to induce the rules and the quality of the induced rules is also deteriorated. For such a case, use article associations instead of user associations in case user associations minimum support is below a threshold otherwise use user association. With this heuristic, the possibliity of inferring new users and new articles for whom minimum support is always low. For majority of users, only a few calls to the main association rule algorithm ASARM2 needs to be called as compared to the multitude of times it has to be called in the conventional A Priori algorithm.