Models for Recommender Systems in Web Usage Mining Based on User Ratings
Gopinath Ganapathy, K. Arunesh · 2011
Abstract - In web applications, recommender systems apply statistical and knowledge discovery techniques to predict and make recommendations to the users. Automatic predictions on the interests of the users are made by the collection of ratings and other information from many other users. Collaborative filtering recommender systems make such predictions. Most of the recommendation techniques are based on navigation behaviors and ratings that might be implicit or explicit. Traditional collaborative filtering techniques are quite vulnerable to injection attacks as many provide noisy ratings that can be detrimental to the quality of predictions and also in sensitivity and sparsity problems. To alleviate these issues this paper presents two unique recommendation models namely RANK-RECO and TEST-RECO using ranking and testing measures respectively. These models evolve into algorithms that were experimented and results were provided. The approach to the recommender system is based on either individual's past behavior, which is personalized recommendation, or on the past behavior of similar users, which is social recommendation or on the items of interest, which is item recommendation. The combination of the three approaches can also be used for predictions. It is illustrated in Table 1. Amazon uses all the three approaches that are based on individual behavior, item and the behavior of other users for its recommendations. For the predictions, a few commercial sites focus on some specific methods for RS. For instance Pandora.com deeply analyses the items. It's a site for music and the technology adopted in the system recommends songs to the users based on the structural data. Basically the technology is based on a deep structural analysis of music files. The subtle musical patterns are detected and the groups are formed based on the patterns. So, the recommendations are made based on the structure of songs.