Methodology for recommender system information source selection based on intrinsic characteristics

Silvana Aciar

Universitat de Girona

17/04/2007 à 14h00

Grand amphithéatre, INRIA Rhône-Alpes, Montbonnot Saint-Martin

Abstract

The methods used by recommender systems to carry out recommendations, such as Content-Based Filtering (CBF), Collaborative Filtering (CF) and Knowledge-Based Filtering, require information from users to predict preferences for certain products. This may be demographic information (gender, age and address), evaluations given to certain products in the past or information about their interests. There are two ways of obtaining such information: users offer it explicitly or system can retrieve the implicit information from purchase or search history. For example, the recommender system MovieLens (http://movielens.umn.edu/login) asks users explicitly to rate at least 15 films on a scale of * to * * * * * (awful, ..., must be seen). The system generates recommendations based on these evaluations. When users are not registered into the site and it has no information about them, recommender systems make recommendations according to the site search history (Amazon.com). Nevertheless, these systems suffer from a certain lack of information [Adomavicius, 2005]. This problem is generally solved with the acquisition of additional information. The solution proposed by us is to look for that information in various sources, specifically those that contain implicit information about user preferences. These sources can be structured like databases or they can be unstructured sources like review page where users write their experiences and opinions about a product. We have found three fundamental problems to retrieve such information:

• The identification of sources of suitable information for recommender systems.
• The definition of the criteria that allow the comparison and selection of the most suitable sources.
• Retrieving the information from unstructured sources.

• A methodology that allows the identification and selection of the most suitable sources. Criteria based on the characteristics of the sources and trust measure has been used to solve the problem of identifying and selecting sources.
• A mechanism to retrieve unstructured user information available on the Web. Text mining techniques and ontologies have been used to extract information and structure it appropriately for use by the recommenders.