Multicriteria User Modeling in Recommender Systems

Abstract

Recommender systems are software applications that attempt to reduce information overload. Their goal is to recommend items of interest to the end users based on their preferences. To achieve that, most Recommender Systems exploit the Collaborative Filtering approach. In parallel, Multiple Criteria Decision Analysis (MCDA) is a well established field of Decision Science that aims at analyzing and modeling decision maker's value system, in order to support him/her in the decision making process. In this work, a hybrid framework that incorporates techniques from the field of MCDA, together with the Collaborative Filtering approach, is analyzed. The proposed methodology improves the performance of simple Multi-rating Recommender Systems as a result of two main causes; the creation of groups of user profiles prior to the application of Collaborative Filtering algorithm and the fact that these profiles are the result of a user modeling process, which is based on individual user's value system and exploits Multiple Criteria Decision Analysis techniques.Experiments in real user data prove the aforementioned statement.

A Neural Networks Approach for Improving the Accuracy of Multi-Criteria Recommender Systems

“…The tables have shown the decrease in RMSE and MAE and also increase in F-measure AUC, NDCG, and FCP, that highlight the positive performance of ANN-based MCRSs. The experimental results showed clear accuracy improvements over the ones observed in earlier studies [6,27,37]. Furthermore, to emphasize the good performance of our models, we conducted some statistical analysis by extracting some of the predicted ratings of all the algorithms and comparing them with the actual ratings from the dataset.…”

“…The Yahoo!movie dataset: It is a multi-criteria dataset where preference information on movies was provided by users on the strength of four different movie attributes (criteria); namely, the direction (k 1 ), the action (k 2 ), the story (k 3 ), and the visual (k 4 ) effect of the movie [27]. Ratings of each criterion were measured on a 13-fold scale starting from F representing the lowest preference to A + , which stands for the highest preference.…”

A Personalized Location Aware Multi-Criteria Recommender System Based on Context-Aware User Preference Models

“…This algorithm is a regression-based technique that infers preference models from given global preferences (e.g. previous user choices) [2,3]. For more information on this algorithm, we recommend [2][3][4][5].…”

“…previous user choices) [2,3]. For more information on this algorithm, we recommend [2][3][4][5]. In order to apply this algorithm we first need to obtain the user's weak preference order, which represents the user preferences for some items, after considering the four criterions for each of them.…”

The Semantics of Movie Metadata: Enhancing User Profiling for Hybrid Recommendation

“…The key to more effective personalization services requires a system able to understand not only what people like, but why they like it. In other words, the ability of creating a more effective preference representation schema, will potentially lead to the design of a recommendation algorithms with increased performance [7]. To go beyond and overcome the common limitations of the use of preferences expressed only in form of ratings, a research trend which can exploit both user preferences and semantic contents, has been emerging [8,9].…”