Michal KOMPAN, Mária BIELIKOVÁ
Institute of Informatics and Software Engineering,
Faculty of Informatics and Information, Technologies,
Slovak University of Technology in Bratislava
Ilkovicova, Bratislava, 842 16, Slovakia
{kompan, bielik}@fiit.stuba.sk

Abstract: The amount of information available on the Web is increasing day by day. Users are overloaded and cannot access desired information in a reasonable time. Plenty of approaches for the web personalization, which try to solve information overload have been proposed in the literature. Important feature of any personalization is its accuracy, or more precisely accuracy of personalized recommendation provided to a user. In this paper we propose a new recommendation approach for the single-user collaborative filtering based on the principles of the recommendations for group of users. We explore the best configuration for such an approach according to the group size used for computation, the aggregation strategy of ratings used within groups or the number of similar users used for the recommendation. We did an experiment over the MovieLens and SME.SK news portal datasets. Proposed approach is compared to the standard collaborative and group recommender respectively. The results support our hypothesis that the proposed approach brings statistically significant improvement and it is applicable on various domains, thus can be used for the single-user recommendation tasks.

Keywords: Collaborative filtering, group recommendation, virtual groups, aggregation strategies.

>>Full Text
CITE THIS PAPER AS:
Michal KOMPAN, Mária BIELIKOVÁ, Personalized Recommendation for Individual Users Based on the Group Recommendation Principles, Studies in Informatics and Control, ISSN 1220-1766, vol. 22 (3), pp. 331-342, 2013. https://doi.org/10.24846/v22i3y201310

Introduction

Recommender systems are an integral part of the Web nowadays. The need for the personalized web increases day by day, while people are generally overwhelmed by the amount of information available over the Web. Personalized web enables users to get access to relevant information matching their interests by filtering or recommending interesting items. Personalized recommendation becomes crucial for the business sector, where it can help to increase profits (e.g., recommending interesting products increases the chances of getting a purchase orderalso by increasing users’ visits of the web sites). Personalized recommendation is the most used approach to satisfy both-users and the business sector respectively.

The recommendation task can be defined as:

                                                                               (1)

where U represents users, S represents the recommendation objects and f is the usefulness function (usefulness of the objects for specific user u).

Several approaches have been proposed for the personalized recommendation. Two principal approaches include content-based and collaborative recommendation. In order to obtain better results these two approaches, are often combined in so-called “hybrid” approaches [5].

The content-based recommendation uses the similarity between recommended items (user liked an article about a new car, thus he/she will probably like a similar article about cars). The similarity between recommended items and a user profile can be computed based on numerous aspects such as a simple text similarity or advanced content analysis is performed. Moreover, various enhancements for specific domains as news have been proposed [11, 24, 25].

The second approach is collaborative recommendation (filtering). This approach to personalized recommendation instead of the content similarity takes advantage of user to user similarity, which is usually computed based on user’s content ratings (e.g., the readerA likes articles about cars and politics, the readerB likes articles about cars, thus the readerB will probably like also articles about politics similarly to the readerA).

Three basic models to collaborative filtering can be distinguished in respect to the focus of the computation process: user-based, model-based and item-based approaches [6]. The user-based approach creates sets of user neighbors (similar users) and then the assumption that similar users like similar items is applied. On the contrary, the item-based approach creates set of similar items, while the rating is computed based on the similar items user ratings. Finally, the model-based approach constructs users sets while the rating is derived based on the other users (within the set) ratings.

Various approaches within the collaborative recommendation have been proposed. The matrix factorization models as SVD, SVD++, PLSA or neural networks are comparable to the state-of-art approaches, while they often offer memory efficient model [13]. Neighborhood based models are used more often, thanks to their simplicity and possibility to easily understand the reason for providing specific recommendations (as this is one of the recommender system important characteristics). Providing explanations for the recommendation is often crucial from the user’s satisfaction point of view.

In order to obtain best results, these approaches are often mixed. For example, in order to create scalable news recommendation Das et al. proposed recommender system based on PLSI and MinHash as one of model-based approaches and item co-visitation as a representative of the user-based approach [8].

From the other point of view we distinguish a single-user andgroup recommenders. In the last years the phenomenon of social networking [18] and mobile devices increase has brought us to the increasing demand for recommendations designed for groups of users [12], because of the group oriented domains increase (e.g., TV, movie, holidays). The possibility of the usage of group recommendation approaches within standard single-user recommendation was raised by Masthoff [14]. In the group recommendation we use not only content or users’ similarity, but inter-group relations are considered (derived explicitly or implicitly [10]) in order to provide recommendation for the whole group instead of single-user.

Most of the proposed group recommendation approaches deal with TV or music domains as these include activities which are usually performed in groups of users [4]. The classic example of such a system is MusicFX [16], which was designed to influence a music played in the gym by actual present users. Ntousi et al. proposed the framework gRecs [19], which uses agglomerative hierarchical clustering in order to compute and predict user’s and thus group’s preferences. Recently the group recommendation was used in new domains as holiday or restaurant recommendation [17, 2].

In the group recommendation members of the group do not include only similar users (groups are formed naturally such as people in the cinema, bus, watching TV etc.), so some kind of aggregation of single-user recommendations or single-user preferences has to be performed in order to produce one list of recommendation for every group member.

As the group recommendation process is highly dependent on the aggregation strategy used, several strategies have been proposed [15]. When the standard plurality voting is used, several users can be highly unsatisfied (the minority is outvoted).

Strategies, when the majority of users take into account (e.g., average, dictatorship) are generally considered as strategies without minimal satisfaction (the minimal level of satisfaction of every group member is not guaranteed).