Back to search

National Technical University of Athens (NTUA)

Συνεργατικά συστήματα συστάσεων με χρήση έμμεσης και άμεσης κοινωνικής πληροφορίας

Abstract

dc:description

The digital revolution of the past decades has had as a direct consequence the rapid increase of the information distributed in electronic form, leading to a situation known as the information overload. Therefore, there exists an immediate need for the development of systems that can search, classify and categorize the available information. Recommender Systems have been proposed as a solution that may deal with the aforementioned problem to a certain extent. The contribution of this thesis lies in the study, development and practical application of algorithms for the most popular category of recommender systems, that of collaborative filtering, aiming at improving their performance.Recommender systems are essentially software tools used for the production of personalized recommendations for each user individually, usually in the form of ordered lists. Recommendations are in most cases items of any kind, like books, movies, music tracks and news articles, that may be useful to the users. The creation of recommendations in collaborative filtering systems is based upon the preferences of each user along with other parameters. Preferences are deduced in two ways; either directly from users (though the ratings they provide in a predefined scale) or indirectly, by the system itself.In the context of this thesis, recommendation algorithms that utilize the explicit user preferences are developed. The notion of collaboration lies in the fact that the new items recommended to a user are based on the ratings provided to them by other “similar” users. That is, collaborative filtering turns each user to the other user's predictor, thereby defining an informal social bond between them. For the exploration and the better exploitation of this kind of indirect relationships between the users, a novel recommendation algorithm is being proposed, whose structural element is a feed-forward multilayered perceptron. The choices made regarding the transfer function of the neurons of the network, the dynamic adaptation of its architecture as well as the boosted training algorithm, allow for the deep analysis of the relationships between the users that lead to the production of better recommendations.Apart form the indirect deduction of similarities and dissimilarities between the users, collaborative filtering can also be extended through the inclusion in the computations of the direct relationships their users may have, in the context of a social network. Examples of such relationships are the bonds of friendship, trust as well as the “follower” ties. Among the most widespread ways of accessing the direct social information in recommender systems is through the performance of random walks, aiming at locating influential users. In this context, a novel methodology for performing random walks in the joint network constructed from the combination of the social network and the “similarity” network (a network of ties between users, based on the ratings they have provided) is proposed. The difference to other approaches lies in the fact that the next step of the walk is not chosen uniformly at random, but a bias towards the most similar users is introduced instead.In addition to the user space, random walks may also occur in the item space. For this reason , the performance of random walks in a novel network, the item consumption network, is also proposed in this thesis. This specific network connects item-nodes with vertices whose weights depict the number of their common access patterns. This network is also personal, in the sense that it is constructed for every user, based on the items that he or she have accessed, as well as other users that are members of either his/her direct social network, or are similar to him/her or belong to both categories. Finally, the direct social relationships are exploited in yet another way in collaborative filtering systems. They constitute the basis of a new bayesian method of non-negative matrix factorization, aiming at categorizing the users in overlapping communities. The novelty of the procedure lies in the choice of the a priori distribution. While in literature a probability distribution based on the conjugate prior theory is selected, in this specific case, the exponential random graph models are used. The reason for choosing this course of action is to balance the locality of the factorization through the introduction of global network properties in the process. Additionally, in the context of the same procedure, a novel approximation method for the hyperparameters of the exponential random graph models has been developed, based on the mean field theory.In conclusion, the actual contribution of this thesis is summarized in the use of intelligent techniques in collaborative filtering recommender systems, for the processing and analysis of the direct and indirect social relationships the aforementioned systems contain. Every proposed technique has been experimentally tested in publicly available datasets that are widely used by the scientific community, while comparisons with other relevant techniques in every research field have also been performed.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2015

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Alexandridis, Georgios
  • Αλεξανδρίδης, Γεώργιος

Subjects

dc:subject × 20

Rights

dc:rights
Statement dc:rights
  • BY_NC_SA
Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/37071

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Alexandridis, Georgios; Αλεξανδρίδης, Γεώργιος. Συνεργατικά συστήματα συστάσεων με χρήση έμμεσης και άμεσης κοινωνικής πληροφορίας. National Technical University of Athens (NTUA), 2015. http://hdl.handle.net/10442/hedi/37071