Συστήματα συστάσεων βασισμένα στην εμπιστοσύνη

Σκιαδά, Μαριάννα Ι.

Master Thesis

Author

Σκιαδά, Μαριάννα Ι.

Date

2014-01-23

Abstract

In this thesis are presented the different techniques that be implemented in a Recommender System with emphasis on Collaborative Filtering because of its implementation in social networks. Malicious attacks are also discussed to understand the implemented strategies and techniques in order to develop more robust systems. In the effort to develop robust systems we focused on the concept of trust and the suggestion that trust is robust which is based only on one paper by O’ Donovan & Smyth, (2005) where its use is ancillary to similarity. They also define as a trusted user someone who has a history of good suggestions to others. In the experiments carried out , using the opinions dataset, variations of computable trust ,as defined by O 'Donovan & Smyth, (2005), were used concerning the range of trust (global and local variables of trust), but also putting forward a new trust metric in which it is not enough for a user to have history of positive suggestions (i.e. being part of neighborhoods that make successful suggestions) but he also has to have contributed positively to these. The evaluation results show that trust can help to increase the robustness and accuracy of the suggestions especially when used as a criterion for selecting the users that will construct each neighborhood in order to generate suggestions. Also our proposal that a user should have contributed positively to the produced suggestions to be considered reliable, seems to have positive effects on system’s robustness and accuracy.