Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Avoiding content bubbles by network-friendly recommendation algorithms

Tzimpimpaki Evangelia

Full record


URI: http://purl.tuc.gr/dl/dias/1C0895C1-E524-4706-AEC5-40A6905B5E4C
Year 2024
Type of Item Diploma Work
License
Details
Bibliographic Citation Evangelia Tzimpimpaki, "Avoiding content bubbles by network-friendly recommendation algorithms", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.98885
Appears in Collections

Summary

Almost all online services encourage users to establish a profile, granting access to personalized content. Having more and more detailed data from the user, allows for the platforms to detect his interests and to create the content that has the greatest chance for success. However, there are instances when recommendations become excessively personalized, especially in (cache-friendly) systems also guiding suggestions towards content with low access cost. This can lead the user in a state where they are consistently presented with content of a singular nature, which may or may not sustain his interest in the long run. This thesis aims to improve recommendation systems, by increasing the diversity of recommended content, thus preventing the creation of content bubbles. First, an overview is provided, initiating with the exposition of Baseline Recommendation Systems (BS-RS), their evolution into Network-Friendly Recommendation Systems (NF-RS), and the representation of the content bubble phenomenon in NF-RS. The setup of BS-RS and NF-RS as optimization problems is detailed, and the introduced Diverse NF-RS is presented, addressing the content bubble phenomenon. The optimization problem for Diverse NF-RS is formulated, demonstrated to be convex, and linearized before being solved. No previously established implementation adequately addresses the diversity issue with comparable cost-diversity trade-offs. The proposed solution incorporates additional fairness metrics from other works, establishing that our proposed Recommendation System can accommodate them without compromising the favourable trade-offs achieved.

Available Files

Services

Statistics