User-to-User Recommendations in a Fashion Portal, Utilizing a Social Network, Implicit Feedback and Clustering Approaches
Authors
Thøisen, Dan Duus ; Kock, Bjarne ; Frazzetto, Davide
Term
4. term
Education
Publication year
2015
Submitted on
2015-06-03
Pages
108
Abstract
I denne afhandling undersøger vi problemet med at forudsige sociale forbindelser—hvilke brugere der sandsynligvis vil knytte kontakt—og bruger resultaterne til at designe et anbefalingssystem til Sobazaar-socialnetværket. Domænet har to særlige kendetegn. For det første er der kun implicit feedback til rådighed: adfærdssignaler frem for eksplicitte vurderinger. For det andet skaber skiftende modetrends komplekse sociale dynamikker i netværkets struktur, så det er afgørende at finde fællesskaber af brugere med lignende karakteristika for at forstå deres adfærd. Vi finder, at en kombination af clustering—at gruppere lignende brugere—og information udledt af brugernes aktivitet kan forbedre forudsigelsen af sociale forbindelser i denne kontekst.
This thesis studies the social link prediction problem—predicting which users are likely to connect—and applies the findings to design a recommender system for the Sobazaar social network. The setting has two notable features. First, only implicit feedback is available: behavioral signals rather than explicit ratings. Second, shifting fashion trends create complex social dynamics in the network, making it essential to identify communities of users with similar characteristics to understand behavior. We find that combining clustering—grouping similar users—with information extracted from user activity can improve social link prediction in this context.
[This abstract was generated with the help of AI]
Keywords
Documents