A Diversity by Design Recommender System for the Movie Application Domain.: A recommender system with diversification by design for the movie application domain.
Translated title
A Diversity by Design Recommender System for the Movie Application Domain.
Author
Zanitti, Michele
Term
4. term
Publication year
2017
Submitted on
2017-09-21
Pages
76
Abstract
Anbefalingssystemer personaliserer indhold på nettet. Når de optimeres hårdt efter præcision—at forudsige hvad du sandsynligvis vil klikke på—kan de blive overpersonaliserede. Det kan skabe filterbobler, hvor brugere kun eksponeres for en lille del af kataloget, så oplevelsen indsnævres over tid. En udbredt modforanstaltning er at øge diversiteten i anbefalingerne, selv hvis det koster lidt præcision. Specialet undersøger en løsning med “diversity by design” til film. Med afsæt i tidligere forskning præsenteres et brugercentreret rammeværk, der øger diversitet på fire dimensioner: global dækning (hvor meget af hele kataloget får eksponering), lokal dækning (hvor blandede anbefalinger en enkelt bruger får), nyhedsværdi (hvor meget nyt/ukendt indhold vises) og redundans (begrænsning af meget ens gentagelser). Løsningen modellerer brugeres profiler ud fra kategoriske præferencer og søger bevidst at variere profiler inden for den samme brugergruppe under selve filtreringen af anbefalinger. Der udvikles et proof‑of‑concept for at vurdere, hvor meget diversitet denne tilgang giver ved at sammenligne udvælgelse af forskellige brugere inden for en gruppe med tilfældig udvælgelse, på tværs af forskellige niveauer af brugerdiversitet.
Recommender systems personalize online content. When they are optimized mainly for accuracy—predicting what you are most likely to click on—they can become overpersonalized. This can create filter bubbles, where people see only a small slice of the catalog and their experience narrows over time. A common countermeasure is to increase diversity in recommendations, even at some cost to accuracy. This thesis explores a diversity-by-design solution for movies. Building on prior research, it presents a user-centric framework to increase diversity along four dimensions: global coverage (how much of the whole catalog gets exposure), local coverage (how mixed an individual user’s list is), novelty (how much new or unfamiliar content is surfaced), and redundancy (limiting very similar repeats). The solution models user profiles using categorical preferences and deliberately varies profiles within the same user group during recommendation filtering. A proof of concept evaluates how much diversification this approach achieves by comparing the selection of diverse within-group users with random selection, across different levels of user diversity.
[This abstract was generated with the help of AI]
Keywords
Recommendation System ; Diversity by design ; Clustering ; Diversity Filtering ; Over-specialisation ; movie application domain ; user-centric framework ; Latent Semantic Analysis ; Singular Value Decomposition ; Coverage ; Novelty ; Redundancy ; profile modeling ; Filter bubble ; Decision Psychology
Documents