Collaborative Filtering in Social Networks
Author
Mateo, Sergio
Term
2. term
Education
Publication year
2010
Submitted on
2010-05-31
Pages
88
Abstract
Dette speciale undersøger, hvordan sociale netværk kan forbedre anbefalingssystemer baseret på collaborative filtering. Arbejdet fokuserer på to hovedspørgsmål: hvordan man mest effektivt beregner lighed mellem brugere i forskellige scenarier, og hvordan brugere kan give feedback, der påvirker de anbefalinger, de modtager. Forfatteren gennemgår og implementerer flere lighedsmetoder (bl.a. Pearson-korrelation og varianter, der inddrager varians og alle ratings) samt formler til at estimere ratings. Der præsenteres tre feedback-teknikker: at brugeren manuelt kan justere lighed, at systemet kan udelade bestemte naboer i anbefalinger, og at bestemte emner kan udelades ved lighedsberegning. En prototypisk recommender implementeres med datasæt bestående af ratings, brugere og items, og den evalueres på tværs af scenarier og datasætsstørrelser ved hjælp af gennemsnitlig absolut afvigelse, præcision, recall og F1 samt målinger af beregningstid, herunder tiden til at beregne ligheder. Specialet rapporterer komparative analyser og diskuterer afvejninger mellem nøjagtighed, brugerindflydelse og effektivitet; konkrete resultater og konklusioner uddybes i de fulde kapitler og fremgår ikke af dette uddrag.
This thesis investigates how social networks can enhance recommender systems that rely on collaborative filtering. It addresses two core questions: how to compute user similarity effectively across different scenarios, and how to let users provide feedback that meaningfully steers their recommendations. The work surveys and implements multiple similarity measures (including Pearson correlation and variants incorporating variance and all ratings) alongside rating estimation formulas. It introduces three feedback techniques: allowing users to adjust similarities manually, skipping selected neighbors in recommendations, and skipping selected items when computing similarities. A prototype recommender is implemented using datasets of ratings, users, and items, and is evaluated across scenarios and dataset sizes using average absolute deviation, precision, recall, and F1, as well as processing-time measurements, including the time to compute similarities. The thesis reports comparative analyses and discusses trade-offs between accuracy, user control, and efficiency; specific quantitative findings and conclusions are detailed in the full chapters and are not included in this excerpt.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents