Cross-Domain Recommendation using Deep Learning /
Rabia Khan
- Rawalpindi, MCS (NUST), 2025
- xix, 170 p
Recommender systemsplayacrucialroleinfilteringvastinformation,buttraditionalsingle- domain modelsstrugglewithdatasparsity,cold-startissues,andlimitedpersonalization. These challengesareamplifiedasuserpreferencesoftenspanmultipledomains,which single-domain recommendersfailtocapture.Toaddresstheseissues,thisthesisexplores cross-domain recommendation using Multi-Layer Perceptron's (MLPs)to model non-linear dependencies andaligndomain-specificfeatures,improvingpersonalizationandscalability. However,MLPsoverlookthegraph-structurednatureofuser-iteminteractions,limiting their abilitytomodelhigher-orderrelationships.Toovercomethis,aGraphConvolutional Network-basedframeworkisproposedtolearnbothdomain-specificandsharedembed- dings throughcollaborativegraphsignals.Additionally,tocapturefine-grainedpreferences, another modelisproposedthatintegratesmetapath-basedgraphswithtransformer-based reviewencoding.Thiscapturesopiniondynamicsfromreviewsandenhancesknowledge transfer insparsesettings. Finally,recognizingthelimitationsofratings-onlymodelsandbiasedsharedembeddings, a sentiment-awarecontrastivelearningframeworkisintroduced.Bydynamicallyadapt- ing userrepresentationsusingreviewsentiment,themodelachievesimprovedcross-domain personalization, mitigatesnegativetransfer,andsupportsaccuraterecommendationsacross heterogeneous domains like movies and books.