16420

1718 دكتري

دكتري

مهندسي كامپيوتر

اصفهان : دانشگاه صنعتي اصفهان

1399

يازده، 102 ص.: مصور، جدول، نمودار

عبدالرضا ميرزايي

مهران صفاياني

عمادالدين فاطمي زاده، زينب مالكي، محمدرضا احمدزاده

1400/02/02

كتابنامه

مهندسي كامپيوتر

مهندسي برق و كامپيوتر

1400/02/08

2685050

CALCF CALCF noContent SQL Rank CTPF CVAE CiteULike a CiteULike a 0 2 0 3 0 15 cs NDCG@M cs Recall@M 0 2 0 1 0 1 0 05 0 0 10 15 20 25 30 40 10 15 20 25 30 40 1 2 3 4 5 1 2 3 4 5 M Number of recommendationsFigure 8 Performance comparison on the test set for different models in the cold start scenario for theimplicit feedback dataset treated as text side information at ML OMDb and ML OMDB 100K by themselves can not show the quality of movies For the movie datasets there are other features suchas actors directors and special effects that can influence user preferences Further itshould be noted that although CVAE works well for ML OMDb and TripAdvisor it hasthe drawbacks that are shared in the models that use deep learning for recommendation such as non interpretability and need for large data ConclusionIn this thesis side information is integrated into the pointwise and listwise CF methods Two novel model are proposed CPFS that incorporates any textual categorical or real value side information into the pointwise CF methods and CALCF that incorporatescontent information into the listwise CF models To our knowledge CALCF is thefirst personalized listwise CF with content side information It is able to recommend apersonalized ranked list of items for the cold start scenario using the side information Both models have been demonstrated on a good variety of data sets movies books scholarly references and hotels which in turn have a rich variety of side informationavailable to them Some of these datasets do not include the required information sowe generate new datasets by joining available ones Side information about words in thedocuments is obtained from GloVe word embeddings CPFS significantly improved overCTPF CVAE HIRE NeuMF and BPMF specially at cold start and sparse datasets Moreover a comparison between variational inference and Gibbs sampling shows that forCPFS Gibbs sampling outperforms the variational method although this result couldnot be expected to generalise to larger datasets Also analysis shows that CALCFsignificantly improved previous works including listwise CF listRankMF SQL Rank and content aware pointwise CF CTPF The effect of content information in the cold start scenario was more visible especially where contents were informative for examplearticle abstracts at CiteULike a We have succeeded in incorporating content side information into listwise CF but inpractice there are other kinds of side information about users and items Integrating thisside information should be undertaken in future work At cold start we focus on the newitems and do not care about the new users whereas information about users may improverecommendation performance at cold start users Also an issue to resolve for future workis determining the number of topics and optimal side information automatically 14

عبدالرضا ميرزايي

مهران صفاياني

عمادالدين فاطمي زاده، زينب مالكي، محمدرضا احمدزاده