Research On Recommendation Integrat-ing Social Relations

In the wake of rapid developments of Internet, websites evolve with more complex archi-tecture while providing much more information; therefore, the difficulty of exploring an-ticipated information under the enormity of online data becomes substantially aggravated.Users are inclined to be overwhelmed by the web resources and, worse still, they may failto reach those websites with potential satisfaction. Consequently, recommendation tendsto be pivotal in the age of Web2.0Traditional recommendation methods are grounded on collaborative filtering, which is ei-ther based on neighborhood assumptions or dependent on a learning model. Moreover,recommendation is handled through matrix completion. The common inherent deficiencyof all conventional recommendation approaches lies in that users are presumed to be in-dependent and social relations among them are totally ignored.In light of that, several remedies have been proposed that details of social relations are in-volved in recommendation, which makes a crucial transition from pure data mining to aframework with more practical sense, namely, social recommendation. With a systematicprobabilistic matrix factorization that jointly concerns users’ preferences and social in-formation, primitive social recommendation has already presented the-state-of-art perfor-mance on large-scale dataset with massive item loss. However, all those previous ap-proaches encounter severe technical difficulty since optimization complexity becomes in-tense as social relations are involved in user-item matrix, especially when an abundance ofuser-trust connections exist.Presently, social recommendations with more maturity are being widely concerned both inacademia and industry. As an original creation, this thesis adopts an augmented Lagran-gian algorithm with low computation complexity, to formulate the Semidefinite programand eventually leverages the matrix factorization. The essence of our work is applyinggraph Laplacian to fully estimate social connections among users, and inferring the opti-mal latent factor matrix under social-relation regularization, which guarantees that thesystem problem is described by a low-complexity Semidefinite program.In the final analysis, experimentation results on large-scale dataset Epinions and Doubandemonstrates the superiority of our proposal, with respect to performance comparison withthose prevailing social recommendation approaches, thus verifies the applicability and ef-ficiency of our work on handling large-scale dataset recommendation.