Graph-based Recommendation Technology Using Combination Feature

The scale of modern commercial ventures based on Internet increases rapidly.Getting through every entity is beyond the reach of users' ability.The recommendation system is a subclass of the information filtering system aiming at predicting the user's rating or preference to the items so that users can quickly find the items they may be most interested in and have not tried.The amount of data of Web is continuing to grow substantially and the utility of semi-structured data is increasingly high.Graph data model can represent multiple relations between entities and greatly enrich the ability of data representation,which is able to unify a variety of recommendation approaches into a unified model.Then we can utilize graph metrics to make recommendations.This dissertation studies recommender system based on graph and proposes a nov-el concept of Combination Feature.The existing graph-based works are often based on the path feature to build the recommender system.We found that those works are often concerned about single item for generating a user's profile.Combination Fea-ture is different from the existing works,which focuses on the combination of items.The basic idea of Combination Feature is that utilizing the combination of items can contribute to the users' profile and have the value of recommendation.We construct features based on tree pattern using the graph of items and their related nodes then compute the feature values based on items set's degree centrality.Based on Combination Feature,we have designed a user-based collaborative fil-tering method(CFC)and a learning-to-rank based method(LRC).CFC exploits Com-bination Feature to search the similar users first and based on the similar users we gen-erate a recommendation list.For LRC,we build a feature space for User-Item pairs,and then we learn a predictor for unknown User-Item pairs.In the last part of this dissertation,we design the system based on those two methods and perform the exper-iments to test our proposals.Experimental results show that our proposals outperform other competitors in recall.