Research On Bipartite Network Structure-based Recommendation Algorithm

The generation of massive data causes the problem of information overload.Recommendation algorithm is an effective way to alleviate this problem,as it can help people to find out the information they are interested in the mass of information.Recommendation diversity is an important indicator of recommendation quality,however,many recommendation ignore the diversity,while only focusing on improving the accuracy.It is of great significance to develop a user-satisfied and diverse recommendation algorithm.Bipartite network structure-based recommendation algorithm has many advantages,such as good accuracy,without limited by project type and etc.This algorithm has been utilized in many recommendations,but it also has the problem of poor diversity performance.After an in-depth study of the relative algorithm,in this paper,the PersonalRank is improved in terms of diversity.The main works are as follows.Firstly,aiming at the problem of insufficient diversity of bipartite network structure-based recommendation algorithm,this paper proposes Incorporating user preferences and network structure for recommendation(IUNR).The algorithm uses preferences to modify the weight of user-item bipartite graph.Random walking on the new bipartite graph gets a recommendation list based on user-item.Then the algorithm predicts user ratings on tags,extract indirect related items according to user's favorite tags,and gets a recommendation list based on user-tag.Finally,the algorithm fusion of lists from the previous two different methods to get the final recommendation list.The preference modify rating to reduce the recommendation probability of hot items,and to improve the recommendation probability of cold items.Label data is fully utilized for recommendation.Recommendation from different dimensions to get a list that meet user's personalized interests and have more diversity.Experiment are conducted on MovieLens dataset,and the results show that the algorithm can effectively improve the diversity of recommendation without losing precision and recall.Secondly,aiming at the situation that many recommendations in the network contain review data and review data contains a lot of useful information,this paper uses the review data to improve IUNR,and proposes incorporating review topic and network structure for recommendation(IRNR).The algorithm uses Topic-Sentiment hybrid model to mine the probability distribution of user-item-sentiment topic and the feature topic of item.The rating model calculates the user's rating on the topic and finds out the list of user's favorite topics.Through these topics,we can find the recommendation list based on user-item,which is fused with the recommendation list based on user-item to get the final recommendation list.Experiments are conducted on Douban movie review dataset,and the results show that the algorithm improves the precision,recall and diversity.