Research On Efficient Hybrid Recommendation Algorithm For Large-scale Social Musics

With the rapid development of computer science and Web 2.0 technology,the Internet has played an increasingly important role in our life,greatly enriching people's lives.However,the continuous development of the Internet has produced a large amount of multimediadata with "noise" and redundant information,making it more and more difficult for users to select network resources.In this context,the recommendation algorithm came into being.This thesis takes social music as an example,facing the explosive growth of network music resources,how to efficiently push high-quality music content of interest to users in mass music resources has become a hot topic in academic research at home and abroad,with high theoretical research value and practical application background.In this thesis,the recommendation system is built on the basis of collaborative filtering,the collaborative filtering algorithm has the following shortcomings:1)the traditional bipartite diagram model of collaborative filtering is too simple and the calculation accuracy is not enough;2)There is a serious cold start problem because social networks are not effectively used to combine social attributes with recommendation algorithms;3)The relationship between accuracy and coverage cannot be effectively balanced,either too high or too low will affect the user experience.To solve these three problems,this thesis proposes three algorithms respectively:A tag-based tripartite diagram recommendation algorithm is proposed.The traditional bipartite diagram(user-item)is used in collaborative filtering only depends on the relationship between users and items,which is not accurate enough and lacks explanation.Therefore,this thesis introduces the element of label and changes the bipartite diagram to the tripartite diagram(user-item-tag)to improve the accuracy of the recommendation system.This thesis also reduces the dimension of labels,and clusters music segments according to the label weight of each item after dimension reduction,creating a dictionary tree index structure,further reducing the I/O and CPU calculation costs of the recommendation system,thus optimizing the performance of the recommendation system.A recommendation algorithm based on social network is proposed.This algorithm integrates the attribute of social relations in social networks into the recommendation system,making up for the defect that social attributes are not considered in traditional collaborative filtering.The trust degree is calculated based on the distance between users in the social network,and the interest preference similarity is calculated based on the user's historical behavior data.Finally,the trust degree and the interest preference similarity are combined to generate prediction scores,and users with higher prediction scores are selected for collaborative filtering recommendation,which can effectively alleviate the cold start problem of users without historical behavior records.A hybrid recommendation algorithm based on tags and social networks is proposed.The algorithm combines the music candidate sets generated by the first two algorithms in a certain proportion to make a hybrid recommendation.The positive and negative effects of the two algorithms can be effectively balanced,so that their accuracy,recall and coverage can reach a relatively balanced and reasonable level.In the experimental part,this thesis adopts the three indicators of accuracy,recall and coverage as evaluation criteria,designs corresponding experiments for the theoretical part.The experiments indicate that the improvements of the algorithms proposed in this thesis are reasonable and effective.