| Nowadays,with the rise of music platform,digital music gradually replaces physical music as music consumption content recognized by more and more people.Among them,personalized music recommendation has also become a hot topic in the field of personal-ized recommendation.However,personalized recommendation requires not only automatic filtering but also content organization.At present,the huge online music content makes this task difficult,and a large number of irrelevant music gradually exceeds the bearing point of users,which will lead to the increase of information fatigue of users,thus losing the user activity.Therefore,how to help users find highly relevant music that conforms to their personal interests is an urgent problem to be solved.The current existing recommender systems are facing various issues:1.The "cold start".2.Slow processing time.3.Poor accuracy of recommended items.These problems are among the key problems that are facing music recommender systems out there.Currently,content filtering and collaborative filtering are the two most commonly used technologies in recommendation systems,which have their fatal disadvantages to different degrees.Firstly,content filtering is mainly to calculate the similarity between music by extracting feature parameters.Although this method is intuitive and can be searched in music libraries with large amount of data,the results obtained are very close to the previous music content,so some music which is different in content and nature may become user's interest can not be obtained.Collaborative filtering is mainly based on user behavior to identify some users who have the same preferences,calculate the distance between them,and use similar users to predict the preferences of target users.However,the music obtained under this method is generally those songs that have existed and scored high,and some new songs can not be recommended because they do not have enough attention.However,these two algorithms have a certain complementary relationship,so the research focus of this paper is to develop a user-oriented recommendation system.The thesis employs algorithms from the field of user-to-user behavior,genre matrix,acous-tic features and audio similarity metrics to recommend personalized music and overcome the problems,through the following ways:1.In content-based filtering.In this paper,the research and application of”music similarity" are carried out,mainly focusing on audio acoustic characteristics and music similarity measurement,in addition to the similarity calculation based on music tag.Through research,I use MFCC in audio acoustical characteristics,which is the most popular feature currently excavated from music audio,and then use Mahalanobis distance to calculate,through comparison,we can conclude that the shorter the Mahalanobis distance is,the more similar the track is,the more we want to recommend the target music.Another point is music similarity measurement.I use weighted Euclidean distance to complete the calculation of song similarity.Here,I add a weight to each user's song difference,so that the results will become more and more accurate as the number of users increases.2.In collaborative filtering.In this respect,the paper uses a genre matrix method to identify the popularity of music and its usage in a community.In these two parts,I used two matrices to determine,which adopted matrix multiplication and proposed a user-oriented model,which captured the popularity of specific music in specific regions and generated user profile,so that users with the same content could recommend music that they might be interested in.Second,by examining user-to-user behavior to show a close correlation between the playlists and songs they listen to,we were able to develop a system to capture user behavior and use its results to recommend relevant music content.It is based on the similarity of the songs the user listens to.The song similarity was calculated by Pearson correlation between two songs that users had heard and rated. |
PDF Full Text Request