Research On Multi-label User Classification Of Social Media

Recently,the rapid development of social media has brought tremendous convenience for people to obtain information such as events and news,and has become an important tool for people to participate in online activities.Social media accounts serve as an important medium for users to obtain information,post comments and communicate with friends on social media.By classifying social media accounts,accounts with common features can be identified as a set,which describing the user's semantic attributes simultaneously.Effective account classification methods can not only help users quickly find information of interest,but also achieve reliable account management and personalized content and user recommendations.Generally,the information sent or followed by users in social media often comes from multiple topics.A single-label classification of these accounts will miss a lot of other topic information.But by multi-label classification of accounts can fully describe the user's Multi-semantic attributes.However,most of the existing researches on multi-label classification algorithms focus on the fields of text and pictures,and there are few studies on multi-label classification of accounts.In addition,the existing multi-label account classification algorithms do not achieve an effective representation of accounts,which makes it difficult to achieve a good classification effect directly in a social media scenario.This thesis analyzes complex social networks and implements multi-label classification of accounts based on single-label classification of accounts.The main contributions are as follows:(1)For user single-label classification,a semi-supervised classification method for accounts based on heterogeneous networks is proposed.In the network construction section,considering that social media contains a variety of information,this information can be effectively fused by constructing a heterogeneous network,which can better describe the real social scene.In the feature extraction section,the user relationship features and special term frequency features are extracted by analyzing the meaning behind the interaction behavior between users in the network;the user entity relationship features are extracted by analyzing the user's post behavior.In the model training section,this thesis uses the idea of semi-supervised to implement model training,which reduces the need for training sets and improves the accuracy of account classification.In actual scenarios,by comparing with some existing account classification methods,it is found that the algorithm proposed in this thesis can effectively improve the classification performance.(2)For user multi-label classification,an account multi-label classification method based on overlapping community detection is proposed.Based on the heterogeneous network-based account single-label classification method,overlapping community detection skill is added.By performing overlapping community detection on users in the network and using the community detection results for user representation,it is possible to effectively evaluate the distribution of a user belonging to multiple labels and the similarity for users who are not directly connected.In the selection of multi-label classifiers,this thesis chooses ML-KNN multi-label classification algorithm to implement model training.In actual scenarios,by comparing with existing multi-label classification methods,it is found that the algorithm can be effectively applied to multi-label classification tasks of social media users.