Research On Reordering Information Streams Of Weibo With Multi-task Learning

Sina Weibo,as a typical representaive of social network,attracts a large number of users,because of its convenient platform,spreading rapidly and interactivity.At the same time,with the increasing scale of users,information in weibo also shows explosive growth.Every day,users will produce more than one hundred million data.For users,Sina Weibo,filled with all kinds of contents,makes the users spend a lot of time selecting useful information.At present,the microblogs are sorted by time in reverse chronological order.The latest microblog is in the front of others.When users read microblogs,they tend to read them from the beginning.The microblogs that users intersted in may be in an arbitrary position.The users need a lot of time to find them.In order to solve this problem,we research the user information and the weibo contents.We reorder the microblog information streams by using machine learning methods.Aiming at the problem of data sparseness,this thesis introduces the multi-task learning framework to improve the reordering model on the basis of learning to rank.The weibo contents and user interests both are important features of the model.In this thesis,we study the user and the microblogs' characteristics and extract the features of them.All these data and features are supported for the research work.In this thesis,we use machine learning to solve the problem of reordering the microblogs.We adopt the learning to rank training models on the training data.However,each user may have few microblogs to reorder.By multi-task learning,we can find the generalities between different reordering tasks and the combine these tasks with co-training and joint learning.The multi-task learning can model the individual behavior,and solve the sparseness of data in the individual modeling by using the global data.In order to establish the link between the tasks and improve the performance of the model,this thesis calculates the similarity of users,and realizes the reordering algorithm with the similarity measure.Different from the traditional document sorting,Sina Weibo,as a social network platform,can not be ignored the interaction between users.The use of Weibo makes the users better interacting with each other because user's behaviors have an impact on another one.Users have a certain similarity between each other.According to the common parts int the contents and structure,we quantify the similarity between users,establish the link between users,and design the users' similarity calculation method.On the basis of the multi-task reordering model,the reorderting algorithm is realized by introducing the users' similarity model.Finally,we perform a lot of experiments on the data.The results show that multi-task learning and user similarity can improve the performance of the reordering model.