Multi-source Social Media Event Classification Using Network Representation Learning

Nowadays,Internet has become one of the necessary tools for people.More and more people are sharing what's going on around them through social media(such as Weibo,friends circle,etc.).As a result,many useful information can be obtained by analyzing and tapping large amounts of data posted by users on social media.By analyzing the social media data from different sources,this thesis classifies the events that are taking place in the world today.There is one of the biggest problems with different source data as the heterogeneity of data.In this thesis,a network topology diagram is constructed by using the relationship between multi-source event data.Then,we use an appropriate network representation learning algorithm to learn the nodes of the network topology diagram formed by the event relationship for obtaining the features between events.Finally,the feature representation of the nodes in the graph is obtained,and the mainstream machine learning method is used to classify them.Specifically,the network topology diagram is built by analyzing the social media event data of multiple sources for extracting common features between events.Then,we use these features to perform relational operations on events to obtain a network topology diagram of the link between multi-source social media events.In addition,the network representation learning algorithm combines the current mainstream network topology graph representation learning algorithm,i.e.,Node2 vec network representation learning algorithm,SDNE network representation learning algorithm and HOPE network representation learning algorithm.These three networks represent the vectors learned by the learning algorithm as the basic vectors to combine to obtain a vector better than the monomer pattern.Finally,we use the stacking model to fuse multiple machine learning algorithms to classify the vectors.In order to evaluate the model presented in this thesis,we collected and experimented with a social media dataset(Flickr-Wiki-YouTube event dataset)containing three different sources.Through a number of experiments,the effectiveness of the network topology charting learning method,the multi-vectors combination and the stacking method is verified.