Research On Weibo Rumor Detecting Algorithm Based On Ensemble Classifiers

Weibo is a popular social networking platform and it's diversity of information,the freedom of speech and the explosive speed of propagation encourage the emergence and diffusion of rumors.Therefore,research on automatically detecting of rumors on Weibo comes into being.Some achievements have been made among state-of-the-art rumor detecting algorithms while there are still some improvements to reach: the detecting accuracy as well as the early detecting accuracy of rumor detecting algorithms still can be improved;existing rumor detecting algorithms ignore the class imbalanced problem of the training data and the learner can't effectively learn the distribution of minority samples when the training data is imbalanced.Some related research has been conducted to solve the above two problems in this thesis.A model ensemble based algorithm GTB-RD for rumor detecting and a generative adversarial networks based over-sampling algorithm GAN-SMOTE for class imbalance problem are proposed.The main work of this thesis is summarized as follows:1.A model ensemble based rumor detecting algorithm is proposed.Most existing Weibo rumor detecting algorithms are based on a single predicting model,a model ensemble based rumor detecting algorithm is proposed to obtain a higher detecting accuracy.The proposed detecting algorithm GTB-RD combines the detecting results of multiple base detectors by gradient boosting method to form the result of the final predicting algorithm.2.A feature selection algorithm for GTB-RD is proposed and a number of new features are constructed and proposed.A higher detecting accuracy is obtained by using the proposed features and GTB-RD detecting algorithm.Most existing rumor detecting algorithms use many features based on propagation and comments which result in poor rumor early detecting accuracy.In this thesis,feature engineering is re-conducted by the proposed feature selection algorithm,the selected features are used as input for GTB-RD algorithm,experiment results on real world dataset show that the detecting accuracy and early detecting accuracy are improved.3.Generative adversarial network based algorithm for class imbalanced problem is proposed.Inspired by the generative adversarial networks,a generative adversarial network based algorithm GAN-SMOTE is proposed to estimate the distribution of the data and to generate synthetic samples.The GAN-SMOTE algorithm is consists of a generative model and a judging model,the generative model aims to estimate the distribution of the data and the judging model aims to evaluate the quality of the generated data from the generative model.The generative model and the judging model train alternately and the generated samples are used to balance the training data.Experiment results on several real-world datasets illustrate the effectiveness of the proposed GAN-SMOTE algorithm.