Individual Anomaly Detection And Emotional Modeling Based On Multi-source Data

Social media platforms contain a large amount of text data,including user feedbacks,comments,dialogues,opinions and other information,which is the basis of emotional analysis.Individual anomaly detection and emotional modeling are important components of emotional analysis and are also the emerging directions in affective computing.Individual anomaly detection mainly refers to the detection of abnormal emotional transfer patterns of users.Personality modeling is to quantify the relatively stable emotional patterns or personalities of users by mining the emotional transfer patterns in dialogue and interaction.At present,the individual anomaly detection and emotional modeling often neglect the emotional personality of the interlocutor while pursuing logic and accuracy.Based on conversational interactive text data of different platforms,this thesis summarizes the individual anomaly detection and emotional modeling of multi-source data(multi-platform conversational text data),including the concepts,backgrounds,relevant methods,research status and some shortcomings.Main tasks are as follows:(1)For individual anomaly detection,the emotional transition probability matrix and a tensor network model based on dialog interaction are proposed,and individual anomaly detection model based on tensor similarity is implemented.(2)Based on the traditional Markov Monte Carlo method,an E-MCMC algorithm is proposed for personality emotion simulation.The algorithm can simulate the samples corresponding to the probability distribution of conversational emotion transfer.(3)For personalized emotion generation,a GEN-MCMC algorithm is proposed,which implement dynamic transfer process of user's emotional interaction in dialogue and the generation of user's personalized emotional transfer sequence.(4)For personalized emotional guidance,a stimulus path sampling algorithm GUI-MCMC is proposed,which implements the framework of emotional guidance to guide user's emotion from a given emotion to a specified target emotion.Experiments show that the proposed tensor-based anomaly detection method can detect hidden anomalous emotions of individual and verify the effectiveness of the improved MCMC algorithms.Among them,the larger the convergence limit of EMCMC,the closer the sampling sequence and the probability distribution of source data is in a certain range.GEN-MCMC algorithm can effectively generate the user's personality emotional sequence,which performs better in consistency and diversity than the baseline model.GUI-MCMC algorithm can effectively guide individual emotion to the desired direction,providing a new idea and a feasible solution for emotional guidance,psychological diagnosis and treatment,chat robots and other research.