Research On Personalized Malicious Comments Filtering Algorithms Based On Reinforcement Learning

Malicious comments are affecting hundreds of millions of Internet users,some of whom are depressed or even commit suicide because of malicious comments.There are two traditional algorithms to filter malicious comments.One is through simple keyword filtering technology to filter comments.Secondly,a general machine learning model is trained on the malicious comments dataset,and then the model is used to discriminate the comments.Neither of the above two algorithms takes into account user's personalized characteristics.Because different users have different backgrounds and personalities,their criteria for judging malicious comments are also different.Whether it is keyword filtering or training out of the general model is far from meeting the user's personalized needs.The main work of this thesis is to introduce user personalized profile into malicious comment filtering algorithm,and use reinforcement learning algorithm to construct learnable user profile.This thesis divides the malicious comment filtering algorithm into two sub-networks,one is feature extraction network,the other is user profile.In the part of feature extraction network,we experimented with a variety of choices,including models based on CNN,RNN and self-attention mechanism.In the user configuration section,because this thesis not only considers the differences between different users,but also considers the differences of the same user's evaluation criteria for bad reviews in different mood states,so this thesis proposes three methods based on reinforcement learning to construct learnable user profile.One is to use DQN algorithm or PG algorithm to fine-tune feature extraction network on user-related data,which makes network parameters contain personalized information.Secondly,user feature vectors are introduced,and DQN or PG algorithm is also used to adjust user feature vectors to enable the user vector to contain personalized information.Thirdly,regarding users as environment,comments as observation,malicious comments filtering algorithm as decision maker,user's feedback as reward,whether to show comments to users as actions,do MDP modeling.The interaction between users and comments is regarded as a sequential process rather than a one-step process,and the user's criteria for judging malicious comments are inferred by the user's feedback of comments.The model mentioned in this thesis achieves an optimal accuracy rate of.95 when it is not fine-tuned by reinforcement learning.After fine-tuning by reinforcement learning,it achieves a accuracy rate of.995.The experimental results show that the strategy proposed in this thesis is superior to the scheme without considering personalized features in the filtering task of personalized malicious comments.