Research On Abstractive Tibetan Summarization Based On Pre-training Model

With the rapid development of the Internet,people have entered the era of big data with information overload.In this context,recommendation system plays an increasingly important role in our life.The recommendation system needs to use a large amount of user data and inevitably involves the privacy of user data.Solving the problem of user data privacy in the recommendation system is not only to protect the user privacy,but also to help the service provider to achieve benefit growth the condition of service quality and user loyalty.During the development of recommendation systems,many important models,such as graph neural network and multi-task model,have appeared.However,while they are widely used in recommendation systems,the work of privacy protection is not fully considered.In this thesis,I studied the problem.Firstly,an edge-level graph neural network system based on differential privacy is designed for the graph neural network in recommendation system.In this model,by focusing on the friends or similar relationship between users and combining with the needs of attackers,appropriate assumptions are made to overcome difficulties that the graph neural network are sensitive and easy to be affected and improve the balance between the privacy protection and model performance of the current graph neural network based on differential privacy.Through the evaluation of three common graph neural networks,the proposed graph neural network based on differential privacy can not only guarantee the performance of node classification tasks,but also guarantee the privacy of graph edge information.Secondly,aiming at the multi-tasking model in the recommendation system,this thesis designs a data-level multi-tasking model based on differential privacy.The multi-task model can improve the performance of the model by simultaneously predicting the tasks with strong correlation,sharing part of the information,and adjusting the weights among different tasks.However,there are some risks of model feature leakage in this process.Therefore this thesis proposes a multi-tasking model based on differential privacy.In the process of task weight adjustment,the model automatically adjusts the noise to ensure the privacy of the model.Compared with the traditional differential privacy multi-tasking model,this model does not need to calculate the task correlation process,and meets the requirement of task weight adjustment at any time.