Research On Recommendation Via Data Debugging And Curriculum Adversarial Training

Recommender systems are becoming more and more popular in people's lives and have become a research hotspot.To accurately model users' preference,a large amount of existing research work mainly focuses on building more complex recommendation models or incorporating side information such as user social relationships and item attributes.In contrast,there are relatively few research work considering the use of original data and the control of model training,and these factors also have an important impact on the performance,which is mainly reflected in the following two aspects.First,the training data of the recommender system is contributed by users in the system,and there will inevitably be noisy data.Compared with those accurately labeled training data,data quality issues may affect the performance of the recommendation model.Second,the existing research work shows that the quality of the user/item embeddings of many recommendation models trained with normal training process is not high,which easily leads to insufficient generalization ability of the model.In response to data quality issues,existing work mainly uses some anomaly detection methods to help filter noisy data.This type of work does not explicitly consider the impact of the detected noise data on the recommendation performance.Aiming at the problem of model training,existing work shows that applying adversarial training to the training process of some recommendation models can improve the generalization ability of the model.However,the existing work uses adversarial training with a fixed perturbation strength,and the learning curve is not smooth enough.Therefore,for data quality issues,this thesis explicitly considers the impact of training data on the recommendation performance of the model,while for model training issues,this thesis considers adversarial training with multiple perturbation strengths inspired by curriculum learning.Specifically,the main content of this thesis is as follows:? Recommendation based on data debugging.This thesis proposes a data debugging method for recommender systems,which can identify the data in the training data that affects the overall recommendation performance,and then edit these data to improve the performance of the trained model.The experimental results show that on the two rating datasets,the proposed method in this thesis can significantly improve the performance by editing the original training data.? Recommendation based on curriculum adversarial training.This thesis proposes a recommendation method based on curriculum adversarial training.Adversarial training with different perturbation strengths is sequentially performed on the training process,which can exert the effect of adversarial training better,and the proposed method improves the performance.The experimental results show that on the two implicit feedback datasets,the proposed method in this thesis can achieve better performance compared to adversarial training with only one strength.? A prototype recommender system for movie recommendation.To verify the rationality and feasibility of the above techniques,we also design and implement a prototype recommender system based on the above-mentioned techniques for movie recommendation scenarios.