Deep Recommendation Algorithms Based On User-adaptive Strategy

Recommender systems have become a very prosperous research field in recent years because of alleviating the information overload problem caused by the explosive growth of information in the big data era,which have also attracted much attention from the academia and industry.At present,recommender systems have achieved rapid development with the help of deep learning techniques.In particular,deep recommendation algorithms can easily model the users' preferences with the powerful feature extraction and feature modeling ability of deep neural networks,so as to provide users with accurate recommendation services,which has very important practical significance and broad application prospects.However,the existing deep recommendation algorithms often need to stack very deep network depths to achieve a better recommendation effectiveness,which is a big overload for computing resources and practical application deployment.In addition,when the users' historical behaviors are limited,it will cause the user cold-start problem,and the conventional deep recommendation algorithms will struggle to play a role.Moreover,artificially designed fixed network structure greatly limits the scalability and diversity of deep recommendation models,and the models' unified treatment of all user inputs can not really highlight the characteristics of personalized users,and have a great impact on the recommendation effectiveness and efficiency,which has become an urgent pain point to be solved in the field of recommender systems.This thesis aims to study the deep recommendation algorithms based on useradaptive strategy,which truly treats each user independently and personally,so as to achieve user-adaptive,fast and accurate recommendation goal.The main work and contributions of this thesis are summarized as follows.1.Proposing a deep recommendation algorithm based on the user-adaptive network depth selection strategy.Aiming at the problem that the network depth of the existing deep recommendation models are too deep and all user inputs are treated uniformly,resulting in sub-optimal recommendation effectiveness and efficiency,a useradaptive network depth selection strategy is designed,which can adaptively select the network depth according to different user inputs.In such a way,it can reduce the calculation cost of the model and accelerate the overall inference process to a large extent,which could provide users with fast and accurate recommendation services.2.Proposing a cross-domain deep recommendation algorithm based on the user-adaptive fine-tuning strategy.Aiming at the problem of the user cold-start dilemma when there are limited users' historical behaviors,a user-adaptive fine-tuning strategy is designed in the pre-training and fine-tuning paradigm solution,which can adaptively fine-tune pre-trained model parameters based on the personalized user inputs during fine-tuning process.Through such a way,the finally learned model can achieve good performance in the cross-domain deep recommendations,which alleviates the problem of accurate recommendations in user cold-start scenarios.3.Proposing an adaptive deep recommendation algorithm based on the knowledge distillation and neural architecture search.The deep recommendation models involved in the previous two parts are too deep and large,and the artificially designed fixed network structure greatly limits the scalability and diversity of the models.Here,through the ingenious combination of knowledge distillation and differentiable neural architecture search,a general adaptive deep recommendation framework is designed to provide a new design concept and paradigm for the recommendation community,that is,a smaller,faster and more flexible deep recommendation framework is designed,while maintaining excellent recommendation performance,which could promote the better development of the field of deep recommender systems.