Research And Application Of Deep Matrix Factorization Method In Recommender System

With the development of the economy,the increasing variety of commodities can bring users a better experience,but it also brings problems such as information overload.It is the key for each platform to recommend the most suitable products for users to improve their dependence.The most important of these is the matrix completion technology,which is a recommendation system to predict user preferences for items.At present,there is a large amount of theoretical knowledge supporting the known matrix completion technology,but it still has certain limitations.It requires adding nonlinear operations to the matrix-filled model,since linear models cannot capture unstructured information in the real world.For example,the existing matrix completion models and matrix completion techniques are only combined with deep networks or filled with algorithms alone.They do not involve the integration of the recommended background,or the representation vectors and interaction representations of users and items which are only learned by nonlinear means without considering the important role of user and item bias in recommendation.In order to solve the above problems,we integrate matrix factorization technology with deep neural network,by learning latent factor representation vectors for users and items from matrices.At the same time,we learn user-item biases to build a deep matrix factorization module and train.This solves the problem that matrix factorization cannot capture unstructured information,and takes into account user preferences and item personalization in recommender systems.Simultaneously,we train with explicit data and uninteracted implicit data to convert the user's preference for items into implicit feedback,which perfectly utilizes the normalized cross-entropy of the heuristic loss function.In order to solve the situation in the sparse matrix,that some user information is abnormal and the data is insufficient or even missing.We propose a matrix factorization model based on generative data augmentation in this paper,inspired by the idea of generative adversarial and discriminative imputation of missing data,it takes the pretrained deep matrix factorization module as the initialization input of the model prediction part.Then,we continue to train the matrix factorization model using the generated data of the generative model.Lastly,the training results are used to assist the generator to generate data for the selected positions.The matrix factorization module and the generative model complement each other,which enables the prediction model results to learn more accurate user representations using the generated near-real data,so as to achieve the purpose of improving the effect of the model.Finally,we respectively apply the deep matrix factorization model constructed in this paper and the generative enhancement-based deep matrix factorization model to the public audience movie rating dataset and the real short-term user-visited urban function zones datasets.We perform matrix filling and predict user preferences for items on the two datasets,compared with other state-of-the-art methods,the two models proposed in this paper have achieved the best experimental results and application effects.Moreover,our model can still play a good role in real-world application scenarios such as function zone recommendation by analyzing the performance on the dataset of user access to urban function zones.