Research On Deep Learning Recommendation Algorithm Based On BERT

The recommender system is an important method that can solve the problem of information overload,it can help people find information that may be of interest from many complex data.However,the recommended technology still has many historical problems,such as data sparseness and cold startup.In order to solve these historical problems,it has been to be a very effective method to incorporate item description information in the recommender system.However,in the current research of recommender systems,the features of text content are usually filled into collaborative filtering algorithms to alleviate the problem of data sparseness.When extracting text content,most researches use the traditional text feature extraction method,which can only analyze and calculate a single word,but cannot effectively combine the context in the sentence.This thesis proposes to apply BERT to extract text features,which can deeply mine the semantic relevance in sentences.This thesis mainly uses the natural language processing model BERT built by the neural network to model the text information.This method replaces the original one-hot vector with the description text information of the item as the input of the item,and then the BERT model performs feature extraction on the text vector matrix.The extracted item feature vectors and user feature vectors are sent to a matrix decomposition machine learning model for training,so as to realize the interaction of the two from time to time to form a new latent vector.Finally,the latent vector is learned through the neural network to obtain the predicted score,and then the back propagation training model is calculated with the actual score to calculate the loss function.Finally,in order to verify the recommendation performance of the BERT-based deep learning recommendation model constructed in this paper,we tested on three real datasets,MovieLens 1M,MovieLens 10 M and Amazon Instant Video.The RMSE values were 0.7667,0.7758,and 0.7803,respectively.Experimental results show that the DeepBert model proposed in this thesis has good recommendation performance in the score prediction task based on item description text information.