Bi-LSTM Commodity Recommendation System Based On Word Embedding

Recommendation system is widely used in e-commerce,online news,medical treatment,new media and other fields.This technology of finding satisfactory goals for users from massive data has developed rapidly in recent years.The commodity recommendation system studied in this paper establishes a user preference model by analyzing user comments on commodities and purchase related behavior,so as to recommend related commodities.In preference modeling,in addition to the analysis of text content,the weight of emotion analysis is supplemented to improve the accuracy.The main research contents are as follows:1)Emotion classification algorithm based on word embedding.An improved word embedding model is proposed.TF-IDF algorithm is used to obtain the word vector of each statement after text preprocessing,increase the weight value of the word vector,construct the word distribution characteristics of the weighted word vector,and model the user's comments.Experimental results show that the effectiveness of the algorithm is verified by comparing with different word representation models in emotion classification task.2)Product recommendation based on sentiment information.An improved user rating matrix model is proposed.A collaborative filtering algorithm is used to calculate the user similarity by adding the user's emotional tendency value to the product,and the user rating is predicted by the user similarity and user product matrix,and finally the user rating prediction matrix is constructed.The experimental results show that the collaborative filtering algorithm based on sentiment information is more accurate than the traditional collaborative filtering algorithm in predicting results.3)Product recommendation based on user behavior sequence.An improved user purchase behavior model is proposed.The Markov chain is used to establish the user's historical purchase sequence model,and the sequence model is trained into the neural network to obtain the recommendation result.The experimental results show that Bi-LSTM model has higher F1 than CNN,XGBoost and SVM.The experiments show that the method of using weighted word vectors can solve the problem that word embedding models cannot distinguish when dealing with semantically similar word vectors;the user rating matrix generated by combining sentiment information can improve the accuracy of predicting rating results;the recommendation model based on user behavior sequences can capture users' historical purchasing behavior to make better recommendations for users.