Hybrid Recommendation Models Based On Multi-source Heterogeneous Data

With the development of the Internet,the data in the Internet is growing rapidly,which makes it difficult for users to find the desired information.The emergence of recommendation system greatly reduces the burden of users to find information.With the abundant of data,recommendation system is gradually developing to combine multiple data.Single rating data can not adequately express users' preferences,While review data can supplement more user information,and social networks can reflect the similarity between users.However,there are few methods of personalized recommendation using machine learning technology combined with rating,review and social network,and there are some problems such as inaccuracy and difficulty in feature representation.Combining rating,reviewing and social network data,this paper proposes two personalized recommendation models for dealing with multi-source heterogeneous data:one is a multi-source heterogeneous data recommendation model based on traditional machine learning.The model combines rating,reviewing and social network data,and predicts ratings of items through community detection,feature extraction and regression algorithm.In the training phase,topic or word embedding model is used to extract the topic information of user review text,and community is divided for users by community detection algorithm,and then community model is trained by regression algorithm.In the prediction stage,the user features are sum by the review features they send out,and the business features are sum by the review features they receive.Finally,the features are input into the regression model for rating prediction.Model 2 is a multi-source heterogeneous data recommendation model based on deep learning.The model uses deep representation learning algorithm to extract rating and text features,and combines social networks to recommend personalized items for users.In the training stage,PV-DBOW is used to represent and learn the text features of user reviews and get the feature vectors of reviews,so as to get the text features of users and items;neural network is used to learn the rating features,so as to get the rating features of users and items;text features and rating features are fused in series to get the final user and item features;Pair-based learning is used to update the network parameters and get a better feature representation of users and items.In the prediction stage,the recommendation probability can be obtained by multiplying the user and the item features and ordering them in descending order,then the list of items recommended to the user can be obtained.In the experimental stage,the model compares the impact of different text processing methods on the accuracy of rating prediction,and compares with other relevant models.The experimental results show that model 1 can improve the accuracy of rating prediction.Model 2 first compares the impact of adding social networks on recommendation results,and then compares it with other related algorithms.The experimental results show that model 2 can improve the quality of recommendation results.The two models proposed in this paper use different algorithms combined with multi-source heterogeneous data to improve the accuracy of rating prediction and ranking recommendation respectively.