Research And Implementation Of Personal Data Platform Recommendation System Based On Machine Learning

With the development of information technology,people are increasingly using the Internet to get news,shop,watch movies,and so on.With the rapid increase in the amount of data in the network,the recommendation system has become an important method to solve information overload.At the same time,the “Personal Data Management Platform” crawls the information of users on various platforms and has a large amount of data.How to use this data to accurately recommend users is also the research field of recommendation system.Currently,the most widely used recommendation algorithm is the collaborative filtering algorithm.However,the algorithm also faces many problems,such as the similarity model is relatively simple,the prediction process does not consider the user preference model,and there are potential performance problems when the project increases.Based on the above problems,this paper studies the similarity model,user preference model and expansiveness problem,and proposes an improved collaborative filtering recommendation algorithm,which has achieved the following main research results:(1)On the similarity measurement problem,a new measurement model is proposed.The model is inspired by the word embedding idea in the NLP field.By mapping the co-occurrence information of the item and the information of the item to the low-dimensional vector space respectively,two vector representation methods for the item are obtained.Finally,the two representations are combined and the similarity is calculated by weight.In the co-occurrence of the item,the f-item2 vec model is proposed,which introduces the item scoring factor to increase the similarity of the highscoring item.For the information of item,the doc2 vec is used to train the vector.Compared with the existing similarity model,the proposed model in this paper can not only capture the scoring features,but also capture the co-occurrence features and content features,and the results are more accurate.(2)In the prediction process,the user's preference model is introduced.This paper proposes a user preference model based on long-term and short-term interests.The model will be the user's preferences into short and long term two parts,calculate shortterm interests and long-term interest in weight,and finally fusion generated preference weights.After introducing the user preference model,when the user's historical data is used to predict the current score,the item with the same kind as the current forecast item accounts for a larger proportion,and the predicted result is better than the traditional method.(3)A recommendation method based on item clustering is proposed to solve potential performance problems.Specifically,after obtaining the item vector by the similarity model,the clustering algorithm is used to cluster the items,and the items in the unified clustering with the items to be predicted are loaded in the prediction score,and the nearest neighbor is found in the cluster set.And predict the score.This method avoids reading all item data when searching for the nearest neighbor,saving both time and memory overhead.