The Personalized Recommendation System Based-on Online Reviews

With the rapid development of the Internet,mass network information plays a more and more important role in people’s daily life.Especially in the field of e-commerce,consumers increasingly incline to online shopping,which result in a large number of product information and comments.It may greatly enhance the consumer’s shopping experience if we get valuable content from the reviews on products.What’s more,it can also promote the turnover rate of goods.In a word,it has important significance both in academic and business circles.The recommendation system builds a model between the user’s historical behavior and the related information of the products,so as to recommend products that may attract the users.In brief,in the practical application,it achieves business growth by recommending products that users may be interested in.The traditional content-based recommendations mainly recommend according to the similarity,which is calculated between products that users have bought and the rest.In this paper,we not only consider the objective data of the product,but also take the reviews and other information into account,which improves the accuracy of the recommendation system.In this paper,we firstly need to collect the information of the products by web crawler,and to do preprocessing work such as word segmentation,which gives a set of feature words.Because of the large number of feature words,this paper applies the improved LDA topic model and TF-IDF algorithm to multi-granulation feature dimension reduction,then by which mining the theme information and calculates the distribution probability of the text on each subject.Combined with the user-interest model,Euclidean distance formula is used to calculate the similarity among the texts.Besides,a sigmoid function is adopted to improve the transition between the user comment and the user attributes information under the cold start state.Finally,a recommendation list of products with high similarity will be provided for the users.To verify the proposed model,books information from Amazon site was selected as the experimental data.The experiment also discusses the effect of probabilistic distributions on latent topics when changing the number of the topics.In addition,the recommendation performance indicators,including the accuracy rate,recall rate and F-Measure index,are evaluated for different feature items and different featureextraction methods.The experimental result shows that,compared with traditional recommendation method,the improved one in this paper provides users with more accurate recommendations,which considers the information of comment text.