Research On Content-based Personalized Recommendation Systems

In this thesis,we conducted a research on personalized recommendation systems. The main focus is to select the content-based recommendation algorithms. A content-based recommendation algorithm utilizes the products features that a target customer likes to find top-k most similar products among all the products to generate recommendations[1]. Then, in order to improve the recommendation precision of the traditional content-based recommendation systems, we try to combine k-means with the traditional content-based method and propose a new recommendation algorithm which implements cluster algorithm on the products features vectors that the user is interested in and find the closest products to each cluster center to do recommendation.These two algorithms only exploit the product information that the target customer likes and then according to these products to find similar products to generate recommendation. In this thesis, Naive Bayes is introduced as a recommendation algorithm. This method uses the products that the target customer likes and dislikes to learn a Naive Bayes classifier. Then, according to this classifier, the probability that the customer likes one product can be calculated to recommend the products with the highest probability values.Finally, in the evaluation part, the personalized recommendation systems based on these three different recommendation algorithms are tested on the real data sets. The data are real user transaction records and products information provided by an e-commerce company in UK. Before the experiment, we introduce how to transfer these data in the mongo database into the format that can be used by recommendation algorithm directly. Then, we give a comparison and analysis of these three recommendation algorithms, i. e., a traditional content-based recommendation algorithm, a method combining with k-means and a recommendation algorithm based on Naive Bayes. The experiment results show that the recommendation algorithm combining k-means improves the recommendation precision of the traditional content-based method, while the performance of the recommendation algorithm based on Naive Bayes is the best among three methods.