Preference-Feature Based Hybrid Recommendation Algorithm

Since human beings stepped into the21stcentury, information has been growing in anexplosive way that makes people surrounded by enormous amount of various kinds ofinformation. Sarcastically, though people are now handling with so many kinds of information,it is difficult for people to extract useful ones. People either benefit from the massiveinformation explosions or obtain the conveniences large amount of information had brought tothem, on the contrary, people are puzzled and befuddled by them. Only thinking to relieve thistricky problem that scientific research groups are trying their best to invest a method torecommend the most appropriate items to users, this method would provide a perfectexperience to those users who are not very clear of what they need or what they really want topurchase. And this problem solving and considerate method is Recommendation System.The recommendation systems are being comprehensively used by e-commerce websites.And with the growing demand from the users, the research of recommendation systems hasgradually become the dominant and pivotal project in the computer science research field. Thedominating techniques of recommendation systems are Collaborative FilteringRecommendations, Content-based Recommendations (CN), Demographic-basedrecommendation(DM), Knowledge-based Recommendation(KB) and Utility-basedRecommendation(UT). The Collaborative Filtering is commonly used due to itsrecommendation accuracy and easy deployment. It is, however, not an immaculate method. Ithas some drawbacks such as data sparsity, scalability, grey sheep problem and so on.Just because we are perfectly aware of the imperfections of collaborative filtering, I am goingto continue research and analyze CF algorithm in a more deep and thorough perspective, andwork out my own improved my own algorithm which is the user preference-based and itemcharacteristic-based hybrid collaborative filtering. This particular algorithm improves thesimilarity calculation method of user-item pair, also it would generate recommendations basedon the item characteristic values it extracts and user preference data. Using this algorithm, it ispossible to uplift the accuracy of each recommendation that is, in another word, improve theperformance of the algorithm significantly. At last but not least, this dissertation takes MovieLens data set as the experimental dataset and makes a comparison between my algorithm and the traditional collaborative filteringalgorithm. The experiment results would be shown through the evaluation metrics MAE andRMSE. As the results have shown that the user preference-based and item characteristic-basedhybrid collaborative filtering technique my thesis proposed makes great progress on real-worldperformance compared with the previous original collaborative filtering technique.