Study On Recommendation Algorithms Based On Item Interaction Characteristics Of Users

With the popularity of smart phones and the rapid development of mobile Internet in recent decades,people are suffering from the problem of "information overload" while enjoying the convenience of "information abundance".As an important tool of information filtering,recommendation system has been deeply studied and widely used in academia and industry.Based on the characteristics of the user and the items and the interactions between them,the recommendation system recommends the items that the user may be interested in.The current mainstream recommendation algorithms are completely from the user's point of view,committed to providing accurate and diverse item recommendation list.However,a recommendation system usually includes users,the platform that provides the recommendation system,and the item provider.A good recommendation system should consider the different needs of all three participants at the same time.Considering the different requirements of the recommendation platform and the item provider,this thesis,based on the user's item interaction characteristics,conducts research on improving the time efficiency of the recommendation algorithms,improving the average of the numbers recommended and the purchase opportunities of different items.The user's item interaction Characteristics refers to the feature of the collection of items that the user interacts with in the historical record,such as the number of items purchased by the user(user degree)and the user's preference for items with different popularity.In this thesis,the main recommendation algorithm is improved by mining the users' item interaction Characteristics,so as to meet the needs of different participants in the recommendation system.The main work and contributions of this thesis are as follows:(1)An efficient core user extraction method for recommendation system is proposed to improve the operation efficiency of the recommendation algorithm.As the most commonly used recommendation algorithm,one of its core steps is to calculate the similarity between users/items.In large-scale systems,however,the large number of users/items makes calculations of similarity time-consuming.The existing research shows that the efficiency of the algorithm can be effectively improved by only using the information core formed by a small number of users to make recommendations.In order to reduce the computation time of the existing extraction methods for information core,this thesis proposes two new generation methods for core users,the public interest superposition and the iterative descending voting.Compared with the existing best methods,the two methods have similar accuracy performance.The accuracy of the original method can be more than 90% with the information core composed of 30% users.At the same time,depending on the number of users and the number of items,the time efficiency of generating the information core can be up to thousands of times higher than the existing methods.In addition,the relationship between different information cores is analyzed to explain the mode of action of information core.(2)The ranking aggregation method of recommendation results is proposed to improve the average degree of item recommendation times.In order to avoid recommending a small number of popular items to the majority of users,item vendors want a balanced recommendation of different items.Existing studies have shown that reverse recommendation from the point of view of items can effectively increase the number of times a large number of niche items are recommended.Therefore,from the point of view of both users and items,this thesis proposes a reordering method of recommendation results.For the user-item recommendation score matrix generated by the existing recommendation algorithm,column ranking and row ranking are carried out respectively,and a linear aggregation method of row ranking matrix and column ranking matrix is designed to obtain new recommendation results.Experiments on real data sets show that the ranking aggregation method can greatly improve the accuracy of recommendation results and the average degree of item recommendation times.In addition,considering that different users have different preferences for item popularity,this thesis also proposes a ranking aggregation method of parameter personalization,so as to produce recommendation results of different personalization degrees for different users.(3)A new equity evaluation index is proposed to measure the average degree of the probabilities of items purchased.Item providers hope to improve the sales of long-tail items with the help of recommendation system.Among them,the most simple and direct method is to increase the recommendation times of long-tail items.The commonly used index to measure the ability of recommendation algorithm is the Gini coefficient of the number of recommended items,also known as coverage fairness.However,being recommended is not the same as being purchased,and the probability of different items' recommendation being converted into items' purchase is not the same.It is generally believed that active users who interact with the recommendation system frequently have higher trust in the system and higher probability to buy the recommended items.Therefore,this thesis proposes the purchase power of users' purchase of recommended items which is positively correlated with the activity of users.Based on this,the Gini coefficient of the probabilities of items purchased is calculated to measure the fairness of the sales of different items brought by the recommendation system.Through the analysis of various diversity evaluation indexes on different data sets,we show the uniqueness of the proposed sales fairness index and the coverage fairness index.