Research On Evaluation Method And Optimization Algorithm Of Diversity In Recommendation Results

Recommendation system,as one of the important tools for mining user's historical data,has been widely used in commercial platform.However,in the past research,accuracy is the most important performance that recommendation system focused on,which led to a large number of hot commodities piled up in front of users.The overall recommendation results are homogeneous,so diversity has become one of the important evaluation indicators to measure the quality of recommendation algorithm in the industry.At present,most of the methods for evaluating the diversity of recommendation list or recommendation system focus on the process of simply calculating the average of the dissimilarities between items or lists,which is not enough to fully reflect the details about the diversity in recommendation(list)system.At the same time,in the process of recommendation,most algorithms only consider the score,not the attribute information of users and items.When recommending with score,we only care about the value of score,ignoring its ranking in all the scored users.Based on the diversity of recommendation system,this paper proposes two improved methods and corresponding indicators for list-level and system-level evaluation.To solve the problem of attribute information and ranking,a diversity optimization recommendation algorithm combining attribute embedding neural network and re-ranking is proposed.The main work of this paper is as follows:(1)In view of the impact of project location and popularity in list-level evaluation,a diversity evaluation method combined with normalized cumulative performance-location factor and NCIP index are proposed.NCIP index absorbs the idea of novelty index and NDCG index,and combines the normalized item popularity and the location information in the recommendation list as the weight of the dissimilarity between the items in ILD index.By using the number of categories in the recommendation list as a measure of diversity,it is verified that NCIP index can fit the curve of the number of categories in recommendation list under different recommendation lengths.(2)In view of the diverse demand of user groups with different activity in recommendation system,a diversity evaluation method,combining diversity demand factor and accuracy,and a PD index are proposed.The PD index fuses Precision index and user-level Hamming distance linearly by normalized user activity.By taking the category hit rate of recall as the basis of diversity consideration,it is verified that PD index can predict the category hit rate of recall under different user activity groups.(3)To enhance the diversity of recommendation algorithm,a diversity optimization recommendation algorithm combining attribute embedding neural network model and re-ranking is proposed.The algorithm is divided into two parts:attribute embedding neural network recommendation model(EA-CNN)and re-ranking(FBRR).EA-CNN is a convolutional network composed of unstructured title attributes and other structured attributes.FBRR introduces the idea of backward recommendation into recommendation set that EA-CNN generates.Combining with TOPSIS method,FBRR ranks recommendation set in two-dimensional recommendation space and reconstructs recommendation list.Compared with the traditional recommendation algorithm,the recommended list generated by the EA-CNN+FBRR recommendation algorithm can improve Coverage,ILD and NCIP indexes by at least 10.97%,8.88%and 6.95%,respectively on the premise that the Precision and HD indexes are basically unchanged.