Design And Implementation Of Evaluation System For The Research Of Novelty And Diversity Recommendation Algorithm

The current mainstream recommendation algorithms are all committed to improving the accuracy of the recommendation,ignoring novelty and diversity,and recommending items similar to the user's historical preferences.After investigation,these projects are often recommended by popular or users.Users may be dissatisfied with the recommendation results because they have already known these items.In addition,niche projects will become more unpopular because it is difficult to enter the recommendation list,and the organizers of these projects will be dissatisfied with their items not being included in the recommendation list.This phenomenon is called the "Matthew Effect".In addition,the research on novelty and diversity recommendation algorithms requires the calculation of accuracy,novelty,diversity,coverage and other evaluation metrics and experimental comparisons for some benchmark algorithms,witch adds redundant workload to scholars.In order to solve these problems,the thesis proposes an improved random walk recommendation algorithm based on novelty factor.The proposed algorithm is uploaded to the recommendation algorithm evaluation system and compared with recent benchmark recommendation algorithms.The main contents of this thesis are as follows:(1)A formal definition of novelty factor is proposed.The thesis proposes the conception of novelty factor by considering the activeness of the candidate set items and the similarity of the the items' topic,and based on this conception,a formula for calculating the novelty factor is given.Finally,it is verified on the Douban and MovieLens datasets that the novelty factor can improve the novelty of the original recommendation algorithm without affecting the accuracy of the recommendation.(2)In order to improve the diversity performance of event recommendation,and to alleviate the problem of lower accuracy caused by the cold start problem of event recommendation,the thesis proposes a random walk recommendation model GT-RW based on group relationships and event topic.This method adds the event group(host)and event topic context information to the energy diffusion bipartite graph,and uses an improved random walk algorithm for model training.Finally,the validity of the proposed algorithm is verified on the Douban and Meetup datasets.(3)In order to improve "Matthew Effect" in GT-RW,the thesis proposes an improved random walk recommendation algorithm based on novelty factor NGT-RW based on the previous research points.The last step was improved,so that the energy diffused to the user dimension is diffused according to the user's preference for popular items,the energy diffusion weight of those entities that like to participate in popular events is appropriately reduced,and the probability of niche activities being recommended is increased The prediction score of the candidate set is incorporated into the novelty factor for event recommendation.Finally,the effectiveness of the proposed algorithm is verified on the Meetup dataset.(4)In order to facilitate the research of the first three research points,the thesis designs a performance evaluation system for recommendation algorithms,which includes four modules:experimental preparation,performance evaluation,experimental record retrieval,and experimental result display.Using the thread pool technology,multiple users can conduct multiple experiments at the same time,and provide multiple evaluation index calculation services,as well as retrieval and display services of experimental records.Finally,the system was tested for function and performance.