Research On Recommendation Algorithms And Evaluation Measures In Multi-stakeholder Environment

With the development of the Internet and the popularization of intelligent terminals,the information on the Internet has grown exponentially.In order to alleviate the problem of "information overload" and help users efficiently obtain the required information,recommendation systems emerge as the times require.The current mainstream recommendation systems are committed to improving the accuracy of recommendation and optimizing the user experience.However,there are other participants such as platforms and item providers in the actual recommendation scenario.The interests of participants other than users are not proportional to the accuracy of the recommendation system,so too much emphasis on the accuracy of the recommendation results will lead to damage to the interests of participants other than users.In order to alleviate the conflict of interests of multi-participants,thesis improves the existing algorithm,and improves the diversity of recommendation results while ensuring that the interests of users are not lost or slightly damaged,thereby ensuring that all participants in a multi-participant environment Interests.The research results show that the user's click behavior is not only affected by interest factors,but also by conformity factors.Aiming at this phenomenon,thesis improves the existing random walk model and latent factor model,and corrects the popularity bias problem in the recommendation results.The results show that the improved algorithm alleviates the phenomenon of popularity bias and improves the diversity of items in the recommendation results.The main work and contributions of thesis are as follows:(1)Two concepts of user-item interaction reliability and user conformity are proposed.These two concepts are quantitatively introduced into the recommendation algorithm based on random walk,and the way of energy random walk in the algorithm is changed.For users with different degrees of conformity,the way of energy transmission is no longer equal,but preferred.Experiments show that the method proposed in thesis can effectively alleviate the phenomenon that popular items have high scores and a large proportion of recommendations in the random walk model,and improve the diversity and fairness of items in the recommendation list.(2)The training sample generation method of the latent factor model DICE is improved,and the error problem of the interest partial order relationship and the conformity partial order relationship between positive and negative samples in the training data set is solved.At the same time,it innovatively proposes a dynamic representation method of item crowd vector,which solves the problem that static crowd vector is not suitable in actual recommendation scenarios,and further improves the accuracy of recommendation.In addition,by reducing the influence of conformity vector on click behavior and improving the role of user interest in recommendation,it also alleviates the popular bias problem of recommendation and improves the diversity of recommendation results.(3)Propose a comprehensive index to measure the accuracy and diversity of recommendation.The diversity of recommendation results and the accuracy index are contradictory,and improving one tends to reduce the other.thesis proposes a comprehensive index that can simultaneously evaluate the diversity and accuracy of the recommendation results,so as to better measure the protection of the interests of the recommendation algorithm for all participants in the multi-participant environment,and also provide recommendations for decision makers of the recommendation algorithm.provides a criterion for selecting parameters.