Recommendation Cold Start Method Based On Multi-armed Bandits

With the rapid development of the Internet,a great deal of information is produced and disseminated every day,which leads to the problem of information overload.It is difficult for users to find the information they need on their own initiative.Therefore,recommendation system emerges as the times require,and it has become a research hotspot because it can effectively filter information and provide accurate personalized recommendation service for users.Especially in the fields of e-commerce,news,audio,video,advertising and others,personalized recommendation,which helps users get information more effectively,also increases the amount of browsing and sales for the platform,is playing an increasingly critical role.In this paper,we mainly study the problem of user cold start in recommendation system.For the scenario of only using commodity history text evaluation,we propose an improved bandits cold start recommendation algorithm,and propose a method of product potential feature extraction compatible with this algorithm.The traditional algorithm of bandits does not make full use of the feedback information from users and does not make use of features and user collaboration in the recommendation process.Specifically,this paper firstly proposes to extract the potential features of commodities by LDA topic model in the field of natural language processing and demonstrates the similarity between topic extraction and feature extraction based on commodity evaluation from the point of view of model principle,as well as the feasibility of this method,and verifies this method by comparing with other methods through experiments.Then,based on the traditional Bandits algorithm,we propose an improved Bandits recommendation algorithm which combines collaborative filtering and context information.Based on the contextual Bandits algorithm,the improved algorithm introduces fine collaborative filtering effect.When a product is recommended for the target user,the target user and its neighbor users will jointly determine the recommendation results,and through the neighbor user similarity weight factor to control the contribution of neighbor users to the recommendation,thus ensuring that the target user's own characteristics play a leading role,and the introduction of synergy effect to optimize the recommendation performance.After the recommendation is completed,the user features are updated according to the user's real feedback and the features of the recommended products,so that the user preferences can be quickly fitted with the product features within a limited number of recommendations to solve the user cold start problem.Finally,we use offline simulation online recommendation method to compare our algorithm with the latest methods in this field based on real data sets Delicious and Last.fm.Experimental results show that the proposed method based on LDA topic model can effectively extract commodity features by using commodity text evaluation and avoid the shortcomings of traditional feature construction methods.Compared with other topic models,the proposed method is more effective and suitable for user cold start problem.Experiments also show that the improved Bandits recommendation algorithm which combines collaborative filtering and context information can improve the recommendation effect and can obtain higher click-through rate and less cumulative error.