Cold Start Recommendation Method Based On User Preference Evolution

With the gradual improvement of relevant research on recommendation system,the problems also gradually reveal,and the cold start problem is the most common and difficult to solve.The existing mainstream solutions need to integrate context information or relevant domain knowledge during training.However,in the real cold start recommendation scenario,due to the user's personal privacy,it is often impossible to access additional auxiliary knowledge.This dissertation mainly studies how to use the user's interactive behavior sequence in the cold start scenario to mine the user's preference evolution law,and finish the user's cold start recommendation without more context information.The main work of this dissertation includes:(1)Dual task cooperative trainingThe attention mechanism is used to extract the effective subsequence of user interaction behavior sequence,so as to reduce the impact of invalid interaction in the sequence.The dissertation designs user interaction behavior sequence prediction task,the next recommendation task,and the user interaction behavior sequence generation task to capture the short-term sequence and long-term sequence dependency in the sequence respectively,so that the performance of sequential recommendation can be improved.Experiments show the effectiveness of establishing learning tasks for long-term and short-term dependence,and lays a foundation for the processing of user sequences in cold start recommendation scenarios.(2)Cold start recommendation method based on meta learningThis dissertation studies that the existing methods can't make reliable recommendation,due to the lack of effective auxiliary knowledge,and combines the sequential recommendation method to solve the user cold-start problem.In this dissertation,we propose a sequential recommendation method based on user preference evolution.In this method,we model the sequential recommendation as a meta learning problem to quickly adapt to cold start users,and construct a user preference evolution tree,simulate the user preference evolution process,and establish a meta learning task.Based on the dual task collaborative training model,we improve the model,and make it adapt to the cold start recommendation scenario,and capture the user's personal conversion mode(i.e.short-term dependence)and the preference evolution law of the user group(this law represents long-term dependence in the cold start scenario),so as to improve the cold start recommendation performance.The sequential recommendation method based on user preference evolution can quickly adapt to new users and provide fast and effective recommendation with few of interactive information.(3)Experimental verification and analysisThe dual task collaborative training recommendation method is verified on Taobao advertising data set and Movielens dataset.The experimental results show that the dual task collaborative training can effectively improve the performance of recommendation,but only relying on a single task can't achieve the best effect.Using electronic product recommendation dataset electronics,film recommendation dataset movie and Book Recommendation dataset book,the sequential recommendation model based on user preference evolution and the comparison method are experimentally verified.The experimental results show that adding the learning of the evolution law of the global preference of the user group to the model will optimize the recommendation results.In addition to considering the transformation mode of user behavior,the model of user preference evolution tree for user group can still achieve the best performance.With extensive experiments on electronics,movie and book,we verify effectiveness of the evolution law of the global preference of the user group in cold-start sequential recommendation.In addition to considering the transformation mode of user behavior,the establishment of user preference evolution tree for user groups helps to improve the accuracy of recommendation.