Research On The User Cold-start Problem Of Recommendation System

Personalized recommendation system is an important method to solve the overload of user information,which is widely used in the Internet field.However,with the existence of user cold-start problem,it will seriously affect the recommended results.At present,researchers have proposed a decision-tree algorithm based on question and answer to solve the user cold-start problem.However,the existing algorithm has the problem of low efficiency and long training time.How to solve these problems and further improve the efficiency and the accuracy of the algorithm,is still a subject worthy of study.This paper will do some research on how to improve the efficiency and training speed of Q&A decision-tress algorithm.First,a clustering-based QA decision tree algorithm CQA_DT(clustering-based Q&A decision-tree)is proposed to solve the problem of low efficiency of Q&A decision-tree algorithm.The basic idea is that by clustering the items,the item set is used to carry out question and answer.In order to improve the prediction precision of the algorithm,a latent matrix variable is added to each node in the decision tree to adapt the user's more precise way of Reply.Compared with the existing algorithms,CQA_DT algorithm can reduce the number of questions to answer for users and improve the prediction accuracy.Second,a distributed hierarchical iterative stochastic gradient descent method(DHISGD)is proposed to solve the problem that the training time of the Q&A decision-tree algorithm is too long.The basic idea is that in order to reduce the time of matrix decomposition in the training Q&A decision tree,the decomposition of the matrix is carried out by the distributed hierarchical iterative stochastic gradient descent method.In order to speed up the convergence rate of DHISGD,a SPCS learning rate update method is proposed.This method can accelerate the convergence speed without increasing the time complexity of single iteration.Compared with the existing distributed matrix decomposition algorithm,this algorithm can accelerate the matrix decomposition,and can reduce the shuffle amount in the matrix decomposition iterative calculation process.Third,in order to give those users who do not want to answer those questions or do not login personalized recommendations,we propose a Real-time recommendation engine framework.According to the user's real-time behavior,it can constantly update items nearest neighbor list and recommend similar items to those users.