Research On Cold Start Problem And Evaluation Method In Recommendation System

With the continuous development of the Internet and the accumulation of human knowledge,people are in an era of information overload.The traditional tool which can solve the problem of information overload can not meet the personalized needs of information retrieval.In this case,a recommendation system is proposed,which can be used to analyze the user's historical behavior records,and to make personalized recommendations for each person's historical preferences.However,most of the existing recommendation algorithms face the cold start problem.Cold start problem is that when a new user enters the system,because the system does not have the behavior records of the user,so the system can not recommend for that user.Most of the existing algorithms solving the cold start problem of recommender systems require the user's demographic information or social network information.Without user's demographic information or social network,the recommend system can not recommend for user properly.To solve this problem,this paper proposes a EGIF algorithm based on the combination of N arm slot machine model and human immune feedback model.The algorithm can quickly find the user's preferences without the user's demographic information or social network information.Finally,Monte Carlo simulation method is used to compare the EGIF algorithm with Softmax algorithm,Annealing algorithm and UCB algorithm.The results show that the EGIF algorithm is superior to Softmax algorithm,Annealing algorithm and UCB algorithm in solving the cold start problem.At present,most of the scholars use the offline metrics of recommendation algorithm,such as accuracy,recall and coverage to evaluate the performance of the proposed algorithm.But in the industry,people are more concerned about the online performance of the algorithm,such as CTR metric.But the CTR algorithm requires a long test process.In order to solve this problem,this paper proposes a linear regression algorithm based on recommendation algorithm.As the offline metrics of the recommendation algorithm is easy to be calculated,the off-line metric of the algorithm can be used as the input to the model,and the CTR metric can be predicted quickly.Experiments show that the algorithm we proposed can use offline metric to predict the metric of CTR.By continuously adjusting the off-line evaluation metric we can get more accurate prediction.