On Optimizing Cross-entropy Based Periodicity Detection Model And Its Application

Periodic pattern mining is a data mining method to find the recurrence of the sequence pattern of repetition in time series data.It is mainly used to describe the stable user behavior patterns.In many application scenarios,the periodic pattern mining of time series data is of great significance.Based on the Cross-Entropy based Periodicity Detection(CEPD)method,this paper has done some further improvement and research on the three problems of the CEPD method.And we make recommendation application combined with the user behavior in e-commerce and mobile internet environment.The main contents of this paper are as follows: First,a comparative study of the periodic detection method is carried out,and the theoretical time complexity of the CEPD algorithm is given.The CEPD method is compared with other three existing methods in terms of running time,resilience to noise and the accuracy of detected period in the synthetic data set and real data set.Meanwhile the theoretical time complexity of the CEPD algorithm is given.The feasibility of this method is proved from both the experimental and theoretical aspects.Second,the objective function of the CEPD method is optimized by introducing the regularization technique in machine learning,which solves the problem that the cross entropy decreases systematically with the increase of the division period and improves the feasibility of the method and the availability of the detected period.Synthetic data and real data are used for experimental verification.On a dataset with a distinct period,the model after adding the regularization term makes it easier to detect the appropriate period value.Third,after obtaining the period of user behavior,we make a recommendation application using the detected periodicity.Adding periodic information in recommendation introduces the time information of user behavior and depicts the regularity of user's own character.The CEPD algorithm efficiency analysis proves the feasibility of this method from the theoretical and experimental aspects.The introduction of the regularization technique solves the problem that the cross entropy increases systematically with the increase of the division period,which improves the feasibility of the CEPD method,and lays the foundation for the futhur application.The contributions of this paper are,first,we proved that the time complexity of the CEPD algorithm is better than other three algorithms theoretically,the feasibility of the periodicity method is proved from the experimental and theoretical aspects.Second,the CEPD method is optimized to solve the problem that the cross entropy decreases systematically with the increase of the division period after introducing the regularization technique in machine learning.And which improves the feasibility of this method and the feasibility of detected period.Third,combined with the periodic detection results of CEPD method,a recommendation strategy considering the periodicity of user behavior is constructed,which takes into account both the user's intention and recommended time.The effect of recommendation is improved to a certain extent.