Point Process Based User Return Time Prediction

With the coming of Internet era,the Internet technology is also constantly changing.Endless software makes the information on the Internet more and more.At the same time,how to make good use of this information has become an important issue for companies.User return time prediction is one of the essential research direction.Accurate prediction help companies take effective measures to retain customers who may be lost and make some personally recommend to the users who have strong desire to buy in the near future to maximize profits.In academic circles,user return time prediction has become one of the research topics for many experts and scholars.There are two common way,the one is based on traditional machine learning methods,the other is related to point process.The traditional machine learning based method regard this problem as a common classification or regression problem without considering the sequence of events.Because of the lack of information about the sequence of events in modeling,the prediction accuracy of this kind of method is not high.The second method is point process method.Point process modeling is a common way for time series modeling.The core of this method is to model the conditional intensity function.Poisson process and Hawkes process are two common point process methods.Poisson process is a relatively simple method,while Hawkes process considers the influence of historical events on current time based on Poisson process.However,this kind of method ignores the influence of other users or commodity features on user return time prediction.Based on the above facts,this thesis carried out the following work:1.In this paper,a point process model considering the user preference factor and price factor is proposed.The model makes use of the user preference factor and the very important price factor,which is an essential element in the process of commodity trading.And this model also makes use of the advantage of point process in time series modeling,which improves the accuracy of user return time prediction when the maximum return time is short.2.On the bases of considering user preference factor and price factor,aim at the low accuracy when the maximum return time is long.This paper add the social factor to the model,which improves the prediction accuracy when the maximum return time is long.3.On the bases of considering the price factor and social factor,this paper talk about the way that three factor influence the performance.By comparing experiments result between one linear model and two non-linear models and considering the common model performance,this paper makes the conclusion that the influence three factors on user return time tends to be independent.However,it is not sure that there may be a better nonlinear model.4.Experiments are conducted on Amazon and Online Retail datasets to verify the Performance of proposed models.