手机用户上网时段的混合Markov预测方法

手机用户上网时段研究与预测对手机用户行为与模式分析、网络服务内容设计、网络黏性与心理、移动互联商业智能等具有重要意义。本文结合Markov模型和关联规则模型,提出一种手机用户上网时段的混合Markov预测方法——Lift-Markov（LM）方法,并采用中国某城市4G手机用户流量上网产生的流量收费数据进行实验验证与分析。研究发现：该实验区域37.66%的手机用户个体存在明显的以天为周期的周期性特性;本文所提出的LM方法在10、20、30、40、50、60 min间隔时的平均预测准确率都优于Markov模型和Mostvalue模型,其中在60 min间隔时能达到79.75%的平均准确率,优于Markov模型（74.64%）和Mostvalue模型（64.44%）;LM方法的预测准确率分布相比于其他2种模型都要窄,而且密度分布峰值最高、标准差最小,说明本文方法对人群的上网时段预测准确率较为集中与稳定,具有较好的预测性能。

A mixed Markov Method to Predict the Surfing Time Period of Mobile Phone Users

In recent years, big data of mobile phones has become a great data source for researches and applications. It has been widely used to understand the human behaviors in cyberspace space. Researching and forecasting the surfing time of mobile phone users have great significance for analyzing mobile phone users’ behaviors and patterns, designing network service, and understanding the relationship of surfing behaviors, website stickiness, users’ psychology, mobile Internet intelligent business. We proposed a mixed Markov method (Lift-Markov method. LM), combining the traditional Markov model and association rule model, to predict the surfing time period of mobile phone users. A dataset of surfing records of 4G mobile phone users collected by Hubei Mobile within twenty days is used to demonstrate the capability of predicting web-surfing time periods of users. LM method has a better prediction accuracy when it is compared with the traditional Markov model and the Most-value model. There are two main findings here: the first one is that there is obvious periodicity in surfing time periods of 37.66% mobile phone users in experimental area by Fourier transformation and periodic tests, which could help us understand the surfing characteristics of users. Also, the second one is that the average accuracy of our proposed method is better than the Markov model and the Most-value model in 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes and 60 minutes intervals. LM method can perform an average accuracy of 79.75% in predicting web-surfing time on a scale of 60 minutes, better than the Markov model (74.64%) and the Most-value model (64.44%). Compared with the other two models, the accuracy distribution of the LM method is narrower, the peak value is higher, and the standard deviation is smaller, which means that the prediction accuracy of the LM method is more concentrated and stable, with good predictive performance.