Research On The Prediction Of The Lifetime Value Of Mobile Game Users

With the rapid development of the mobile game industry,more and more game manufacturers purchase user traffic and manage users from social platforms based on user life time value indicators.User life time value refers to the sum of economic benefits that users bring to the company in the whole purchase cycle.Unlike traditional industries,in order to adapt to the market trend,mobile games need to quickly update the version to add functions,so they often only rely on prediction to obtain user life time value in advance.Due to the rapid changes in the market and products,the data often have strong timeliness,resulting in a certain degree of difference between the training data of the model and the data of the prediction,that is,the concept drift problem.The existence of concept drift has greatly reduced the prediction performance of the model.It is particularly important for game manufacturers to mitigate the impact of concept drift as a core indicator of game adjustment,but this problem has not been fully considered in previous studies.Based on the multi-task learning framework,this paper uses the hard sharing mechanism to split the 30-day user life time value prediction task into six 5-day prediction subtasks,and then uses the user churn tag to assist the main task user life time value prediction learning through the soft sharing mechanism,and innovatively proposes a multi-task learning prediction framework called VC-Pro MM.Compared with traditional prediction methods,VC-Pro MM can effectively alleviate the concept drift problem,significantly improve the prediction effect,and obtain more robust prediction effect,which can be widely migrated to other projects.This paper takes the user profile data of a mobile game project owned by T Company as an example.First,users are divided by registration month,and then multiple time interval data sets are manually constructed.The VC-Pro MM is compared with the prediction effect of traditional models on all data sets.The experimental results show that the data in the mobile game scene has obvious timeliness,the larger the training set capacity is not the better,and the longer the time span will reduce the model performance because of concept drift;The multi-task learning framework proposed by VC-Pro MM can deal with incomplete tag data information,so the prediction effect is stable and significantly improved compared with traditional prediction methods;Finally,the ablation experiment proved that the two sub-modules under the VC-Pro MM framework improved the model effect to a certain extent,and proved the stability of the model.This paper believes that in the field of mobile games,the VC-Pro MM framework can be used to predict the user life time value to help game manufacturers purchase user traffic and adjust games.