User Churn Prediction In Music Streaming Based On Time Series Data

With the rapid development of the mobile Internet,the music streaming market is expanding.In order to seize the market share,music streaming service platforms attract new users in various ways.At the same time,there is little cost for old users to switch to other platforms according to their own preferences.This means that old users are easy to churn from the music streaming service platforms.Slight variations in churn have a great impact on the profits of enterprises.It is important to accurately identify potential churn users and implement retention operations on these users.Currently,the industry primarily uses data mining and machine learning techniques to predict user churn.In view of the characteristics of churn data in music streaming and the problems existing in the current churn prediction methods,this paper improves the performance of churn prediction from the two aspects of models and features.The main contributions in this paper are listed as follows:(1)Since the churn dataset in music streaming field usually contains a large amount of time series data and the LSTM model is often used for sequence modeling,this paper proposed an ensemble model based on LSTM to improve the performance of churn prediction from the perspective of ensemble learning.On the one hand,the proposed model uses LSTM as the base learner;on the other hand,it improves the Snapshot ensemble method according to the characteristics of the real data.Specifically,we introduce a method to adjust sample weight for the Snapshot ensemble method,and use the learning method to combine the output of sub models.The experimental results show that PR-AUC of the proposed model is increased by 4.21%compared to the original LSTM model.(2)At present,the common models used in churn prediction is relatively simple and their representation ability is not strong.These models cannot make full use of the time series data.This paper proposed a churn prediction model based on LSTM and CNN,which aims to improve the performance of churn prediction from the perspective of model's structure.The proposed model enhances the ability of learning features by combining LSTM and CNN.In addition,it can learn long-term dependency features and local important features at the same time.The experimental results show that compared with the original LSTM model and CNN model,the proposed model increases PR-AUC by 5.05%and 6.08%respectively.(3)The construction of business features in current churn prediction tasks requires a lot of time and the common strategies of selecting training samples will lead to the problem of wasting historical data.To improve the performance of churn prediction from the perspective of features,this paper proposed two non-manual feature construction methods that are independent on business,and a model based on fused features.First,we use the existing model to transform the historical data into features,aiming to make full use of the historical data.Then,we use the tree model to perform high-order combination of existing features to construct tree model features.Finally,we fuse the newly constructed features and the original features.In addition,we improve the existing model's structure by using fused features.The experimental results show that the model that fuses all features increases PR-AUC by 2.71%compared to the model based on the original features.