Research On The Method Of Member Loss Prediction And Content Recommendation Of Music Website Based On RBM

As people's demand for audio-visual entertainment increases,the music industry is booming,and the number of various music platforms has surged.However,the current forms and content of various music products are homogenized seriously,and fierce competition has led to an increase in the loss of members,and the profit of the music industry mainly depends on member fees.Even a small amount of loss of members will cause huge economic losses.In summary,it is necessary to predict the loss of music members and take pre-retention measures in advance.This research measure will help music platforms better maintain the platform user ecology and bring more profits.This article mainly cuts into the field of music industry,selects the member behavior data of music streaming media websites as the data set to predict the loss of members and conducts the research of user recall strategy through personalized recommendation after the loss.Two user models are mainly built,the first is the prediction model of membership churn,and the second is the recommendation model of music content.The two models complement each other to form a closed loop in the business.Taking retention measures in a timely manner after receiving a churn warning has greatly prevented the churn of core users,and at the same time increased user stickiness through content recommendations.The main research of this paper is as follows:(1)Aiming at the problem that there is a high degree of behavioral similarity between lost members and non-churned members,and traditional classification models are difficult to effectively distinguish,a Restricted Boltzmann Machine model(RBM)is designed to extract features from member users,Build an Extreme Gradient Boosting(XGBoost)model to learn member behavior characteristics,and combine the two models to propose an RBM-XGBoost combined prediction model to predict member loss.Experiments have proved that the RBM-XGBoost combined prediction model is superior to the traditional classification model in terms of prediction accuracy and generalization ability.(2)Sparrow Search Algorithm(SSA),as an emerging group intelligent optimization algorithm,has outstanding optimization capabilities and is suitable for hyperparameter optimization problems,but it still has limitations.Aiming at the problem that SSA is easy to fall into the local optimal value and premature in the later stage of the iteration,an Improved Sparrow Search Algorithm(ISSA)is proposed.The algorithm initializes the population through PWLCM mapping to make the initial population distribution more uniform;when the iterative process falls into the local optimal value,it starts the mixed mutation strategy of t distribution mutation and Gaussian mutation to increase the diversity of the population and jump out of the local optimal value.Experimental verification shows that the improved sparrow search algorithm ISSA is more prominent in accuracy and stability.(3)Aiming at the problem that the XGBoost model has many hyperparameters,it is difficult to select by manual experience and the traditional parameter search method is time-consuming and prone to combinatorial explosion,the RBM-ISSA-XGBoost model is proposed,and the improved sparrow search algorithm ISSA is used to perform the XGBoost model.Hyperparameter optimization.Experiments prove that the prediction accuracy and convergence ability of the prediction model after parameter optimization have been effectively improved.(4)After predicting the members with a tendency to churn,divide the members into groups with different retention difficulties according to their churn probability,and make targeted retention recommendations.For member groups that are easier to retain,music recommendation is used as a retention method for churn recall,a hybrid recommendation model is designed,and the prediction results of RBM and user-based collaborative filtering models are combined through linear weighting,which can achieve more accuracy compared with a single recommendation model recommend.