Research On The Construction Of Customer Behavior Prediction Frameworks Based On Ensemble Strategies

Customers are the foundation for the survival and development of enterprises,as well as the source of value enhancement and benefit creation.Customers refer to natural persons or organizations who exchange money or some valuable goods for property,services,products,or ideas.Customer behavior refers to the decision-making behavior of customers under the influence of various factors,such as whether to purchase enterprise services,cancel current service bookings,or stop subscribing to current services.Factors influencing customer behavior can be divided into two aspects:self-factors and external environmental factors.Self-factors can directly influence customer behavior,while external environmental factors need to influence customer behavior through external media.Customer behavior prediction refers to the adoptation of customer historical behavior and performance-related information to predict their future consumption decision-making behavior,in order to judge their possible decision-making behavior.Enterprises can provide personalized management strategies for customers by conducting customer behavior prediction,maintain sustainable and healthy development by enhancing insight into customer behavioral tendencies,reducing customer churn rate,increasing customer retention rate,and improving customer satisfaction.However,in the field of customer behavior prediction,researchers typically use accuracy metrics to evaluate the performance of customer behavior prediction,and select customer behavior prediction models based on accuracy metrics,ignoring the pursuit of enterprises for profits.Some scholars have studied customer behavior prediction from the profit-driven perspective,but have not considered the practical application of multiclassifier ensemble frameworks from this perspective.Therefore,developing more advanced multi-classifier ensemble frameworks for profit-driven customer behavior prediction has important theoretical value and practical significance.Predicting customer behavior in terms of purchasing,booking canceling,and churn can help businesses to achieve more accurate targeting of marketing objectives,increase the number of effective bookings,reduce cancellation rates,and prevent customer churn by following their needs and improving product information.This is of great significance for improving business efficiency and expanding operational scale.Considering that customer behavior is often influenced by multiple factors,this study takes profit-driven approachs to fully consider the characteristics of customer behavior and constructs fusion ensemble,linear stacking ensemble,and weighting ensemble frameworks to predict customer purchasing,canceling booking,and churn behaviors.Furthermore,Sharpley additive explanation values are used as a post-explanation method to interpret the prediction results of the three ensemble frameworks.The developed profit indicators in this study have expanded the modeling research on profit-driven modeling in some areas of customer behavior prediction,while the proposed ensemble prediction framework has overcome the shortcomings regarding low predictive ability of individual classifiers in previous research.The Sharpley additive explanation values also addresses the lack of interpretability in multi-classifier ensemble frameworks.This study enriches the theoretical and methodological research on customer behavior prediction from a profit-driven perspective and provides support for business decision-makers in developing accurate marketing strategies,booking management strategies,and customer retention policies.The fusion ensemble framework is constructed to predict customer purchase behavior.First refer to the profit evaluation indicator for predicting customer churn and design a profit evaluation indicator for customer purchase prediction as the profit-driven basis.Then,bagging ensemble classifiers and boosting ensemble classifiers are combined to construct a fusion ensemble framework in order to simultaneously reduce variance and bias.Grid search is employed to select hyperparameters for the proposed fusion ensemble framework to achieve profit-driven customer purchase prediction,and explanatory analysis is conducted by calculating the Sharpley additive explanation values based on the prediction results.Finally,after the verification from Bayesian perspective,conclude that customer purchase prediction performance is superior to benchmark classifiers.This study extends customer purchase prediction research from the profit-driven perspective,further enhances profitability with fusion ensemble framework,and clarifies the characteristics of customers with potential purchase behavior in the customer group based on Sharpley additive explanation values.The linear stacking ensemble framework is built to predict customer booking cancellation behavior.First step is to reference the profit evaluation indicator for customer churn prediction and transform it into applicable evaluation indicator for predicting hotel booking cancellations.Then,grid search is used to determine the hyperparameters of ensemble members,and the optimal prediction results of each member with the strongest profit-making ability are obtained.Next,grid search is used to select the hyperparameters of the linear regression methods,and the prediction probabilities of the ensemble members are integrated based on maximizing profits.Two prediction datasets are then selected to evaluate the customer hotel booking cancellation prediction performance of the linear stacking ensemble framework and baseline classifiers.Furthermore,by setting the hyperparameters of the profit evaluation indicator,the sensitivity of the linear stacking ensemble framework to hotel business scenarios is analyzed.Finally,based on the prediction results of the ensemble members,the Shaipley additive explanation values are calculated,and the Sharpley additive explanation values of the linear stacking ensemble framework are calculated using the weight coefficients of linear regression methods to analyze the main factors affecting customer booking cancellations.This study not only addresses the shortcomings of individual classifier prediction abilities in customer booking cancellation prediction but also optimizes the interpretability using post-explanation methods.The weighting ensemble framework is constructed for predicting customer churn.Firstly,grid search is adopted to determine the hyperparameters of ensemble members,ensuring that their profit capacity is maximized.Secondly,the artificial hummingbird optimization algorithm is introduced to optimize the weight coefficients corresponding to the prediction probabilities of ensemble members,and the weight coefficients with the strongest profit are obtained through a global optimization perspective.Then,eight public datasets from various industries are selected to evaluate the prediction performance of the weighting ensemble framework,and the Sharpley additional explanation value is calculated using the optimal weight coefficients to analyze the interpretability of the prediction results.Finally,the weight coefficients of the proposed weighting ensemble framework are analyzed by the evaluation indicator on prediction performance of ensemble members,verifying the conclusion that the artificial hummingbird optimization algorithm has adaptive tuning capabilities for optimizing the weight coefficients.Compared with existing research,the proposed weighting ensemble framework has higher prediction ability than the individual classifiers without sacrificing interpretability.Based on the existing public datasets,this study predicts customer behavior at three scenarios:customer purchases,customer booking cancellations,and customer churn,and draws the following main conclusions:(1)In customer purchase prediction,the proposed fusion ensemble framework integrates the advantages of bagging and boosting ensemble classifiers.Compared with the baseline classifiers,the fusion ensemble framework combining categorical gradient boosting and random forest achieves the highest predictive profitability.The Sharpley additional explanation value calculated based on the prediction results can provide support for decision-makers to formulate precise marketing strategies.The results of Bayesian A/B test indicate that the proposed fusion ensemble framework is more likely to become the optimal model for customer purchase prediction compared to baseline classifiers,and result in the lowest expected loss.(2)In customer booking cancellation prediction,the linear stacking ensemble framework outperforms the baseline classifiers,and the linear regression methods can effectively integrate the predicting probabilities of ensemble members.Sensitivity analysis of the hyperparameters in profit evaluation indicator reveals that the linear stacking ensemble framework surpasses baseline classifiers in different hotel scenarios.Analyzing the factors that affect customer booking cancellations based on the prediction results can help companies optimize booking management.(3)In customer churn prediction,the proposed weighting ensemble framework can bring higher profits to the customer retention of enterprises.Combined with the Sharpley additional explanation value,it can further improve explanatory power and provide more information for company decision-makers.The artificial hummingbird optimization algorithm can adaptively optimize the weight coefficients of predicted probabilities corresponding to ensemble members globally.The main innovations of this research are as follows.First,three prediction frameworks based on different ensemble strategies were constructed and applied to customer purchase,booking cancellation,and churn scenarios,which can provide enterprises with customer behavior prediction and optimize personalized content provided during customer relationship management,helping decision-makers to develop accurate customer service strategies.Second,from profit-driven perspective,the hyperparameters of customer behavior prediction models were selected through grid search,and the weight coefficients in the weighted ensemble framework were determined using artificial hummingbird optimization algorithm,which helps achieve profit-driven customer behavior modeling and optimize customer relationship management in enterprises.Third,the Sharpley additional explanation value was used for post-explanatory analysis of the three customer behavior prediction frameworks,enabling interpretable customer behavior prediction,and providing decision-makers with more information.Although this study increased the complexity of model structure,post-explanatory analysis was performed in different forms based on Sharpley additional explanation value,which can improve customer behavior prediction performance without compromising model interpretability.Due to limitations in the dataset,this study has the following shortcomings that require further research in future.First,selecting the hyperparameters in profitability evaluation metric based on previous research cannot fully measure the profitability of classifiers in customer behavior prediction,but only serve as a standard for comparing the profitability of various classifiers.In future studies,it is necessary to find datasets that contain more customer information to improve the scientificity of selecting the hyperparameters of profitability evaluation metrics and provide decision-makers with more valuable empirical results.Second,priority was given to numerical and categorical variables,while ignoring unstructured data such as texts and images.Future research should consider finding texts or images that can assist in predicting customer behavior,using deep learning models to extract useful information and improve predictive performance.