| Customer churn has become a common problem faced by businesses due to factors such as shortened product lifecycles and intense market competition.Customer attrition not only results in a decline in sales revenue but also impacts the reputation and brand image of the company.Severe customer churn poses a threat to the survival and growth of businesses;therefore,it is crucial for enterprises to predict and manage customer attrition.With the rapid development of computer hardware and the continuous growth of data volume,deep learning has achieved significant success in various domains,such as computer vision,speech recognition,and natural language processing.However,for the specific field of customer churn prediction,building churn prediction models using deep learning faces numerous challenges.This is mainly because customer data is heterogeneous tabular data with complex and irregular dependencies among its features,and it often exhibits high dimensionality and class imbalance issues.Thus,constructing high-performance deep neural network models for customer churn prediction has emerged as a critical research topic in the field of customer churn prediction.In this thesis,three effective customer churn prediction algorithms based on the data resampling and deep learning are proposed.The main research contents are as follows:(1)A customer churn prediction ensemble algorithm based on one-dimensional convolutional neural networks(1D-CNN)and gradient boosting decision trees(GBDT)is proposed.Given that the cost of retaining an existing customer is often several times that of attracting a new one,to enhance the detection rate of churning customers,this algorithm first employs Tomek Links undersampling to address class imbalance issues.Next,the processed customer data is fed into a 1D-CNN model,which automatically extracts customer features through one-dimensional convolutional layers and employs fully connected layers to predict customer churn.If the 1D-CNN model predicts churn,the result is directly output;however,if the prediction is non-churn,the customer data is re-entered into the GBDT algorithm for a second prediction,and the new prediction result is output.Experiments on publicly available customer datasets demonstrate that the proposed ensemble algorithm improves the recall rate and F1 score of the customer churn prediction algorithm.(2)A customer churn prediction algorithm based on resampling and interpretable deep neural networks is proposed.Existing commonly used customer churn prediction algorithms generally only provide global feature importance,while a deeper investigation into the differential impact of various factors on customer churn is still needed.This algorithm first applies SMOTE oversampling,SMOTETomek mixed sampling,and Tomek Links undersampling to the dataset,and then divides it into three stages for training in the Tab Net model.Tab Net model performs multiple decision steps.In each decision step,an attention transformation module is employed for feature selection,and a feature transformation module is utilized to learn high-level representations of customer attributes,which are then used as input for the attention transformation module in the subsequent decision step.The outputs from all decision steps are aggregated and used for the final prediction and interpretation.Experiments conducted on publicly available datasets from various industries demonstrate that the prediction performance and interpretability of this algorithm are better than other comparative algorithms.(3)A customer churn prediction algorithm based on ensemble learning and interpretable deep neural networks is proposed.Ensemble learning integrates the predictions of multiple base learners,enhancing the overall accuracy and stability of the prediction.The algorithm initially preprocesses the raw customer dataset,followed by the application of Tomek Links under-sampling and Synthetic Minority Over-sampling Technique(SMOTE)over-sampling to generate two new datasets.Subsequently,these two datasets,along with the unsampled dataset,are employed to train three distinct Tab Net models.The outputs of these three Tab Net models are stacked to form a new training set,which is then input into a logistic regression model for further training,resulting in a comprehensive ensemble model.Experiments conducted on publicly available datasets from various industries demonstrate that the prediction performance of this algorithm is better than other comparative algorithms. |
