Improved Bidirectional-Long Short-Term Memory for Customer Churn Prediction in the Telecom Industry

In competitive landscape of the telecom industry, predicting customer churn is a crucial challenge for service providers. This paper proposes an improved method for churn prediction namely, improved Bi-LSTM model. It starts with preprocessing input data via Tanh-normalization to enhance the data representation. Unlike traditional Bi-LSTM models, this proposed framework incorporates convolutional and max pooling layers for feature extraction with batch normalization to enhanced training efficiency. Moreover, it has strategic dropout layers to reduce overfitting risks. To address the class imbalance, the model employs a modified Focal loss function enhances strength and generalization capabilities. Experimental results shows that the improved Bi-LSTM model not only achieves higher prediction accuracy but also displays improved stability as well as efficiency in handling complex telecom datasets.