Customer Churn Prediction
DiveDeepAI developed a platform that helps the company to identify the customers which are more likely to quit the service in the future.
A water supplying company operating in Canada and USA with turnover more than 2 billion dollars. Company is providing services to both commercial and residential customers.
The initial data was huge, the database consisted of approximately 350 tables, and it was very difficult to merge data into a single dataset. Moreover, few tables had more than 550 million records and our development environment (jupyter notebook and google colab) crashed frequently. Selecting a model for training was a difficult phase. We tested several models to find the best one. Hyper parameters fine tuning took much of the development phase.
Project was completed in several phases. Commercial customers were added to the analysis and churn rate was calculated for different categories of customers. Different parameters such as price change, age group, gender, length of residence, missed delivery rate, customer service were used to calculate the affecting customer churn rate. Moreover, a correlation analysis was conducted. A machine learning pipeline was created and trained several machine learning models but Lightgbm outperformed the other algorithms. In the end, threshold was set for several features and highlighted which customers are more likely to quit.