GSPANN’s advanced analytics team created the model to predict customer churn utilizing its in-house ML solution called ‘Zeolite’ that is based on Keras (Python deep learning library). We leveraged data residing in the existing SQL servers and combined it with Azure cloud’s processing and deployment capabilities to provide a feasible solution.
After running successful POCs on various deployment methods, the model was finally deployed on Azure ML compute nodes and an automated batch-mode scoring pipeline was built. Further tuning and re-training was done in the second version of the model with additional features to achieve more accurate results.
Various data types with possible relations with the churning PCs were collected. These include net sales, order counts, tenure, email-click stream data, sponsor attributes, last purchase details, etc. The prediction made through the Deep Learning Keras Sequential Model showed around 73% of training and test accuracy, and 76% of Recall (the ability to find all the positive samples).
The model was deployed on Azure ML platform and a scoring pipeline was created wherein new data files can be scored in bulk with the help of an API POST call. The model scores the probability of each churning PC on the scale of 0 – 1. Based on the probability, the client segments the PCs into 10 segments, picks 3-5 segments with the highest probability, and targets them with personalized offers. The model is re-trained every 2-3 months. The solution also analyzes the results of the discount campaign that is run on the target customers.
