Step 2: More predictive models
The next step was to introduce another model and to merge it with the one predicting churn. The new model was meant to assess customers’ buying preferences. Once a customer was identified as one who was likely to leave, the model would suggest the consultant what products they should present to increase the chance of retention. The idea was to offer customers only the products that would fit their needs.
Once the two models were ready and successfully merged, it was the time to use some real data collected by the company. We initially worked with the Customer Retention Department, planning to start the pilot after 9 months from the beginning of building predictive models. Working Agile, we quickly discovered what features are really needed to start testing the models in action. After reviewing the initial plan and verifying some of the basic assumptions, we were ready to launch the pilot after less than 3 months!
That was the time when we were able to expand our activity and work with different departments. After achieving the set goals in the most demanding area of the market, we expanded our predictive models to work on the whole database and added some tools to support the Sales and Marketing Departments.
One of the most important challenges of this stage was to make sure that data records of millions of the Telecom clients are secure and no unauthorized person has access to this data – and we achieved both goals by applying proper data anonymization procedures.
As we dealt with sensitive personal data, it was very important to secure all the records from the very beginning. For us, as a third-party contractor, every end customer had to be anonymous. To ensure that, we implemented a solution that generated one-sided hashes based on customers’ data. From our perspective customer was a long identifier like “XXXX-YYYY-ZZZZ-1234” not Mr Jan Kowalski living in Warsaw, without possibility of reversing such hashing to get to know the data of the actual person. When someone from the Client’s team needed to fetch that person scores, he was able to send the request that was generating hash on the fly to get to look up a profile of desired customer – and to receive the right record back. To prevent potential data leaks, we made sure that each time someone requests the data about the specific user, his or her name was stored in the access history. If any data was exposed, it would be easy to track who was responsible for the leak.
Data anonymization was also crucial to comply with the GDPR policy. As every piece of available data was anonymous, we were able to use it to train the models. The outcome, however, was applied only to those customers, who agreed for the profiling – only those ones would receive personalized offers based on their behavior and the predictions about their future choices.
At this stage, we delivered a trained predictive model predicting the risk of churn and assessing clients’ buying preferences as well as all the mechanism that kept churn scores updated. As the model was trained on real, though anonymized, data provided by the Client, it was ready to test.