Let’s get started.
Tell us what you’re working on, we’ll answer right away.
Churn of business customers
| Industry | Category |
|---|---|
| Telecom | User projects |
Client
Telecommunications company with a significant number of business users and group accounts.
88%
clients have issues that can be resolved using just chat
6.000+
hours per year is spent solving issues that can be automited
Challenge
A telco provider approached SmartCat to improve the existing churn model that the telco internal team had been developed. The problem refers to detecting companies (group contracts) that are likely to stop using provider services. The general monthly churn rate is very low (less than 2%) with no obvious or easy-to-detect pattern. Because of this, the client’s internal model had modest results and our goal was to increase the accuracy by 5-10% (this is something the client believed was achievable).
What we did
Our approach to this project included multiple stages, as follows:
- Phase 1: Data cleaning and validation. Exploratory data analysis.
- Phase 2: Feature extraction.
- Phase 3: Implementation and evaluation of predictive models for churn one and two months in advance.
During the first phase, we used historical data to analyze typical patterns, trends, and potential seasonality. Different statistics and visualizations were implemented in R. Also, we validated and cleaned data, since we saw that some related columns had inconsistent values in some cases. These steps were done in permanent communication with a dedicated person on client side. Before the start of the modeling phase, we extracted many features that were used as input to train machine learning models. The accuracy of models was measured using precision and recall for churners (because of highly imbalanced labels in the dataset) and compared with the client’s model. Also, we performed the analysis of seasonalities and anomalies.
Results
A predictive algorithm was being trained with historical data and optimized as we strived for our defined goal of prediction accuracy. Many features that we designed using provided data significantly improved the final accuracy. Comparing to the client’s baseline model, for the same recall values, our final model had a 5-10% higher precision, which satisfies the customer benchmark.
“Very smart people, great company. With detailed preparation and data sharing principles in mind (GDPR and security) they helped us develop algorithms to get the probability that customers will churn.”
Dr. Saša Radovanović
Data Science and CLM Team Leader
← Previous
Email sentiment analysis
Sentiment analyzer is one of the main tools to measure how people participate in products, services, and brand in general. It is great to have huge volumes of input text which needs to be classified (email campaigns, satisfaction surveys, review comments).
Next up →
Search ranking for a large-scale job platform
