Such title is the very topic that Ms. Vo Thi Hong Phuong – Data Scientist, Head of Data Analysis at FPT Telecom, had shared at Solution Fourm – FPT Software Ho Chi Minh. The following article will provide a practical insight into AI applications and Machine Learning in solving this very relevant problem for businesses.
FPT Telecom provides various services, major ones being Internet and FPT Television. At the moment, leaving ratio of FPT Telecom averages around 1%. Is such a low rate worthy of using data analysis?
According to Ms. Phuong, leaving customers is a focus of the management of FPT Telecom due to certain reasons. One, keeping customers’ loyalty would increase the brand reliability, and thus revenue as profits only generate after a period of 12-18 months. Furthermore, investment cost for new customers are 5-10 times higher compares to that of old ones. Even new sales made would be influenced by leaving customers, and thus FPT Telecom always try to keep every one of their users, focusing on those with higher chances of leaving.
The problem is just the question of “Which customers may leave?”, without any metrics or suggestions. The data analysis team has to go through various steps, including: define problem, gather data, filter and standardize data, train model, test, complete, and deploy to the system.
“Define problem is of extreme importance”, Ms Phuong stressed. This lays the foundation to the entire later process. Then, you need to focus on data gathering and extraction (of branches’ statistics, sales policy metrics, customer information, behavior of service users, satisfaction…)
Most effort is spent in the step of preprocessing data, as in, making sure that data is arranged in the best and most logical way. The team’s motto is “To believe nobody”, seeing that only one day of not saving data or one error while inputting data may lead to grave inaccuracy. Even the impossible is possible, and therefore, data filtering and standardization is one of the team’s most important target through the last 3-4 years.
Transforming data is no less time-consuming. From the data gathered, the team have find meaningful data for analysis and prediction to select bright features for the models to learn. The team therefore needs to come up with various elements and simulations to separate the two classes of customers (those potentially leaving and those not), like frequency of contacting technical department, contract elements, age, infrastructure, end date of contracts… Then, the team shall supervise and analysis the real logics for the model to learn of the leaving points, before moving on to testing.
“Subjectivity is not always right” and “needing to consider as many aspects as possible” are adviced give after much experience. Running an algorithm may take a lot of time and cost, and therefore we need to make careful calculations to minimize the number of trials and save time for building models. As such, even the same problem but in different circumstances would require the team to redo their research on businesses, policies, data allocations… for the best result.
After building, testing, and running the model, the data analysis team at FPT Telecom runs into quite many challenges in real deployment. In particular, how to persuade businesses to believe in the system, how to convince them to drop traditional methods to adopt new ones. According to Ms Phuong, she often receives inquiries of “How do we know these customers would leave?”, “What are the signals and conditions?”, “I can’t quite believe”… when trying to introduce the model.
As making people understand and read about another’s model is quite the challenging task, Ms Phuong has to use the most convincing example: actual efficiency over long-term. At the moment, accuracy of prediction of leaving customers for FPT Internet is at 43.1% out of 215,760 customers, while with FPT Television, the rate is 32.5% out of 64,758 customers.
In summary, Ms Phuong said that real data is extremely complex and noisy, as the more people there are, the more noise the data gets. Along with it is the increasing demand for data analysis, and the difficulty, patience, and time involved in solving real life problems.
Ms Phuong has given a full disclosure on her team’s journey from building models to system deployment, from solutions to challenges along the way. She has also shared information on data analysis techniques, data gathering process, exploration and modelling of big data to find information, as well as the hidden value in data relevant to operations of telecommunication businesses.
Below is a livestream video of the event: