Page 62 - Flytxt

CSP level, which results in significant domain is applied to predict customer
overheads. Moreover, within the same churn in the banking domain.

CSP, subscribers’ behavior may change So the question to address here is can we
over time. For instance, subscribers’ formulate a generic framework, which
behavior in pre-WhatsApp and post- would be agnostic to the data set or

WhatsApp periods are significantly domain, automatically recommend an
different. This could potentially lead to optimal model with respect to a given
invalidating the old model and building task, without or minimal involvement

a new model from scratch to account of human Machine Learning experts,
for the new behavior. The new model thereby overcoming challenges

will in turn require manual monitoring mentioned above?’. In this context, we
of its performance on a day-to-day will discuss one of the most promising
basis, identification of abnormalities research areas namely, Automatic

in the model’s performance and Machine Learning (AutoML), which can
perform the steps from scratch. These bring in enormous capability to the data

Data preprocessing layer

One hot Missing value Outer
encoding imputa ons Normaliza on detec on

Feature preprocessing layer

Feature Feature extrac on Hybrid
selec on (e.g. PCA) technique

Model selec on layer

Algorithm Hyper parameter
selec on op miza on

Output layer

Visualiza on
dashboard

Fig - 3-4-1 Source: Flytxt

processes consume a lot of the precious science and Machine Learning arena in

human time. This becomes even more the years to come.
challenging if the same churn model, An AutoML framework seeks to

developed for telecommunication automate the process of designing

62 INSIGHTZ - VOLUME 03, 2018