Over the past decade, focus on data science and machine learning (ML) has been radically shifted from theory to real-world applications as more powerful machines, learning algorithms and the vast amount of data is available. The improved performance of ML algorithms is the direct result of increased complexity. A prime example is the deep learning paradigm, where complex hierarchical models are built by iteratively adding activation layers. This clearly states a trade-off between the performance of a machine learning model and its ability to produce explainable and interpretable predictions. Therefore, on the one-hand, we have the so called black-box models where the complex structure hinders model interpretability, such as deep learning and ensembles. On the other hand, there is the class of so called white-box or glass-box models, which easily produce explainable results, such as linear and decision tree based models.

We have employed several methods when it comes to the emerging landscape of interpretability techniques. Model interpretability can be about generating interpretations for single prediction scores produced by a classifier, i.e., local interpretation or by analyzing the accumulated impact of individual features on the entire customer base, i.e., global interpretability.

A especially important separation of interpretability happened based on the type of algorithms that could be applied. If the application is only restricted to a specific family of algorithms such as deep neural networks, then these methods employed are called model-specific . In contrast, the methods that could be applied in every possible algorithm such as ensemble models, are called model agnostic.

The interpretation methods that do not try to create interpretable models, but, instead, try to interpret already trained, often complex models are known as post-hoc interpretability method. It encompasses methods that are concerned with black-box pretrained machine learning models such as deep neural networks. In contrast to post-hoc approach, there exists interpretation methods for models often called intrinsic, transparent, or white-box models. Such models include the linear, decision tree, and rule-based models and some other more complex and sophisticated models that are equally transparent and, therefore, promising for the interpretability field.

The above figure shows how explainable AI is used in practice to gain actionable insights. We have fit one of the explainable AI methods on a churn model implemented on moke-up data. This is the post-hoc interpretability method in which instead of creating interpretable models, the already trained model is interpreted. The figure gives a very clear and easy to understand interpretation of the results of the churn model, i.e., the major indicator of customer churn is the sudden reduction in data usage. This is then followed by reduction in voice minutes and change in handset. The above three indicators gives a collective understanding that reduction in service usage along with change in handset is the biggest trigger of churn.

Unlike age or gender, the above indicators are actionable, i.e., alarms can be raised following simple decision rules to help approach the customers proactively before any churn decisions are made. Likewise, there can be customized retention campaigns triggered automatically few months before the contract binding period expires.

The above example only summarizes isolated triggers whereas there can be cases in which a combination of certain triggers leads to major churn events. The example signifies how model interpretability leads to actionable insights and can create a considerable impact on EBITDA.