Interpretable Matching

A matching model predicts if a pair of records describes the same entity. Model interpretability is about the WHYs. Why did the model predict a specific pair as a match and another as a no-match? The higher the interpretability of a model, the better we can answer these questions.

Why do we even care about interpretability? Why not just focus on building a top-performing prediction model? Here are three very different reasons:

• Interpretability is a legal requirement in some use cases. Think of a model predicting payment fraud for an e-commerce shop or loan default. Modern systems rely on entity resolution to incorporate multiple loosely connected information sources, like databases of confirmed identity thefts. In many markets, consumers or applicants have the right to know why their transaction has been rejected.

• A credit card company has much experience detecting fraud. Practitioners are less likely to deploy a new model they don’t trust, even if it outperforms the status quo on test sets. Model explanations increase trust by confirming or updating the current beliefs with new evidence.

• An excellent test performance does not always translate to great inference on examples outside the labeled data. The test set might not represent specific scenarios well, or there might be a temporal drift in the data, for example, a worldwide pandemic, which changes the game’s rules. Model explanations should be part of the engineer’s debugging toolbox.

The second and third points are closely related. The engineer’s beliefs are part of the debugging process. Let’s explore how interpretability helps identify issues in a black-box model and the training data.

Explaining predictions with SHAP

In this lesson, we interpret a CatBoostClassifier algorithm fitted to the df labeled dataset of similarity features below: