Model interpretability in Machine Learning: a talk by

Patrick Hall is senior Data Scientist and Product Engineer at This is a company that provides Artificial Intelligence solutions for businesses. As such communication of complex technology is a trademark for This blog has posted about in its brief existence. Then it was about some AI course suggestions for an enterprise audience.

Today I am posting again from, but this time it is about a more technical subject. Interpretation of models is often an important issue with Machine Learning, especially so in business implementations. Within regulated industries like insurance, banking or retail the need for a proper, simple interpretation of extremely complex models becomes a priority. This is more so if there is any advantage in using Machine Learning frameworks within business.

The talk I share deals precisely with Machine Learning Model Interpretability for a business setting. As a case in point these models often rely on simulation of scenarios in order to perform predictive analytics. In this it common to employ what is called a Surrogate Model. These are approximation models of reality, but are often helpful in helping with the sometimes thorny model interpretation task.

Interpretable machine learning  


The video is full of nice slides about how communicates the best Machine Learning models for a variety of business applications. For instance Patrick informs us that Generalized Additive Models have gained some business traction. Also important for business is separable interpretable linear models – a space of breakthrough in recent decades -, in order to accurately perform prediction of variables of interest.

The suggestion of combining models for better accuracy is particularly interesting. The slide that presents small, interpretable ensembles is the part of the talk about this (check the featured image). Here  the main trick revolves around using less complex models and not a big number of models. The point about European regulations against black box decisions was a bit disappointing. Would this work in favor of improved research on this field? We hope so.

From the YouTube description of the video:

While understanding and trusting models and their results is a hallmark of good (data) science, model interpretability is a serious legal mandate in the regulated verticals of banking, insurance, and other industries. Moreover, scientists, physicians, researchers, analysts, and humans in general have the right to understand and trust models and modeling results that affect their work and their lives. Today many organizations and individuals are embracing deep learning and machine learning algorithms but what happens when people want to explain these impactful, complex technologies to one-another or when these technologies inevitably make mistakes?

This talk presents several approaches beyond the error measures and assessment plots typically used to interpret deep learning and machine learning models and results. The talk will include:

– Data visualization techniques for representing high-degree interactions and nuanced data structures.

– Contemporary linear model variants that incorporate machine learning and are appropriate for use in regulated industry.

– Cutting edge approaches for explaining extremely complex deep learning and machine learning models.


Wherever possible, interpretability approaches are deconstructed into more basic components suitable for human story telling: complexity, scope, understanding and trust.


Finally a word of mention to the part of the talk about Sensitivity Analysis. Patrick Hall here cautions everyone about the perils of model validation when many data scientists transition form a linear model to a non-linear model in the quest to boost accuracy. This can be a waste of time and ending up introducing unnecessary risks. Of notice is what Patrick says about Machine Learning models being highly regularized and numerically stable ones but their predictions turning out to be unstable, and how this can have bad consequences. Then the part about trade-off between global variable models and local variable models ends the talk.

featured image: Interpretable machine learning


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s