Machine Learning and Human Interpretability: The shadow price of Complexity

Machine learning can be thought of as sitting at the intersection of computer science and statistics. ‘Computer Science has focused primarily on how to manually program computers, Machine Learning focuses on the question of how to get computers to program themselves …’ – Tom Mitchel (see: ).

For many tasks, it makes relatively little difference if these programs are opaque to human introspection. We are almost exclusively concerned with performance for some task (‘is this a cat?’; Donkey Kong, etc.). Here, high capacity models, like deep learning, suffer little penalty for marginal increases in representational complexity.

However, for several reasons, marketers tend to be wary about ceding control of their customers’ experiences to black box methods. Firstly, they need assurances that they can trust automation to make reasonable decisions over a large space of possible environments. Secondly, companies often have internal and legal regulatory requirements around the how, and which, data may be used for making marketing decisions.  With the EU’s GDPR coming into effect in May 2018, this will be even more of a concern.

While lacking the expressiveness, or capacity, of more complex representations, decision trees have many appealing properties for the marketing use cases:

1)  Human readable – For any given input, a human can easily determine the output.

2)  Auditable – for regulatory review and approval before use.

3)  Loggable – Because each decision node in each decision tree can be uniquely identified, it is possible to register, and retain a history of what policy, or rule, was used for every decision event. This is particularly useful for significant decisions as defined by Article 22 of the GDPR.

Conductrics makes marketing software that helps companies discover what experience to give each customer via trial an error learning. In industry speak, this is generally referred to as ‘Website Optimization’, or simply as ‘Optimization’, of which AB Testing is a subset. This class of problem can be framed as a contextual bandit or, more generally, as a reinforcement learning problem.

In addition to incorporating an interpretability constraint, we also consider more traditional issues surrounding optimization problems (exploration-exploitation trade-off, estimating real world efficacy, drift, etc.)


Matt is co-founder of Conductrics, a marketing optimization platform.  Matt has an MSc. in Artificial Intelligence from the University of Edinburgh, and an MS in Resource Economics from the University of Massachusetts at Amherst.

Thank you to Georgian Partners for the extra space!

Both audiences, those who are interested in machine intelligence / data science, and those who are practitioners of machine intelligence / data science are invited, and can both expect to learn something new.

Attendees can expect to learn what machine intelligence is, its applications, and what’s going on in Toronto’s data science community. Significant getting to know you time, and Q&A time is deliberately set aside.

Back to all