The AI Playbook

“Your mission, should you choose to accept it, is to forge a rare collaboration, enlisting business leaders to weigh in on the caveats and qualifications that determine the prediction goal in all its detailed glory.”

If you’re launching a new ML project, consider creating a binary model, or “binary classifier,” that makes predictions by answering yes/no questions. e.g., if you want to target customers more effectively with your advertising, your binary model could answer the question: “Will the customer buy if contacted?”.

“Headlines about machine learning promise godlike predictive power… It’s all a lie.”

Focus on metrics such as “lift,” which calculates the ratio by which your model performs better than making guesses without the model, and “cost,” which calculates the price of false negatives (FN) and false positives (FP). The “fatal flaw” of using accuracy models to evaluate your model is that it treats FNs and FPs as equally bad, when in reality, one may be costlier than the other.

It doesn’t matter how sophisticated your algorithms are if you don’t use the right data to fuel your predictions, so make sure your data spreadsheet is:

Long — You need a lengthy list of entries to ensure your data is representative. 

Wide — Note pertinent information about each entry in the row connected to the item.

Labeled — Humans often need to manually label data to help train ML software to detect negative and positive cases.

“Machine learning algorithms may be the fun, sexy part — everyone wants to crash that party — but improving the data is where you usually get the greatest payoff.”

ML algorithms “learn” from your data, deriving functional predictive models. Your model emerges from a “rule” or “pattern” that an algorithm derives from your data and uses to make predictions. While it can be tempting to greenlight a model when it aligns with your human intuition, understanding your model is not typically a straightforward process. 

“When a newborn model emerges, it absorbs all your attention. Like counting a baby’s fingers and toes, you examine it thoroughly, poking around to see how well it works and why — what makes it tick.”

Deploying your model means moving from the experimental phase to leveraging its predictive capabilities in the field to drive your operational decisions. Deployment will require full-stack cooperation and buy-in from team members at every level of your organization. Though your leaders may be the ones approving your model, it’s vital that you overcome any resistance and get buy-in from the staff tasked with deploying it in the field. People naturally fear change, so work to build trust in your model.

“ML delivers a rocket, but those in charge still must oversee its launch.”

“When you launch astronauts into space, you commit yourself to a new job: You’ve got to keep them alive. Likewise, once it’s in play, sustaining a model’s viability moving forward takes maintenance, monitoring, and vigilance.”