These 3 AI Algorithms Provide True Value to Marketers

Gradient boosted machines (GBMs) are high-performing predictive models marketers can use to make behavioral predictions about customers to drive personalization. GBMs can answer questions like: Which prospects most resemble current, high-value customers? Which customers are likely to upgrade to a higher service level? Which customers are likely to leave the brand?

GBMs produce more accurate results than other predictive methods because they are often built as a sequence of decision trees. Misclassified items are used as the training data for subsequent rounds to fine-tune the model. This iterative approach makes GBMs more accurate than other algorithms, but it also means they can be made up of hundreds of simple models combined into a single highly accurate (though often highly complex and opaque) model.

Why did this algorithm come up with the answer it did? As AI goes mainstream, organizations are increasingly asked to explain the decisions their tools make. Recent concerns about bias in AI algorithms used to determine credit qualification or criminal sentencing have led to louder calls for strong AI oversight. Proactive companies now have SHAP algorithms to understand how black-box models like GBMs or neural networks produce their recommendations or predictions.

SHAP algorithms answer two important questions:

1. How did the model draw its conclusion for a particular record?

2. Which variables have the biggest impact on the model? 

To answer the question, SHAP often produces a plot that visualizes the various elements that influence a decision and their relative weight. SHAP works best for models that use structured datasets, but be warned: SHAP can tell you what the model is doing, but it can’t tell you whether the model is of high quality. SHAP can also require a significant amount of computing power to operate, making it expensive to run.

For the past decade, content teams have used A/B and multivariate testing to identify the best content to elicit a desired response from various audience segments. Contextual bandit algorithms are the next wave of improvement to content testing. They identify audience preferences more quickly and in a more automated fashion as compared to A/B or multivariate approaches, and thereby reduce the number of people who see low-performing content.

Contextual bandits are particularly useful on high-traffic channels (eBay, Amazon, Microsoft). The high volume of impressions possible in a short time on these sites produces meaningful results quickly, but the results can also apply in real time to enable a brand to stand out from competitors in noisy online environments.

Contextual bandits work by combining information about the customer with contextual data such as time of day or current weather and any available details about the current visit or user history. Using this data, contextual bandits serve up ad content to test and home in on the combination of words and visuals that best produces the desired response.