Why AI Treats People as Clones
Two men, Roger and Stephen, could easily be grouped together in a marketing analysis. Both:
- were British,
- attended the University of Oxford for their undergraduate studies,
- pursued studies in the sciences
- achieved fame before the age of 30, and
- died in 2018.
Yet, while there are similarities between the two, their accomplishments and their lives were quite different. Roger Bannister (1929-2018) was a middle-distance runner who became the first person to run a mile in under four minutes. Stephen Hawking (1942-2018) was a renowned theoretical physicist and cosmologist known for his groundbreaking work on black holes, the development of the theory of Hawking radiation, and his bestselling book, “A Brief History of Time.”
We’re all individuals, yet while the marketing industry seeks hyper-individualization, all too often our AI systems are oversimplistically designed to group people together with thousands, if not millions, of other people. And while that approach may have been best practice twenty years ago, the world has changed. People demand to be treated as individuals, and they have learned to automatically discard the poorly targeted spam that fills their inboxes. Worse still, research shows that we don’t trust algorithms when we believe we are not being treated as individuals.
Sometimes, the most valuable signal you can feature engineer is to quantify how different a person’s behavior is from their apparent peers. This can lead to inspiration for new products for customers with unmet demands, and avoid damaging your organization’s brand from poorly executed digital campaigns.
To feature engineer a similarity signal, here are a few tips:
- Use a metric that standardizes for the number of events e.g. use the cosine similarity of product purchases
- Choose a time window long enough to capture the full breadth of customer behavior, including seasonal variations
- Metrics that compare across multiple events are more robust, so choose a tool that can natively handle one-to-many table relationships across a database
Here at FeatureByte, we’ve built an open-source feature engineering library that makes it easy to create similarity features. Click here to access our open-source SDK, with worked examples in Python.