AI meets Behavioral Science

Our research at Yobi focuses on the exciting intersection of AI and behavioral science. In many ways, this is a new frontier for AI: while previous research has led to excellent models for images and text, we still don’t have great models for people. We are working to change that, combining Yobi’s rich data streams with innovative machine learning.

Predicting Human Behavior

Part of the success of machine learning for modeling images and text comes down to having found models that capture just the right kind of structure to predict the color of a pixel or which word will come next in a document. At Yobi, we are exploring different kinds of machine learning models for understanding human behavior. We can also think about this as a kind of prediction problem: based on their past behavior, what kinds of products might a person be interested in? By making better models, we can help companies find the right customers and help shoppers find their next favorite product.

Preserving Consumer Privacy via Better Modeling

Building better models is also a key part of our approach to consumer privacy. Yobi focuses on data that consumers have opted to share. Making this choice to focus on ethically-sourced datasets reduces the amount of data available for our models. Consequently, we need to make our models smarter to intelligently extrapolate from these smaller datasets. By using powerful machine learning methods, we are able to create excellent predictive models from limited data.

A “Foundation Model” for Human Behavior

The current trend in machine learning is towards building large models that can serve as a foundation for developing more specialized models. For example, a large language model such as GPT can be used as a source of embeddings for words and sentences, which can then be used to build smaller models for specific tasks. To build a sentiment classifier, predicting whether a sentence expresses a positive or negative emotion, an engineer can get the embeddings for those sentences from GPT and then train a machine learning model to classify the resulting embeddings as positive or negative. This new machine learning model benefits from the rich representation of sentences produced by GPT, which is based on training on a massive amount of text.

In the same way, Yobi has created a “foundation model” for human behavior. Our embeddings capture the aspects of human behavior that are key to making future predictions, while concealing the details of the actions taken by any individual consumer. By building machine learning models based on these embeddings, our clients get the benefit of our huge behavioral dataset for solving their own modeling problems. Even better, they get the benefit of our expertise in machine learning and behavioral science, as we have reduced these data down to a single vector of numbers for each consumer that have been pre-selected to be useful in predicting purchases and other actions.

‍