For all of its empirical and theoretical rigor, science often begins with an intuition or inspiration. Long before a new idea becomes a paper that appears in an academic journal, for example, it begins as a hypothesis. These creative suppositions commence with “data” stored in a researcher’s mind, which she then “analyzes” through a purely psychological process of pattern recognition. The “aha” moments that may follow are not necessarily inspiration, but rather the output of the researcher’s brain-driven data analysis. In other words, scientific contributions often derive from a researcher’s idiosyncratic and very human thought process.
Given the importance of scientific research and its many benefits, this raises a question: Is there a better way to generate hypotheses other than a reliance on personal analytical insight? This novel paper examines this question through the lens of machine-learning algorithms and the exploding availability of human behavior data. Second-by-second price and volume data in asset markets, high-frequency cellphone data on location and usage, CCTV camera and police “bodycam” footage, news stories, children’s books,1 the entire text of corporate filings, and so on, is now machine readable. What was once solely mental data in the service of hypothesis generation is increasingly becoming actual data.
The authors posit that these changes can fundamentally change how science gets done, which they demonstrate by developing a procedure that applies machine learning algorithms on rich data sets to generate novel hypotheses. Please see the working paper for more details, but broadly described, the authors extend the human process of generating data-driven correlations to supervised machine learning. Their new approach not only yields far more correlations than a human researcher, but it has the capacity to notice correlations that a human might never discern. This is especially true in high-dimensional data applications, potentially opening the door to research that would otherwise go unexplored.
To illustrate their procedure, which is applicable across disciplines, the authors study a high-stakes issue with profound implications: How pre-trial judges decide which defendants awaiting trial are incarcerated and which are sent free. These decisions are supposedly based on predictions of a defendant’s risk, but mounting evidence from recent research suggests that judges make these decisions imperfectly. In this case, when the authors build a deep learning model of the judge—one that predicts whether the judge will detain a given defendant—a single factor has large explanatory power: the defendant’s face.
The authors find the following:
It is important to note what this work does not reveal. The authors do not claim that a mugshot predicts a defendant’s behavior; rather their analysis reveals that mugshots predict a judge’s behavior: A defendant’s appearance correlates strongly with a judge’s decision to jail or not. And what are those appearance characteristics? There are two, and the authors label the first as “well-groomed” (e.g., tidy, clean, groomed, vs. unkempt, disheveled, sloppy look), with the second as “heavy-faced” (e.g., wide facial shape, puffier face, rounder face, heavier). Importantly, these features are not just predictive of what the algorithm sees, but also of what judges actually do:
Again, please see the working paper for the authors’ careful consideration of the many factors involved in their analysis, including their discussion of the ways that psychological and economic research over the past century has informed our understanding of people’s reaction to faces, including on such factors as race. The point here is that the authors’ algorithm seems to have found something new, beyond what scientists have previously hypothesized, and beyond human capability.
Bottom line: Hypothesis generation matters, and developing new ideas need not remain an idiosyncratic or nebulous process. The authors’ framework reveals that combining supervised machine learning with rich human behavior data can open the scientific world to new and fruitful lines of research. The authors’ example of judicial decision-making, along with other applications that they describe, are just the beginning. Future research will help move hypothesis generation from a pre-scientific to a scientific activity.
1 See “What We Teach About Race and Gender: Representation in Images and Text of Children’s Books,” by Anjali Adukia, et al., for a BFI Research Brief and links to the paper, an interactive tool, and video presentation.