A Large-Scale Image-Based Appearance Analysis on Career Transitions

Discrimination against individuals based on beauty has generated immense interest both among researchers and practitioners. Any discrimination in the labor market based on attractiveness would make the hiring and promotion processes inefficient. Existing literature has often studied beauty bias in an experimental setup. Individuals are asked to play hypothetical roles such as jury member or manager, choosing between likely convicts or able workers respectively. These studies highlight bias when an evaluator has limited information to differentiate candidates on true quality. Other studies that look at archival data show a beauty premium over a significant career length without revealing whether bias consistently affects an individual over his or her entire career or not. Determining at what point beauty bias plays a role is extremely important to understand the mechanism at play. These long-term studies typically find positive premium for both men and women overall. Once again ignoring the time dynamics likely mask different mechanisms at play for the two genders.

In this project, we investigate the dynamic effects of beauty bias in the labor market to identify different factors that result in beauty premium for men and women. Using detailed panel data on career milestones and education background of 7,436 individuals selected from a professional social network, we show that the beauty bias assists attractive men (2 s.d. above mean) in moving up 22.6% faster early in the career when employers have little information about the true quality of the candidates. However, early in their careers, attractive women may not be able to extract the same advantage; they may be perceived to not be serious participants in the job market or face jealous reactions from their competitors. But, if analyzed over mid to late career phase, they would appear to outperform their plain-looking counterparts. We find attractive women (2 s.d. above mean) are 33.6% more likely to move up in any given period in this phase. Studies have shown mixed results on whether beauty is penalizing or rewarding for women. We assert that the confusion likely arises from focus on different career phases in these studies.

The size of our study (7,436 primary and 92,540 auxiliary profiles) provides a unique external validity on impact of looks on professional careers. However, the breadth of our study presents two key challenges: we need to rate the attractiveness of a large number of subjects and establish a measure of success that is comparable across industries and career stages. Using computer vision techniques, we build a supervised machine learning model that is trained to mimic attractiveness scoring done by human raters. This model is invariant to temporary characteristics, such as clothing, hair style and expressions. Secondly, we extend upon traditional labor economics methods that establish preference order between jobs utilizing observed pairwise job switches. This approach works well when the hierarchy is well-defined, allowing quantification of success. Our unstructured data is highly sparse, including over 100,000 unique employers and 100,000 unique titles. Therefore, we improve upon the basic idea by building a dense representation of jobs and running a variant of Page Rank algorithm to calculate a measure of job’s rank. The resulting job ranks are highly intuitive and can be used by other studies that investigate career transitions from large-scale unstructured data.