Detecting disability and ensuring fairness in automated scoring of video interviews

A method comprising:

• identifying, by a processing device using digital interview data stored in a computer storage, a subset of a plurality of interviewees that have a disability;
• labeling, by the processing device, a first group of the plurality of interviewees as disabled and a second group of the plurality of interviewees as not disabled with reference to the disability;
• identifying, by the processing device, features from the digital interview data for the first group that correlate with the disability;
• employing, by the processing device, a convolutional neural network (CNN) to find patterns in a time series audio/video (A/V) file within the digital interview data that are unique to the features that correlate with the disability;
• calculating, by the processing device using the patterns, a set of convolutional feature-pattern recognition filters to apply to the digital interview data;
• performing, by the processing device for a given A/V sample of the digital interview data to be modified, target activation calculations using the set of convolutional feature-pattern recognition filters, to generate a digital fingerprint;
• mapping, by the processing device, the digital fingerprint of the features onto a dataset of an interviewee belonging to the second group of the plurality of interviewees, to generate a mapped dataset;
• quantifying, by the processing device via application of a predictive performance model to the mapped dataset, effects of the digital fingerprint on a job performance score for the interviewee;
• determining, by the processing device, that a difference in the job performance score, due to the plurality of the features being mapped onto the dataset, results in an adverse impact;
• and
• updating, by the processing device, the predictive performance model by removing at least some of the plurality of the features that correlate with the disability, to reduce the adverse impact in application of the predictive performance model to review of the plurality of interviewees.

A non-transitory computer readable storage medium including instructions that, when executed by a processing device of a computing system, cause the computing system to perform operations comprising:

• identifying, using digital interview data for a plurality of interviewees captured during a plurality of interviews, a subset of the plurality of interviewees that have a disability;
• labeling a first group of the plurality of interviewees as disabled and a second group of the plurality of interviewees as not disabled with reference to the disability;
• identifying features from the digital interview data for the first group that correlate with the disability;
• modifying a dataset of an interviewee of the second group using the features identified within the digital interview data as correlative with the disability, to generate a modified dataset for the interviewee;
• applying a performance prediction model to the dataset to generate an original score;
• applying the performance prediction model to the modified dataset to generate an altered score in view of the modifying;
• determining, from a difference between the original score and the altered score, that the features that correlate with the disability impart an adverse impact comprising a bias against the interviewee being hired;
• identifying a subset of the features that most strongly contribute to the bias in contributing the most to the difference being greater than a threshold difference value;
• and
• iteratively removing, from the performance prediction model, individual features of the subset of the features until the difference falls below the threshold difference value.