The Good, the Bad, and the Unhirable: Recommending Job Applicants in Online Labor Markets

Choosing job applicants to hire in online labor markets is hard. To identify the best applicant at hand, employers need to assess a heterogeneous population. Recommender systems can provide targeted job-applicant recommendations that help employers make better-informed and faster hiring choices. However, existing recommenders that rely on multiple user evaluations per recommended item (e.g., collaborative filtering) experience structural limitations in recommending job applicants: Because each job application receives only a single evaluation, these recommenders can only estimate noisy user-user and item-item similarities. On the other hand, existing recommenders that rely on classification techniques overcome this limitation. Yet, these systems ignore the hired worker’s performance—and, as a result, they uniformly reinforce prior observed behavior that includes unsuccessful hiring choices—while they overlook potential sequential dependencies between consecutive choices of the same employer. This work addresses these shortcomings by building a framework that uses job-application characteristics to provide recommendations that (1) are unlikely to yield adverse outcomes (performance-aware) and (2) capture the potentially evolving hiring preferences of employers (sequence-aware). Application of this framework on hiring decisions from an online labor market shows that it recommends job applicants who are likely to get hired and perform well. A comparison with advanced alternative recommender systems illustrates the benefits of modeling performance-aware and sequence-aware recommendations. An empirical adaptation of our approach in an alternative context (restaurant recommendations) illustrates its generalizability and highlights its potential implications for users, employers, workers, and markets.