Content-Aware Re-Ranking of Job Recommendations with Gradient-Boosted Trees

Abstract

Online job marketplaces rank openings for each candidate primarily from behavioral signals, which capture popularity and engagement but are largely blind to how well an opening matches a candidate's profile. We study whether a lightweight re-ranking layer can improve such a production ranking by exploiting candidate-opening content-match signals. Using one month of impression logs from Kariyernet, a large Turkish online recruitment platform (13.08M} impressions and 85,698 application events), we train a gradient-boosted decision tree model with the LambdaRank objective on 35 features that combine content-match indicators-overlap in city, sector, position, seniority, experience, education, and salary-with a small set of contextual signals, and we evaluate it under a strictly temporal train/validation/test split with leakage-safe feature construction. Using a single-impression attribution of each application that prevents the metric inflation a naive labeling would introduce, the re-ranker improves the test normalized discounted cumulative gain at rank 10 by approximately 18.6 percent (with a standard deviation of 0.9 points across five random seeds) relative to the production ordering, with the largest gains at the top of the list. A forward ablation and a comparison with learned baselines attribute the improvement to the content-match features rather than to the list wise objective specifically. These results indicate that content-aware re-ranking is an effective and low-cost complement to behavioral recommendation in the labor-market domain.