Improving Reciprocal Job Recommendations via Text Embedding Models

In this paper, we enhance reciprocal job recommender systems by leveraging transformer-based text embeddings and contrastive learning. We introduce textual representations for the Turkish human resources domain, which is currently under-resourced. Reciprocal recommendation systems uniquely benefit both applicants and employers, necessitating balanced, accurate predictions on both sides. Building on the biDeepFM study, we incorporated text embeddings from pre-trained multilingual language models, followed by fine-tuning with a triplet loss objective. This approach includes hard negative mining, which introduces challenging distinctions in candidate-employer interactions to refine model performance. Experimental results on a Turkish job application dataset indicate that our approach improves AUC scores (up to one percentage point), for both employer and job seeker preferences. Our findings suggest that integrating fine-tuned, multilingual embeddings can enhance system performance, advancing the field of reciprocal recommendation.