A Hybrid Machine Learning Approach for Automated Resume Screening and Candidate Ranking

Abstract

In this empirical research, we introduce a new hybrid machine learning system to overcome the limitations of vocabulary and semantic mismatches in automated resume screening and candidate ranking systems. The architecture proposed here integrates three types of features: sparse Term Frequency-Inverse Document Frequency (TF-IDF), static Word2Vec and fine-tuned Bidirectional Encoder Representations from Transformers (BERT) contextual features. The framework was tested with a synthetic Kaggle recruitment dataset of 30,000 candidate profiles and was tested against 5,000 professional resumes. The experimental results demonstrate that the proposed pipeline achieves 93.2% classification accuracy, 93.6% precision, 92.8% F1-score and 0.91 Normalized Discounted Cumulative Gain (NDCG@10) value. The results show better semantic match between resumes and job descriptions, 78.3% reduction in the time spent on processing candidates, and consistent candidate ranking in various organizational contexts.