Web-Based Job Recommendation Based on LinkedIn Profiles Using Domain-Aware SBERT Retrieval and TF-IDF Reranking

Abstract

Online job search often relies on keyword matching, while the semantic relationship between candidate profiles and job descriptions may not be captured adequately. This study develops a web-based job recommendation system based on LinkedIn-style candidate profiles using SBERT retrieval and domain-aware TF-IDF reranking. Candidate profiles are constructed from target role, headline, skills, experience, education, preferred location, and work preference, with non-English input translated into English when needed. The job corpus consists of approximately 1.3 million job postings represented by precomputed 384-dimensional SBERT embeddings. The system retrieves initial candidates using cosine similarity and reranks them using TF-IDF similarity with domain, experience, and location constraints. Manual evaluation on 1,012 judged profile-job pairs shows that the proposed method achieves Precision@5 of 0.428, Precision@10 of 0.368, NDCG@10 of 0.531, and MRR of 0.605. An additional validated pseudo-label evaluation achieves Precision@5 of 0.840, with 83.33% agreement and a Cohen’s Kappa of 0.75 against human-checked samples. These results indicate that semantic retrieval combined with explainable domain-aware reranking can improve the relevance of web-based job recommendations.