JobMatchAI-An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI

Mayank Vyas; Abhijit Chakraborty; Vivek Gupta

cs.AIMar 15, 2026

JobMatchAI-An Intelligent Job Matching Platform Using Knowledge Graphs, Semantic Search and Explainable AI

Authors: Mayank Vyas, Abhijit Chakraborty, Vivek Gupta

Organizations: Arizona State University

Abstract

Recruiters and job seekers rely on search systems to navigate labor markets, making candidate matching engines critical for hiring outcomes. Most systems act as keyword filters, failing to handle skill synonyms and nonlinear careers, resulting in missed candidates and opaque match scores. We introduce JobMatchAI, a production-ready system integrating Transformer embeddings, skill knowledge graphs, and interpretable reranking. Our system optimizes utility across skill fit, experience, location, salary, and company preferences, providing factor-wise explanations through resume-driven search workflows. We release JobSearch-XS benchmark and a hybrid retrieval stack combining BM25, knowledge graph and semantic components to evaluate skill generalization. We assess system performance on JobSearch-XS across retrieval tasks, provide a demo video, a hosted website and installable package.

Figures & tables

Panel of Figure 1 — Figure 1: End-to-end JobMatchAI hybrid search pipeline architecture.

Factor ( $f$ )	$w_{f}$	Feature ( $\phi_{f}$ )
Skill match	0.35	Jaccard + KG relatedness bonus
Experience	0.25	Level-distance penalty
Location	0.15	Tiered: exact / state / remote
Salary	0.10	Expectation-to-midpoint ratio
Semantic sim.	0.10	$\cos(\mathbf{e}_{c},\mathbf{e}_{j})$
Company fit	0.05	Industry & size preference

Table 1: Default reranking weights. Users may adjust these interactively; the utility function recomputes in real time.

Panel of Figure 2 — Figure 2: JobMatchAI Platform Interface

Figure 3: Visualization of the JobMatchAI knowledge graph illustrating the relationships between candidates, jobs, skills, locations, and companies used for hybrid retrieval and explainable reranking.

BM25	Sem.	KG	NDCG@5	NDCG@10	R@50	R@100	P50 (ms)
✓	✗	✗	0.721	0.756	0.49	0.57	79
✗	✗	✓	0.758	0.790	1.00 *	1.00 *	–
✓	✓	✗	0.740	0.771	0.36	0.42	80
✓	✗	✓	0.747	0.776	0.33	0.39	81
✓	✓	✓	0.750	0.780	0.29	0.35	82
✓	✓	✓ +R	0.785	0.810	0.29	0.35	82

Table 2: Retrieval performance on JobSearch-XS (evaluated on human-verified gold labels). ✓/✗ indicate active retrieval channels.

Intent type	n	Mean	Std	IQR
Job title	3	0.90	0.09	0.08
Natural language	6	0.74	0.34	0.39
Skill synonym	6	0.60	0.31	0.46
Overall	15	0.72	0.30	0.45

Table 3: NDCG@10 distribution across gold queries by query intent (Hybrid+Reranker pipeline).

Extractor	NDCG@10	R@100
High-precision (current)	0.791	0.336
High-recall	0.817	0.326
$\Delta$ (recall $-$ precision)	+0.026	$-0.010$

Table 4: End-to-end retrieval quality under two skill-extractor operating points (Hybrid+Reranker).

Task-based evaluation
Metric	Mean $\pm$ Std
Top-5 relevance (Smart Search)	3.92 $\pm$ 0.79
Skill synonym handling	3.75 $\pm$ 0.87
Explanation helpfulness	4.17 $\pm$ 0.72
Weight adjustment usefulness	3.83 $\pm$ 0.94
Usability

Table 5: User study results (

N=20

). Scores are mean

\pm

std on a 1–5 Likert scale.

Feature	JobMatchAI	LinkedIn	Jobright	Teal	Jobscan
Semantic search	✓	~	✓	✗	✗
Knowledge graph	✓	✗	✗	✗	✗
Hybrid retrieval	✓	✗	✗	✗	✗
Explainable scores	✓	✗	✗	✗	~
Factor-level control	✓	✗	✗	✗	✗
Resume-driven search	✓	~	✓	✓	✓

Table 6: Feature comparison with representative job search tools. ✓= supported, ✗= not supported, ~= partial.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Field	Base P	Base R	Base F1	Ours P	Ours R	Ours F1	$\Delta$ F1
Name	–	–	0.33	–	–	0.64	+0.31
Skills	0.41	0.60	0.48	0.94	0.12	0.19	$-$ 0.29
Education	0.74	0.31	0.42	1.00	1.00	1.00	+0.58

Appendix

Table 7: Component-level resume parsing quality on a labeled set of 45 resumes. Baseline uses a single-strategy spaCy NER pipeline.

Model	Acc.	F1	AUROC	Notes
Pairwise MLP	0.70	0.45	0.72	Concatenated embedding features; reasonable baseline
Cross-encoder	0.30	0.34	0.51	Off-the-shelf without fine-tuning; poor calibration

Appendix

Table 8: Resume screening model performance on pairwise preference prediction (1,200 triples, 80/20 train/test split).

Dimension	Slice	NDCG@10	Recall@100	Notes
Location	Same-region	0.78	0.21	Best precision with location match
Location	Cross-region	0.50	0.85	High recall, low ranking precision
Location	Remote-friendly	0.66	0.57	Moderate precision and recall
Salary	Within band	0.77	0.23	Highest NDCG with salary overlap
Salary	Unknown/noisy	0.63	1.00	Perfect recall; fallback when salary missing
Seniority	Junior	0.62	0.92	High recall, lower ranking quality

Appendix

Table 9: Retrieval performance by location, salary, and seniority slices on JobSearch-XS (hybrid + reranker).

Statistic	Value
Total Queries	30
Total Jobs	1,283
Silver Labels	29,013
Gold Labels	30
Train / Dev / Test	15 / 6 / 9

Appendix

Table 10: JobSearch-XS dataset statistics.

	P50	P90	P95	Mean $\pm$ Std
Latency (ms)	81.55	84.66	86.00	81.67 $\pm$ 2.69

Appendix

Table 11: Latency profile (210 queries, 12.24 qps).

Metric	@1	@3	@5	@10	@20	@50
NDCG	0.667	0.720	0.733	0.756	0.759	0.772
Recall	0.010	0.037	0.061	0.131	0.262	0.574
Precision	0.667	0.733	0.747	0.773	0.767	0.703

Appendix

Table 12: Information retrieval metrics on JobSearch-XS (baseline).

Split	Queries	NDCG@10	MRR	P50 (ms)
Train	15	0.890	0.922	79.08
Dev	6	0.402	0.389	77.23
Test	9	0.769	0.749	78.23

Appendix

Table 13: Per-split retrieval results.

Profile	NDCG@10	R@10	MRR	P50 (ms)
Entry SWE	0.721	0.144	0.780	81.90
Senior Data Eng	0.642	0.148	0.747	82.32
H1B ML	0.642	0.148	0.747	83.10
Public Health	0.721	0.144	0.780	82.87
GIS Analyst	0.721	0.144	0.780	82.88

Appendix

Table 14: Per-profile personalized search results.

Job-Candidate Matching

See all 13 papers

ConFit v3: Improving Resume-Job Matching with LLM-based Re-Ranking
May 10, 2026Xiao Yu, Ruize Xu, Chengyuan Xue +6Reinforcement Learning Learning to Rank
Columbia University · New York, NY, USA · John Hopkins University +6
Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization
Apr 2, 2026Ansel Kaplan Erol, Seohee Yoon, Keenan Hom +1Job-Candidate Matching Recommender Systems
Georgia Institute of Technology, Atlanta, GA, USA
PJB: A Reasoning-Aware Benchmark for Person-Job Retrieval
Mar 18, 2026Guangzhi Wang, Xiaohui Yang, Kai Li +4Job-Candidate Matching Information Retrieval
CareerInternational Research Team

May 10, 2026J/K move · Enter open · S save

ConFit v3: Improving Resume-Job Matching with LLM-based Re-Ranking

Xiao Yu, Ruize Xu, Chengyuan Xue, Junyu Chen+5

Columbia University · New York, NY, USA · John Hopkins University+6

Reinforcement Learning Learning to Rank Job-Candidate Matching LLM Reranking LLM Post-Training

Open paper AskPDF arXiv

A reliable resume-job matching system helps a company find suitable candidates from a pool of resumes and helps a job seeker find relevant jobs from a list of job posts. While recent advances in embedding-based methods such as ConFit and ConFit v2 can efficiently retrieve candidates at scale, the lack of controllability and explainability limits their real-world adaptations. LLM-based re-rankers can address these limitations through reasoning, but existing training recipes are developed on short-document benchmarks and do not account for noise in real-world recruiting data. In this work, we first conduct a systematic analysis over the LLM re-ranker training pipeline for person-job fit, covering inference algorithm design, RL algorithm selection, data processing, and SFT distillation. We find that using multi-pass re-ranking, training with listwise RL objectives, removing noisy samples, and distilling from a stronger LLM before RL significantly improves re-ranking performance. We then aggregate these findings to train ConFit v3 with Qwen3-8B and Qwen3-32B on real-world person-job fit datasets, and find significant improvements over existing best person-job fit systems as well as strong LLMs such as GPT-5 and Claude Opus-4.5. We hope our findings provide useful insights for future research on adapting LLM-based re-rankers to person-job fit systems.

Abstract

Job-Candidate Matching

Explore similar work

ConFit v3: Improving Resume-Job Matching with LLM-based Re-Ranking

Synapse: Evolving Job-Person Fit with Explainable Two-phase Retrieval and LLM-guided Genetic Resume Optimization

PJB: A Reasoning-Aware Benchmark for Person-Job Retrieval