EXAI-ResumeIntel
Explainable AI for automated resume analysis
- 70.73%
- overall accuracy
- 100%
- SHAP–LIME agreement
- 0.993
- best ROC-AUC
- 2,484
- resumes evaluated
“Automated screening assigns match scores without any rationale — candidates can’t act on them and recruiters can’t audit them.”
Resume screening systems output a number and nothing else, preventing candidates from understanding which skills to develop and preventing recruiters from auditing shortlisting decisions.
A five-layer framework. A hierarchical domain ontology with 346 alias mappings across 22 canonical skill nodes detects implicit skills — a resume listing 'YOLOv8', 'COCO dataset' and 'anchor boxes' is correctly read as carrying computer vision expertise even though neither canonical term appears. TF-IDF with Truncated SVD produces 150-dimensional semantic embeddings, a four-component engine yields an interpretable score, exact Shapley values are independently validated by LIME, and counterfactual what-if explanations quantify the gain from acquiring each missing skill.
What makes it
work.
Exact Shapley values satisfying all four game-theoretic axioms, cross-checked by LIME across 300 perturbation samples.
Evaluated on 2,484 real resumes across 24 job categories with 5-fold stratified cross-validation.
Ontological inference alone lifts the Machine Learning Engineer role from a 38% semantic-only baseline to 74%.
Models and dataset published to the Hugging Face Hub alongside a live Space.
Built with.
- Python
- SHAP
- LIME
- scikit-learn
- Truncated SVD
- FastAPI
- Streamlit