KG Job & Candidate Matcher (HireMatch AI)

Streamlit UI for ranking jobs <=> candidates over a Neo4j knowledge graph using Rasch scores, TransR embeddings, and skill coverage. The app also integrates optional LLM helpers (OpenAI, Gemini, or local Ollama) for explanations and skill-coaching guidance.

This project is also referred to as HireMatch AI, highlighting its role as a two-way conversational job–candidate matching assistant built on knowledge graphs and statistical learning models.

System Architecture

User Interface

Candidate View

• Upload resume (PDF/TXT)

• Link resume to candidate node

• View top-K job recommendations

• Identify missing skills for target jobs

• Optional AI explanations and coaching tips

Employer View

• Search and select job postings

• Rank best-fit candidates

• Recruiter-friendly AI summaries

DATASET in Drive

https://drive.google.com/drive/folders/1cZb0mlXbhNcTOQklmxNYmzxMt7WGBtkr?usp=sharing

Problem Statement & Goals

Hiring platforms often rely on keyword matching, which fails to capture skill difficulty, candidate strength, and contextual relevance. This project aims to:

• Match candidates to relevant jobs and jobs to relevant candidates
• Use a Neo4j knowledge graph to model skills, jobs, candidates, companies, and locations
• Combine Rasch statistical modeling with TransR graph embeddings
• Provide interpretable and explainable rankings for recruiters and candidates
• Support optional LLM-based explanations and coaching

High-level pipeline:

1. Data ingestion: Job postings and resumes are parsed to extract skills, locations, and metadata
2. Knowledge graph construction: Entities and relationships are stored in Neo4j
3. Metric computation: Rasch model -> skill difficulty, candidate ability, job difficulty. TransR -> relational graph embeddings
4. Recommendation engine: Composite scoring using coverage, Rasch, and TransR
5. Streamlit UI: Interactive candidate and recruiter experiences

Core Algorithms & Models

1. Knowledge Graph (Neo4j)

The Neo4j graph represents:

• Nodes: JobPost, Candidate, Skill, Company, Location, Chunk
• Edges: HAS_SKILL, REQUIRES_SKILL, POSTED_BY, LOCATED_IN, etc.

This structure enables rich graph queries and embedding-based reasoning.

2. Rasch Model (Skill & Ability Scoring)

The Rasch model converts skill rarity into interpretable difficulty and ability scores:

• Rare skills → higher difficulty
• Candidate ability increases with possession of difficult skills
• Job difficulty aggregates required skill difficulties

This allows principled comparison between candidates and jobs beyond raw skill counts.

3. TransR Graph Embeddings

TransR embeds entities and relations into relation-specific vector spaces:

• Captures latent semantic compatibility between jobs and candidates
• Supports similarity-based ranking when direct skill overlap is limited
• Trained offline and written back to Neo4j for fast retrieval

4. Composite Ranking Function

Final ranking score is a weighted combination of:

• Skill Coverage (α) – overlap between candidate skills and job requirements
• Rasch Signal (β) – candidate ability vs. job difficulty
• TransR Similarity (γ) – embedding proximity

Weights are adjustable in the Streamlit UI.

Datasets

Job Postings Dataset

• 124,000+ LinkedIn job postings (2023–2024)
• Scraped and stored originally as CSV
• Fields used include job title, salary range, employment type, location, and skills

Resume Dataset

• Real anonymized and synthetic resumes
• Stored in JSON format
• Includes experience, education, skills, and embedded text chunks

All data is cleaned, normalized, converted to JSONL, and integrated into Neo4j.

Evaluation

Evaluation performed at K = 5:

• Precision@5: 0.86
• Recall@5: 0.82
• NDCG@5: 0.78
• MAP@5: 0.67

These results indicate strong ranking quality and effective blending of Rasch, TransR, and coverage signals.

Quick Start

git clone <repo>
cd SWMP
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt