matty6409/job-gnn-recommender

GNN Job Recommendation System v3 — Jobs & Skills

Simplified heterogeneous GNN for job-skill recommendation, trained on Hong Kong job market data.

Changes from v2

• 2 node types (dropped company & course — too noisy/volatile)
• 3 edge types: job→needs→skill, job→relatedto→job, skill→relatedto→skill
• Better dedup: Embedding-based clustering (paraphrase-multilingual-MiniLM-L12-v2) instead of exact-match
• skill↔skill edges: Derived from co-occurrence in same job listing

Architecture

• Text Encoder: matty6409/job-classifier-minilm-v4 (384-dim)
• GNN: GraphSAGE (hidden=256, 2 layers, residual + dropout)
• Predictor: Relation-conditioned MLP with learned relation embeddings

Graph

Type	Count
Jobs	44,673
Skills	44,835

Edge Types

ID	Relation
0	job → needs → skill
1	job → relatedto → job
2	skill → relatedto → skill