Scribd
Upload
13 views2 pages

Graph Recommender Programming Report

The document outlines a graph-based job recommendation system that utilizes SBERT for semantic profiling, PyTorch Geometric for graph learning, and Streamlit for deployment, offering personalized job suggestions with interpretability and fairness insights. Key components include feature extraction, graph construction, GNN training, recommendation generation, and fairness auditing. Future enhancements aim to improve model performance and address cold-start challenges.

Uploaded by

nasrin.sabet.gh
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
13 views2 pages

Graph Recommender Programming Report

The document outlines a graph-based job recommendation system that utilizes SBERT for semantic profiling, PyTorch Geometric for graph learning, and Streamlit for deployment, offering personalized job suggestions with interpretability and fairness insights. Key components include feature extraction, graph construction, GNN training, recommendation generation, and fairness auditing. Future enhancements aim to improve model performance and address cold-start challenges.

Uploaded by

nasrin.sabet.gh
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd

Graph-Based Job Recommendation

System
Prepared by: Nasrin Sabet

1. Project Overview
This project implements a personalized and fair job recommendation system using a graph-
based approach. By integrating SBERT for semantic profiling, PyTorch Geometric for graph
learning, and Streamlit for deployment, the system provides top-N job suggestions along
with interpretability and fairness insights.

2. System Components
Below is a breakdown of the system's components by module:

• feature_extraction.py:
- Uses SBERT (Sentence-BERT) to embed resume and job descriptions into high-
dimensional semantic vectors.
- Output is used for initializing user/job/skill node features.

• build_graph.py:
- Loads CSVs ([Link], [Link], [Link]).
- Constructs a heterogeneous NetworkX graph: nodes = users, jobs, skills; edges = User–
Skill, Job–Skill.
- Graph is saved in .gpickle format for reuse.

• graph_gnn.py:
- Defines and trains a GNN model (GraphSAGE or LightGCN) using PyTorch Geometric.
- Learns node embeddings from the constructed graph.
- Stores embeddings in gnn_node_embeddings.pt.

• graph_recommender.py:
- Loads trained node embeddings.
- Computes cosine similarity between user and job embeddings.
- Returns ranked job recommendations per user (Top-N).

• explanation_builder.py:
- Generates explanation text for each recommendation.
- Example: 'User matches 82% of required job skills.'
• fairness_metrics.py:
- Audits fairness of recommendations using Fairlearn / AIF360.
- Outputs metrics: Demographic Parity (DP), Equalized Odds (EO), Disparate Impact (DI).

• ui_app.py:
- Implements a Streamlit interface where users can upload resumes and receive suggestions
interactively.

3. Datasets and Preprocessing


- Real and synthetic resumes ([Link] to [Link])
- Job postings from Kaggle and [Link]
- Skills mapping in [Link]
- Users metadata in [Link]

Preprocessing includes:
- SBERT embedding generation for text fields
- Graph construction via node and edge creation
- Node IDs are mapped and normalized for training

4. Results
- Final user and job embeddings stored as .npy/.pt
- Recommendations saved in gnn_recommendations.csv
- Fairness metrics (from fairness_report.txt):
- DP: 0.000 | EO: 0.000 | DI: 1.000 (Perfect fairness on education)
- All users receive interpretable job match reports
- Streamlit UI successfully demonstrates complete pipeline

5. Future Extensions
- Upgrade GNN to Relational GCN or HGT for better heterogeneous learning
- Implement temporal split (LinkSAGE-style) to simulate chronological training/testing
- Integrate SHAP or interpretable attention into explanation_builder
- Handle cold-start more robustly using GNP or TIMBRE
- Add support for few-shot user/job generalization

You might also like