Graph-Based Scientific Paper Recommendation

From scientific citation graphs to an interactive paper recommender

Graph learning and metadata-aware recommendation over a scientific citation network

This project studies scientific-paper recommendation by combining graph representations, learned node embeddings and bibliographic metadata. It includes exploratory data mining, classical machine-learning baselines, Graph Neural Network experiments and an interactive Streamlit application.

Motivation: paper relevance is not captured by title keywords alone. Citation-network proximity can reveal relationships between publications, while bibliographic attributes provide interpretable ways to adapt recommendations. The project explores how to combine learned graph-based similarity with metadata-aware ranking.

Architecture and project walkthrough

• System design — offline graph experiments, online recommendation, two Mermaid diagrams and technical decisions
• Guida italiana — problema affrontato, spiegazione semplice e preparazione al colloquio

Important boundary: GNN/model experiments run offline. Streamlit loads a precomputed node-embedding tensor and a metadata table; no GNN training or message passing happens during an interactive title search.

What the project explores

The experimental workflow compares several approaches:

• classical ML baselines;
• structural graph features;
• PCA / t-SNE analysis;
• Graph Convolutional Networks;
• GraphSAGE;
• Graph Attention Networks;
• learned node embeddings;
• ensemble comparisons.

The recommendation application uses learned graph embeddings together with publication metadata to rank papers related to a user-selected title.

Recommendation pipeline

Paper / title query
        ↓
Exact or fuzzy matching
        ↓
Graph embedding lookup
        ↓
Cosine similarity
        ↓
Metadata-aware re-ranking
        ↓
Interactive recommendations

The additive score is cosine similarity + configurable same-category bonus + shared-author bonus + PageRank contribution. Citation thresholds and scientific concepts act as filters, not direct score terms.

Interactive application

The Streamlit application is implemented in streamlit_app.py and provides:

• GitHub OAuth identity integration (prototype; recommendations are not fully authorization-gated);
• exact and fuzzy title search;
• embedding-based similarity;
• configurable ranking weights;
• metadata filters;
• recommendation cards;
• session-only favourites (not persisted in a user database);
• CSV export.

OAuth credentials are read from Streamlit secrets and are never committed.

Representative results

Model comparison

Citation-network neighbourhood

Additional figures retained for documentation cover degree distribution and XGBoost feature importance.

Data enrichment

scripts/openalex_enrichment.py enriches publication identifiers through the OpenAlex API with metadata including:

• title;
• authors;
• institutions;
• concepts;
• citation count;
• publication date.

Repository structure

Data-Mining-Streamlit/
├── streamlit_app.py
├── notebooks/
│   └── graph_recommendation_analysis.ipynb
├── scripts/
│   └── openalex_enrichment.py
├── data/
│   └── README.md
├── checkpoints/
│   └── README.md
├── docs/
│   └── figures/
├── requirements.txt
├── runtime.txt
└── README.md

Large experiment outputs, prediction arrays, training checkpoints, generated HTML visualizations and intermediate datasets are intentionally excluded from the portfolio repository.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure GitHub OAuth:

mkdir -p .streamlit
cp .streamlit/secrets.toml.example .streamlit/secrets.toml

Then provide the two required but not tracked runtime artifacts described in:

• Metadata table: data/df_final3.csv
• Embedding tensor: checkpoints/deepgcn_node_embeddings.pt

Without these files, the application stops with an explanatory error. The stored node_idx mapping must match the rows of the embedding tensor. The GitHub OAuth integration additionally requires appropriate secrets in .streamlit/secrets.toml.

Run the app with:

streamlit run streamlit_app.py

Tech stack

Python · PyTorch · Graph Neural Networks · scikit-learn · Pandas · NumPy · NetworkX · Streamlit · OpenAlex · OAuth · RapidFuzz

Author

Paolo Pangallo
M.Sc. Computer Engineering — Artificial Intelligence
University of Calabria