🔍 ResearchRover

ResearchRover is a full-stack, AI-powered Research Paper Recommendation System that combines Graph Neural Networks (GNN) and Semantic Embeddings (SBERT) to recommend relevant academic papers. It includes an interactive Streamlit UI for real-time use.

---

📌 Features

• 🧠 GNN-Based Recommendations using paper similarity graph
• 🗣️ SBERT Embeddings for deep semantic understanding of abstracts
• 🧪 Hybrid Recommender: combines structural (GNN) + semantic (SBERT)
• 🌐 Streamlit Web UI to explore recommendations
• 🔗 Link Prediction using GCN + MLP to suggest hidden connections

---

📁 Project Structure

ResearchRover/ │ ├── streamlit_app.py # Streamlit UI frontend ├── sample_1000_arxiv.csv # Cleaned and preprocessed paper data (1000 samples) ├── gnn_embeddings.pt # Learned GNN embeddings for each paper ├── sbert_embeddings.pt # Precomputed SBERT embeddings ├── requirements.txt # Required Python packages └── README.md # This file

---

⚙️ How It Works

🧾 Step 1: Data Preparation

• Load and clean two datasets: arxiv_data_210930-054931.csv and arvix_data.csv
• Standardize column names (terms, title, abstract)
• Combine datasets and remove duplicates/nulls
• Sample 1000 papers for demo scale

🔁 Step 2: Graph Construction

• Compute TF-IDF embeddings of abstracts
• Calculate cosine similarity between papers
• Construct an undirected similarity graph (NetworkX):
  • Nodes: papers
  • Edges: if similarity > 0.3

🧠 Step 3: Graph Neural Network (GNN)

• Train a 2-layer GCN (Graph Convolutional Network) using torch_geometric
• Node features: TF-IDF vectors
• Output: 64-dim GNN embeddings per paper

class GCNRecommender(nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels):
        ...
🔗 Step 4: Link Prediction
Positive edges = actual graph edges

Negative edges = random non-edges

Train an MLP (LinkPredictor) to classify whether an edge should exist

python
Copy
Edit
class LinkPredictor(nn.Module):
    def forward(self, x_i, x_j):
        x = torch.cat([x_i, x_j], dim=1)
        ...
💬 Step 5: SBERT Semantic Embeddings
Use sentence-transformers to encode abstract text

all-MiniLM-L6-v2 model outputs 384-dim embedding per abstract

sbert_model = SentenceTransformer("all-MiniLM-L6-v2")
sbert_embeddings = sbert_model.encode(abstracts)
🔀 Step 6: Hybrid Recommender
Combine GNN + SBERT embeddings:

hybrid_embeddings = [GNN | SBERT]  --> shape: [1000, 448]
Use cosine similarity over hybrid vector to recommend papers

🌐 Streamlit UI
Features:
Select paper from dropdown or paste a new abstract

View top-5 hybrid recommended papers

Supports cold-start (no graph) queries

Usage:

streamlit run streamlit_app.py