IGDHRS: An Intelligent Graph-based Deep Hybrid Recommender System

This repository contains the official implementation of IGDHRS, a novel hybrid recommender system that integrates graph-based user similarity, a Deep Denoising Graph Convolutional Autoencoder (DDGCAE), and an automata-driven adaptive thresholding mechanism. IGDHRS effectively addresses key challenges in recommendation systems, such as data sparsity and the cold-start, by combining auxiliary user/item metadata with graph-derived structural features.

📖 Overview

IGDHRS constructs a user similarity graph (SG) using a dynamic threshold ( T_s ), which is automatically optimized via Learning Automata (LA) based on feedback from recommendation performance (e.g., RMSE, MAE, precision, recall). The system extracts graph-based features, merges them with auxiliary user/item metadata, and applies a DDGCAE to learn robust representations. Clustering and collaborative filtering are then applied to generate accurate recommendations.

🚀 Key Features

• Adaptive similarity threshold tuning via Learning Automata (LA)
• Six comprehensive graph features: PageRank, Degree Centrality, Closeness Centrality, Betweenness Centrality, Load Centrality, and Average Neighbor Degree
• Fusion of structural and auxiliary (demographic/content) features
• Deep GCN-based autoencoder with corruption-aware denoising
• Spectral clustering on graph embeddings
• Handles both cold-start and sparse data scenarios
• Fully modular, Pythonic, and configurable via config.yaml

⚙️ Installation

1. Clone the repository:

git clone https://github.com/yourusername/IGDHRS.git
cd IGDHRS

2. (Optional) Create and activate a virtual environment:

python3 -m venv venv
source venv/bin/activate       # Linux/macOS
venv\Scripts\activate          # Windows

3. Install required dependencies:

pip install -r requirements.txt

Note: Dependencies include: numpy, scikit-learn, scipy, networkx, PyYAML.

⚙️ Configuration

All parameters are defined in the config.yaml file. Key options include:

graph:
  similarity_threshold_init: 0.015
  threshold_range: [0.001, 0.03]
  automata_type: LRP
  reward_learning_rate: 0.1
  penalty_learning_rate: 0.1

ddgcae:
  corruption_probability: 0.3
  layers: 3
  regularization_lambda: 0.01

clustering:
  num_clusters: 8

▶️ Usage

The system is organized into modular components:

File	Description
src/graph_construction.py	Constructs the user similarity graph using adaptive LA mechanism
src/feature_extraction.py	Extracts graph-based and auxiliary features
src/ddgcae.py	Learns latent embeddings using the DDGCAE model
src/clustering.py	Performs spectral clustering and new user assignment
src/recommender.py	Generates predicted ratings and recommendation lists
main.py	Orchestrates the full training and recommendation pipeline

To run the complete system:

python main.py --config config.yaml

📂 Outputs

The system produces the following output files (saved in /outputs/ directory):

• embeddings.npy: Final latent user embeddings
• clusters.npy: Cluster assignment of users
• predicted_ratings.npy: Full user–item prediction matrix
• logs/: Evaluation logs and metadata

🧠 Model Components

Similarity Graph Construction

• Threshold ( T_s ) is adjusted using Learning Automata (LA)
• Actions: Increase, Decrease, or Unchanged
• Feedback: Based on RMSE, MAE, precision, recall
• Variants: LRP, LRI, LRεP schemes supported

Feature Extraction

• Structural features (x6): PR, DC, CC, BC, LC, AND
• Auxiliary features: One-hot encoded demographics/items
• Combined into a unified feature matrix ( F_t )

DDGCAE: Deep Denoising Graph Convolutional Autoencoder

• Inputs: corrupted node features ( \tilde{X} ), graph adjacency matrix ( A )
• Multi-layer GCN with spectral convolutions
• Closed-form weight updates; no backpropagation
• Robust embedding learning via corruption marginalization

Clustering & Recommendation

• Spectral clustering on ( Z_2 = 0.5(|Z| + |ZZ^T|) )
• New users: assigned to nearest cluster using cosine similarity
• Predictions: item scores estimated via cluster-wise rating averages

📊 Evaluation

• Supports: RMSE, MAE, Precision@k, Recall@k
• Per-epoch evaluation logging
• Early stopping supported (optional)

📚 Citation

If you use this repository in your research, please cite the corresponding paper:

[Insert BibTeX or citation information here once published]

🪪 License

This project is released under the MIT License. You are free to use, modify, and distribute this work with attribution.

📬 Contact

For questions, suggestions, or collaboration:

Milad Payandeh
📧 milad71payandeh@gmail.com
🌐 https://miladpayandeh.com

---

Thank you for using IGDHRS! If you find this repository helpful, please consider starring ⭐ the project.