Knowledge-Aware Graph Prompt Tuning for Cross-Domain Recommendation

A cross-domain recommendation system that combines knowledge graph embeddings, graph attention networks, and prompt tuning to transfer user preferences across Amazon product categories.

Table of Contents

• Overview
• Architecture
• Results
• Project Structure
• Getting Started
• Dataset
• Technical Details
• References
• Authors

Overview

Cross-domain recommendation addresses the data sparsity problem by transferring knowledge from a data-rich source domain to a target domain. Traditional approaches require retraining the full model for each domain pair. This project proposes adaptive prompt tuning — small learnable modules that adapt frozen graph embeddings — to efficiently bridge domains without costly retraining.

The system constructs a unified knowledge graph across two Amazon product categories, learns structural and semantic embeddings via TransE and a Graph Attention Network, and applies prompt-based adaptation for the final recommendation task.

Architecture

┌─────────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────────┐     ┌───────────────┐
│  Amazon Reviews  │────►│ Knowledge    │────►│   TransE     │────►│  GAT Encoder     │────►│ Prompt Tuning │
│  (Two Domains)   │     │ Graph (KG)   │     │ (Embeddings) │     │ (Contrastive)    │     │ + BPR Ranker  │
└─────────────────┘     └──────────────┘     └──────────────┘     └──────────────────┘     └───────────────┘

Stage 1: Knowledge Graph Construction

• Identifies overlapping users between two Amazon product categories
• Constructs triples with three relation types:
  • Purchase (user → item)
  • Produced by (item → brand, from product metadata)
  • Described as (item → keyword, via TF-IDF extraction)

Stage 2: TransE Embedding

• Learns 64-dimensional entity and relation embeddings
• Trained with margin ranking loss and negative sampling over 20 epochs

Stage 3: GAT Encoder

• 2-layer Graph Attention Network with contrastive self-supervised learning
• Dual-view generation via stochastic edge dropout
• Gradient checkpointing for GPU memory efficiency

Stage 4: Prompt-Tuned Recommendation

Three prompt strategies are compared in an ablation study:

• Baseline (No Prompt): Raw GAT embeddings fed to the BPR predictor
• Static Prompt: A single learnable vector added uniformly to all user embeddings
• Adaptive Prompt (Ours): A density-aware gating mechanism that blends global domain knowledge with personalized user signals derived from neighbor item aggregation

Results

Experiment 1: Electronics ↔ Cell Phones & Accessories

Method	HR@10	Improvement
Baseline (Pure GAT)	0.3403	—
Static Prompt	0.3452	+1.45%
Adaptive Prompt (Ours)	0.4033	+18.53%

Experiment 2: Electronics ↔ Beauty

Method	HR@10	Improvement
Baseline (Pure GAT)	0.4662	—
Static Prompt	0.4701	+0.83%
Adaptive Prompt (Ours)	0.4667	+0.09%

Key Findings

• The adaptive prompt achieves a significant +18.53% improvement on the Electronics ↔ Cell Phones pair, where domain overlap is structurally strong (both are technology categories).
• On the more distant Electronics ↔ Beauty pair, all methods perform similarly, suggesting that prompt tuning benefits are largest when domains share meaningful structural and semantic connections.

Project Structure

├── WSC_Cell_Phones_electronics.ipynb   # Experiment 1: Electronics ↔ Cell Phones
├── WSC_Electronics_Beauty.ipynb        # Experiment 2: Electronics ↔ Beauty
├── assets/                             # Result visualizations
│   ├── results_cell_phones_electronics.png
│   └── results_electronics_beauty.png
├── requirements.txt                    # Python dependencies
└── README.md

Getting Started

Prerequisites

• Python 3.9+
• CUDA-compatible GPU (recommended for training)

Installation

git clone https://github.com/deepanshu-prog/Recommendation_System.git
cd Recommendation_System
pip install -r requirements.txt

Running the Experiments

The notebooks are designed for Google Colab with a GPU runtime:

1. Open either notebook in Google Colab
2. Set runtime to GPU (Runtime → Change runtime type → T4 GPU)
3. Run all cells sequentially

The notebooks handle data download, preprocessing, training, and evaluation end-to-end.

Dataset

All data is sourced from the Amazon Reviews 2023 dataset (McAuley Lab, UCSD).

Domain Pair	Overlapping Users	Items	KG Triples
Electronics ↔ Cell Phones	100,000 (capped)	249,277	826,177
Electronics ↔ Beauty	355,607	467,363	3,253,649

Data files are downloaded automatically within the notebooks. No manual setup is required.

Technical Details

Component	Configuration
Embedding Dimension	64
TransE Epochs	20
GAT Layers	2 (1 attention head each)
GAT Training	10 epochs, contrastive loss
Dropout	0.2
BPR Batch Size	2048
Optimizer	Adam
Evaluation Metric	HR@10 (99 negative samples)
Seed	42

Training Strategy

1. Warmup Phase: Train only the BPR predictor with frozen embeddings (10–20 epochs)
2. Joint Phase: Train prompt module and predictor jointly with differential learning rates (25 epochs)

References

• Knowledge-Aware Graph Prompt Tuning for Cross-Domain Recommendation
• Amazon Reviews 2023 Dataset
• Bordes et al., Translating Embeddings for Modeling Multi-relational Data (TransE)
• Velickovic et al., Graph Attention Networks (GAT)
• Rendle et al., BPR: Bayesian Personalized Ranking from Implicit Feedback

Authors

• Deepanshu — @deepanshu-prog
• Zaed Rizwan — @ZR12345

IT752 Term Project — Group 13