GNN_Book_Recommendation

A GNN-based book recommendation system models users, books, and their interactions as a graph, leveraging graph neural networks to capture complex, high-order relationships for more accurate and personalized recommendations.

Dataset

This project utilizes the Amazon Book Dataset. The raw data consists of the following files:

• t10k-images-idx3-ubyte.gz
• t10k-labels-idx1-ubyte.gz

---

Dataset Preview

Before diving into model training, it is highly recommended to get familiar with the data distribution and content.
I have written a utility script for this purpose: You can run datasetpreview.py to quickly preview the dataset details, including data volume, label mapping, and sample structures.

---

Model Architecture

After becoming familiar with the dataset, you can start training your own recommendation engine.

This project adopts LightGCN, a lightweight yet powerful variant of Graph Neural Networks (GNN). By removing unnecessary feature transformations and non-linearities, LightGCN efficiently captures high-order collaborative filtering signals between users and books.

---

Results & Visualization

The trained recommendation model successfully generates personalized suggestions, providing the top 10 books with the highest confidence scores for every user.

To make the results intuitive, I have applied visual analytics to the recommendation outputs. You can check out the following visualization materials included in the repository:

About Graph Neural Networks (GNN)

1. Graph Construction (Data Modeling)

The foundation of the system is how we define the graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$.

Element	Description
Nodes ($\mathcal{V}$)	Two main types: User nodes ($U$) and Book nodes ($B$). Optionally add Knowledge Graph entities (Author, Publisher, Genre).
Edges ($\mathcal{E}$)	Interaction behaviors: rate, purchase, add_to_cart, like, review.
Node Features	User: age, location; Book: title/blurb embeddings (BERT), price, avg. rating.
Edge Weights	Implicit (0/1) or explicit (rating score, confidence level).

Typically modeled as a Bipartite Graph, where edges only exist between $U$ and $B$.

---

2. GNN Architecture Design

2.1 Core Idea: Message Passing

GNNs learn node representations by iteratively aggregating feature information from neighboring nodes.

$$ \mathbf{h}_v^{(l+1)} = \sigma \left( \mathbf{W}^{(l)} \cdot \text{AGG}\left({\mathbf{h}_u^{(l)} : u \in \mathcal{N}(v)}\right) \right) $$

2.2 Recommended Backbones

For recommendation scenarios, simpler and more efficient models usually outperform complex ones:

1. LightGCN (Highly Recommended)
   • Removes feature transformation matrices and non-linear activation functions.
   • Purely uses neighborhood aggregation to refine embeddings—lightweight and state-of-the-art for collaborative filtering.
2. NGCF (Neural Graph Collaborative Filtering)
   • Exploits high-order connectivities in user-item bipartite graphs explicitly.
3. GraphSAGE / GAT
   • Useful when incorporating rich side information (e.g., book abstracts) and needing attention mechanisms to weigh neighbor importance.

---

3. Training & Prediction Pipeline

Step 1: Embedding Propagation

Stack $K$ layers of graph convolution to capture $K$-order connectivity (e.g., "Users who read this also read that").

Step 2: Representation Fusion

Combine embeddings from all layers (usually via weighted sum or concatenation):

$$ \mathbf{e}_u = \frac{1}{K+1} \sum_{k=0}^{K} \mathbf{e}_u^{(k)} $$

Step 3: Interaction Prediction

Calculate the matching score between user $u$ and book $i$:

$$ \hat{y}_{ui} = \mathbf{e}_u^\top \mathbf{e}_i $$

Step 4: Optimization Objective

Use BPR (Bayesian Personalized Ranking) loss to ensure the predicted score of a positively interacted book is higher than an unobserved one:

$$ \mathcal{L}_{BPR} = -\sum_{(u,i,j) \in \mathcal{O}} \ln \sigma(\hat{y}_{ui} - \hat{y}_{uj}) + \lambda|\Theta|^2 $$

---