GNN Model with 93% Accuracy for Facebook Page-Page Network Node Classification with TSNE Visualization

[liammulhern]

GNN Model with 93% Accuracy for Facebook Page-Page Network Node Classification with TSNE Visualization

This project introduces a multi-layer graph neural network (GNN) for semi-supervised, multi-class node classification on the Facebook Large Page-Page Network dataset, achieving 93.14% accuracy. The network classifies nodes (representing Facebook pages) into four categories: Politicians, Government Organizations, Television Shows, and Companies.

Key features of PR:

Modules:

• dataset.py: Loads and preprocesses data.
• main.py: CLI for training and inference.
• modules.py: Defines GNN architecture.
• train.py: Manages training, validation, and metric logging.
• predict.py: Runs model inference and visualizations.

Execution:

Supports training (--train --save --load), inference (--inference <index>), and visualization (--display) through CLI.

GNN Architecture:

1. Uses multilayer perceptrons (MLPs) and sparse layers, transforming node features into learned embeddings for classification.
2. Incorporates ReLU activation and log softmax for output.

Training:

• Learning Rate: 1e-4
• Epochs: 100
• Optimizer: Adam
• Loss: Cross-Entropy

Results:

Achieves 93.14% accuracy; training and validation metrics show potential overfitting.
TSNE visualizations show clearer clustering post-training, indicating successful categorization.

[gayanku]

This is an initial inspection, no action is required at this point
GNN FB Page-Page Network dataset ---- Normal Difficulty

Category		Marks	Comments
Algorithm solves the problem	5	4.5	Some overfitting seen. No early stop
Implementation functions as intended	3	3	2 layer GCN, torch_geometric/GCNConv
Good design	1	1	Modular, Reusable
Commenting	1	1	Meaningful docstrings and comments
Algorithm above Normal Difficulty	5	5
Algorithm is Hard difficulty	5	0	Normal Difficulty
Section IV : Max mark 15 from 20		14.5

TSNE / UMAP:
Good, though would have expected to see a bit more separation at the reported accuracy.

• Discussion: Good, with a bit more depth useful.

Suggestions:
Redundant code in repo - SparseLayer. Is this needed?
Also doc mentions the use of the sparse layer, though the implemented model does not use this layer.
Code can have some early stop, ex: train loop should save best model by keeping track of current best accuracy.
Could add a hyper parameter search or discuss how the values were obtained.

[gayanku]

Marking

Good/OK/Fair Practice (Design/Commenting, TF/Torch Usage)
	Adequate design and implementation. Redundant code in repo - SparseLayer.	-1
	Spacing and comments.
	Header blocks.
Recognition Problem
	OK solution to problem. Some overfitting seen. No early stop	-2
	Driver Script present.
	File structure present.
	Good Usage & Demo & Visualisation & Data usage.
	Module present.
	Commenting present.
	No Data leakage found.
	Difficulty : Normal. GNN Task.	-5
Commit Log
	Good Meaningful commit messages.
	Good Progressive commits.
Documentation
	Readme :Good.
	Model/technical explanation :Acceptable.	-1
	Description and Comments :Good.
	Markdown used and PDF NOT submitted.	-2
Pull Request
	Successful Pull Request (Working Algorithm Delivered on Time in Correct Branch).	-2
	Feedback action required: Feedback marks possible +2 if the requested changes are made. Update PR to correct branch.	-2
	Request Description is good.
TOTAL		-15

Marked as per the due date and changes after which aren't necessarily allowed to contribute to grade for fairness.
Subject to approval from Shakes

[gayanku]

Feedback marks possible +2 if the requested changes are made (see above).