CallosumNet

Unlearning on Spatio-Temporal Graphs through Subgraph Virtual Edge Reconstruction

CallosumNet is a partition-and-integrate framework for spatio-temporal graph unlearning, biologically inspired by the corpus callosum structure. It partitions a spatio-temporal graph into enhanced subgraphs, stitches them via a lightweight meta-graph, and enables exact (100%) node unlearning by retraining only the affected subgraph—without touching the rest.

Paper: Unlearning on Spatio-Temporal Graphs through Subgraph Virtual Edge Reconstruction (Under review at PVLDB 2026)

Key Features

• 100% Exact Unlearning — deleted nodes have zero residual influence on model outputs, fully compliant with GDPR/CCPA
• Sub-linear Unlearning Cost — retrain only the affected subgraph(s) + lightweight GGB layer (1–3 epochs), not the entire model
• Near Gold-Model Accuracy — 3.7% average relative MAE degradation vs. the gold model, compared with 29%–304% for baselines
• Backbone Agnostic — supports STGCN, ST-GAT, ST-GATV2, and ST-SAGE
• Four Real-World Datasets — RWW (23 nodes), PeMS08 (170 nodes), Global Weather (1,000 nodes), Human Mobility Flow (3,220 nodes)

Architecture

CallosumNet consists of two core components:

1. Enhanced Subgraph Construction (ESC) — partitions the graph along a correlation-driven backbone and inserts virtual ganglion edges + K-Ring boundary connections to preserve local spatio-temporal structure
2. Global Ganglion Bridging (GGB) — promotes key/boundary/ganglion nodes into a sparse meta-graph and fuses cross-partition information via a Transformer-based integration layer

Installation

# Create environment (tested on Python 3.10)
conda create -n callosumnet python=3.10 -y
conda activate callosumnet

# Install dependencies
pip install -r requirements.txt

Requirements: PyTorch, PyTorch Geometric, NumPy, SciPy, scikit-learn. See environment.yml for the full specification.

Data Preparation

PeMS08 is included in the PeMS08/ directory and ready to use.

For the other three datasets:

Dataset	Nodes	Time Steps	Source
RWW (Sewage Water Depth)	23	18,000	HydroNet
PeMS08 (Traffic Flow)	170	17,856	Included
Global Weather (Temperature)	1,000	18,000	NOAA PSL
Human Mobility Flow	3,220	3,000	Kang et al. 2020

Place each dataset folder in the project root directory.

Quick Start

Full-Graph Baseline (No Partitioning)

python callosumnet.py \
  --datasets PeMS08 \
  --model_type STGCN \
  --experiment_mode full_graph \
  --hidden_features 128 \
  --epochs 60 --patience 8 \
  --unlearn_rate 0.00 --seed 40 --profile_time

CallosumNet (ESC + GGB)

python callosumnet.py \
  --datasets PeMS08 \
  --model_type STGCN \
  --experiment_mode subgraph_aggregation \
  --num_partitions 4 \
  --hidden_features 128 \
  --epochs 60 --agg_epochs 30 \
  --patience 8 --agg_patience 6 \
  --agg_dmodel 48 \
  --agg_fusion cross \
  --ablation none \
  --unlearn_rate 0.00 --seed 40 --profile_time

Unlearning (10% Node Deletion)

python callosumnet.py \
  --datasets PeMS08 \
  --model_type STGCN \
  --experiment_mode subgraph_aggregation \
  --num_partitions 4 \
  --hidden_features 128 \
  --epochs 60 --agg_epochs 30 \
  --patience 8 --agg_patience 6 \
  --agg_dmodel 48 \
  --agg_fusion cross \
  --ablation none \
  --unlearn_rate 0.10 --seed 61 --profile_time

Ablation Studies

# ESC only (no GGB aggregator)
python callosumnet.py \
  --datasets PeMS08 --model_type STGCN \
  --experiment_mode subgraph_only --num_partitions 4 \
  --hidden_features 128 --epochs 60 --patience 8 \
  --ablation no_aggregation --unlearn_rate 0.00 --seed 41

# GGB only (no intra-subgraph ganglion nodes)
python callosumnet.py \
  --datasets PeMS08 --model_type STGCN \
  --experiment_mode subgraph_aggregation --num_partitions 4 \
  --hidden_features 128 --epochs 60 --agg_epochs 30 \
  --patience 8 --agg_patience 6 --agg_dmodel 48 \
  --agg_fusion cross --ablation no_intra_ganglion \
  --unlearn_rate 0.00 --seed 42

# Ganglion-path only (no Token-Transformer)
python callosumnet.py \
  --datasets PeMS08 --model_type STGCN \
  --experiment_mode subgraph_aggregation --num_partitions 4 \
  --hidden_features 128 --epochs 60 --agg_epochs 30 \
  --patience 8 --agg_patience 6 --agg_dmodel 48 \
  --agg_fusion ganglion --ablation none \
  --unlearn_rate 0.00 --seed 44

# No virtual edges
python callosumnet.py \
  --datasets PeMS08 --model_type STGCN \
  --experiment_mode subgraph_aggregation --num_partitions 4 \
  --hidden_features 128 --epochs 60 --agg_epochs 30 \
  --patience 8 --agg_patience 6 --agg_dmodel 48 \
  --agg_fusion cross --ablation partition_m_4 \
  --lambda_1 0 --lambda_2 0 --unlearn_rate 0.00 --seed 51

Key Arguments

Argument	Description	Default
--datasets	Dataset name (PeMS08, RWW, Weather, Mobility)	PeMS08
--model_type	ST-GNN backbone (STGCN, STGAT, STGATv2, STSAGE)	STGCN
--experiment_mode	full_graph, subgraph_only, or subgraph_aggregation	—
--num_partitions	Number of ESC subgraphs (M)	4
--hidden_features	Hidden dimension of the ST-GNN backbone	128
--epochs	Training epochs for subgraph models	60
--agg_epochs	Training epochs for GGB layer	30
--agg_dmodel	Transformer hidden dimension in GGB	48
--agg_fusion	Fusion mode: cross (full) or ganglion (ganglion-only)	cross
--ablation	Ablation flag (none, no_aggregation, no_intra_ganglion, etc.)	none
--unlearn_rate	Fraction of nodes to unlearn (0.0–1.0)	0.0
--lambda_1	L1 sparsity regularization on meta-graph	0.01
--lambda_2	L2 regularization on ganglion embeddings	0.001
--seed	Random seed for reproducibility	40
--profile_time	Enable wall-clock timing	False

License

This project is licensed under the MIT License. See LICENSE for details.