MSGCN: Multiplex Spatial Graph Convolution Network for Interlayer Link Weight Prediction

This repository contains the code for the Multiplex Spatial Graph Convolution Network (MSGCN), a multilayer link weight prediction method for spatial multiplex networks. The code also includes synthetic data generation to test the MSGCN method with different multiplex network types and sizes. The supported network types are complete, random, and small-world.

Acknowledgement

Please cite the following paper if you use this code.

Steven E. Wilson and Sina Khanmohammadi. "MSGCN: Multiplex Spatial Graph Convolutional Network for Interlayer Link Weight Prediction." arXiv preprint arXiv:2504.17749 (2025).

https://arxiv.org/abs/2504.17749.

Authors

1. Steven Ewan Wilson:

2. Sina Khanmohammadi:

Instructions

Installation

1. Prerequisites: Make sure you have the following installed:
   • Python (3.6 or later)
   • Required libraries: argparse, numpy, pandas, pickle, logging, matplotlib, seaborn, scipy, sklearn, networkx, torch, torch_geometric
   • Required folders: Be sure the following folders are present to save files that are created during experimental runs: data, plots, results, trials

MSGCN Experiments

Usage

To run all experiments use the main.py script with default arguments as shown below:

python main.py

Any of the default arguments can also be overridden as shown in the following example command:

python main.py -num_trials 5 -data_path "./datasets" -epochs 100 -lr 0.001

Command-Line Arguments

The table below describes the available command-line arguments that can be used:

Argument	Type	Default	Description
-num_trials	int	1	Number of trials to repeat for the same set of parameters
-data_path	str	"data"	Path to the dataset
-epochs	int	40	Number of training epochs
-lr	float	0.0005	Learning rate
-dim_coor	int	3	Number of spatial coordinates
-out_dim	int	1	Number of output classes for classification or output nodes for regression
-label_dim	int	1	Number of input node features to SGCN layer
-layers_num	int	1	Number of SGCN layers
-model_dim	int	16	Number of output dimensions from each SGCN layer
-out_channels_1	int	32	SGCN hidden layer size (inner convolutions)
-dropout	float	0.0	Dropout percentage
-use_cluster_pooling	bool	False	Flag to use cluster pooling
-mse_option	float	1.0	Custom loss MSE percentage
-spread_option	float	0.0	Custom loss spread percentage
-range_option	float	0.0	Custom loss range percentage
-graph_types	str	["complete"]	Allowed network types: ["complete", "random", "smallworld"]
-num_graphs	int	0	Number of synthetic networks to generate
-num_nodes	int (list)	[4]	List of node counts for the graphs
-num_layers	int (list)	[2]	List of layer counts for the graphs
-num_neighbors	int (list)	[2]	List of k-nearest neighbors (small-world)
-link_probability	float (list)	[0.3]	List of edge addition probabilities for small-world networks

Plotting Results

To visualize all experimental results use the plot.py script with default arguments as shown below:

python plot.py

Command-Line Arguments

The table below describes the available command-line arguments for plot.py:

Argument	Type	Default	Description
-trials	int	1	Number of trials to plot
-num_nodes	int (list)	[4]	List of node counts for the graphs
-num_layers	int (list)	[2]	List of layer counts for the graphs
-graph_types	str (list)	["complete"]	Allowed network types: ["complete", "random", "smallworld"]
-data_path	str	"trials/"	Path to the dataset

References

[1] Danel, T. et al. (2020). "Spatial Graph Convolutional Networks." In: Yang, H., Pasupa, K., Leung, A.CS., Kwok, J.T., Chan, J.H., King, I. (eds) Neural Information Processing. ICONIP 2020. Communications in Computer and Information Science, vol 1333. Springer, Cham. https://doi.org/10.1007/978-3-030-63823-8_76