MLIM

This is our C++ & Python implementation for the paper:

Xueqin Chang, Ruize Liu, Qing Liu. Theoretical and Learning-based Approaches for Influence Maximization in Multilayer Social Networks. 2025.

Environment Requirements

STARIM

For STARIM, we implement the algorithm in C++, and the detailed environment requirements are as follows. The project is developed and tested under a Linux environment. We recommend the following configuration for stable compilation and execution:

Component	Recommended Version	Description
Operating System	Ubuntu 20.04 / 22.04 (or other modern Linux distributions)	Verified under Ubuntu
Compiler	g++ ≥ 7.5.0	Must support the C++14 standard
Build Tool	make (optional)	If you prefer to use Makefile for compilation
Memory	≥ 16 GB RAM	Large-scale graph processing may require more memory
Dependencies	None	Only requires a standard C++14 environment

LGQIM

For LGQIM, we implement it using Python. The specific environment requirements are as follows. We strongly recommend using Anaconda to create and manage the Python environment required for running this project.

You can create a new environment as follows:

conda create -n LGQIM python=3.9
conda activate LGQIM

Recommended Versions

Component	Recommended Version	Description
Python	≥ 3.9 (tested with 3.9)	Python 3.9 or higher is recommended
PyTorch	2.5.1	Tested with torch==2.5.1
CUDA (for GPU use)	12.1	Verified to work with torch==2.5.1
cuDNN	7.4.2	Verified to work with CUDA 12.1

Required Python Packages

You can install all dependencies with pip or conda. Below are the essential and auxiliary libraries:

Essential libraries torch, networkx, scikit-learn, argparse, subprocess

Auxiliary libraries time, random, sys, os, re

⚙️ CPU and GPU Support This code can run on both CPU and GPU:

CPU: Simply install the CPU version of PyTorch using: pip install torch==2.5.1

GPU: Install the GPU-enabled version of PyTorch (recommended CUDA 12.1 and cuDNN 7.4.2). Both versions have been verified to work correctly in our experiments.

Dataset

Dataset Statistics

The following datasets are used in our experiments. Each dataset is represented as a multilayer or multiplex network, characterized by the number of nodes, number of edges, and number of layers.

Dataset	Abbreviation	#Nodes	#Edges	#Layers
Venetie	VT	1,380	19,941	43
FF-TW-YT	FTY	11,863	87,396	3
Citeseer	CS	15,533	122,903	3
DBLP	DBLP	41,892	661,883	2
Twitter	TW	47,807	657,456	3
MoscowAthletics2013	MA	133,364	309,952	3
Frienfeed	FF	1,531,015	20,204,535	3
ObamaInIsrael2013	OB	3,452,114	6,711,448	3
SF	SF	4,975,888	63,497,050	9

Due to the size limitation of the datasets, only three datasets have been placed in this repository as examples. They are: Citeseer, DBLP, and ff-tw-yt, located under the dataset folder of STARIM.

Dataset Format

Each dataset is organized as follows:

1. Layer Files

For each layer, three files are provided: layer{x}.txt, layer{x}model.txt, and layer{x}ov.txt

These files respectively store:

• the intra-layer edges,
• the propagation model of the layer, and
• the inter-layer (overlapping) edges.

Here, {x} is replaced by the corresponding layer index. For example, the Citeseer dataset contains three layers, and the related files are:

layer1.txt, layer1model.txt, layer1ov.txt  
layer2.txt, layer2model.txt, layer2ov.txt  
layer3.txt, layer3model.txt, layer3ov.txt

---

File: layer{x}.txt

• Format:

  • The first line contains two integers:

    <number_of_nodes> <number_of_edges>
    

  • Starting from the second line, each line represents one directed edge:

    src dst weight
    

  • weight ∈ [0, 1] denotes the influence strength of the edge from src to dst on the x-th layer.

---

File: layer{x}model.txt

• Stores the propagation model type for the x-th layer.

• Encoding:

  • 0 → Independent Cascade (IC) model
  • 1 → Linear Threshold (LT) model

---

File: layer{x}ov.txt

• Records the cross-layer (overlap) edges.

• Format:

  node_id overlap_layer_id overlap_node_id weight
  

  This represents an inter-layer edge from (overlap_layer_id, overlap_node_id) → (x, node_id) with weight weight. In other words, a node overlap_node_id in layer overlap_layer_id influences node node_id in layer x.

---

2. max_nodeID.txt

Stores the maximum node ID in the dataset.

---

3. node_score.txt

Stores the initial influence score of each node, which is used for training the learning-based methods.

---

4. total_layers.txt

Stores the total number of layers in the dataset.

Reproducibility & Run

Running STARIM

To run STARIM, navigate to the STARIM directory and compile the code using the following command:

g++ -std=c++14 -g -O3 ./MGRR.cpp ./multiplex.cpp -o executableFile/STARIM

---

Example of Execution

As an example, you can run STARIM on the Citeseer dataset with the following command:

./executableFile/STARIM -mode=M -dir=dataset/Citeseer -seedsize=20 -delta=0.01 -epsilon=0.2

---

Additional Implementations

In this project, we also provide implementations of several baseline methods and utility tools used in the experiments. Their compilation and execution processes are similar — simply replace ./MGRR.cpp with the corresponding .cpp source file, and then compile and run it in the same way.

Running LGQIM

To run LGQIM, please navigate to the LGQIM directory. LGQIM consists of two main components: GNN training and QNet training.

---

1. GNN Training

You can train the GNN model using the following command:

python GNN.py --dataset "$dataset" --num_epochs "$r"

where:

• "$dataset" should be replaced with the path to the dataset,
• "$r" should be replaced with the number of training epochs for the GNN model.

---

2. QNet Training

You can train the QNet model using:

python QNet_model/QNet_train.py --qnet_dataset "$dataset" --qnet_k "$k" --qnet_gamma "$gm"

where:

• "$l_dataset" — path to the dataset used for QNet training.

• "$k" — number of seed nodes to be selected.

• "$gm" — parameter gamma (γ) in QNet, representing the discount factor for Q-learning.

  • Optional values: [0, 0.2, 0.4, 0.6, 0.8, 1.0]
  • Larger gamma implies higher influence.
  • The typical setting is 0.8.