COCO: A Cohesiveness-aware Learning Framework for Community Search over Temporal Graphs

This repository contains the official implementation for the paper: COCO: A Cohesiveness-aware Learning Framework for Community Search over Temporal Graphs.

Overview

COCO is a cohesiveness-aware learning framework designed for community search over temporal graphs. It seamlessly integrates the classical 𝑘-core decomposition with Graph Neural Networks (GNNs). The framework addresses the temporal interval community search problem through a three-stage process: pre-training, query-driven fine-tuning, and community search.

Workflow and Usage

The end-to-end workflow consists of data preprocessing, model pre-training, and query execution.

Step 1: Data Preparation

1. Download Datasets: Obtain raw temporal graph data from public repositories such as SNAP and KONECT and place them in the datasets/ directory.

2. Process Graph Data: Execute the script to convert the raw graph into the required format.

   python process_graph.py

   • Output: datasets/<dataset_name>.txt

3. Perform Core Decomposition: Run the decomposition script to generate the core number file.

   python decomposition.py 

   • Output: datasets/<dataset_name>-core_number.txt

Step 2: Model Pre-training

1. Train HM-Index (MLP Models): Train the MLP models that constitute the Cohesiveness Prediction Index.

   python MLP.py 

   • Output: Trained models saved in models/<dataset_name>/

2. Pre-train the Main GNN Model: Run the main training script to pre-train the primary graph model.

   python main.py 

   • Output: Pre-trained model weights saved in the ./ directory.

Step 3: Model Fine-tuning and Community Search

Perform query-driven fine-tuning on the pre-trained model, and then use the fine-tuned model for search.

python single_query.py 

Repository Structure

Core Logic Files

• data_loader.py: Reads datasets and constructs Graph objects.
• extract_subgraph.py: Handles subgraph sampling and training tuple generation.
• loss.py: Defines loss functions for model training.
• model.py: Defines the AT-GNN and Adapter models.
• MLP_models.py: Defines the series of MLP models for the HM-Index.
• train.py: Implements model training and validation routines.
• utils.py: Contains various utility functions.

Executable Scripts

• process_graph.py: Preprocesses raw graph data.
• decomposition.py: Performs k-core decomposition on a graph.
• MLP.py: Trains the HM-Index models.
• main.py: Pre-trains the main GNN model.
• single_query.py: Executes a fine-tuning and community search task for a given query.
• MLP_search.py: Utility script for analyzing HM-Index prediction targets.