[Paper Review] Contrastive Brain Network Learning via Hierarchical Signed Graph Pooling Model

Haoteng Tang, Guixiang Ma+4 more|arXiv (Cornell University)|Jul 14, 2022

Functional Brain Connectivity Studies4 citations

TL;DR

This paper proposes a novel hierarchical signed graph pooling model for brain functional network representation learning, integrating contrastive learning with data augmentation to improve performance on clinical prediction tasks. The method achieves state-of-the-art results in classification and regression on HCP and OASIS datasets while enabling interpretable biomarker discovery via graph saliency maps.

ABSTRACT

Recently brain networks have been widely adopted to study brain dynamics, brain development and brain diseases. Graph representation learning techniques on brain functional networks can facilitate the discovery of novel biomarkers for clinical phenotypes and neurodegenerative diseases. However, current graph learning techniques have several issues on brain network mining. Firstly, most current graph learning models are designed for unsigned graph, which hinders the analysis of many signed network data (e.g., brain functional networks). Meanwhile, the insufficiency of brain network data limits the model performance on clinical phenotypes predictions. Moreover, few of current graph learning model is interpretable, which may not be capable to provide biological insights for model outcomes. Here, we propose an interpretable hierarchical signed graph representation learning model to extract graph-level representations from brain functional networks, which can be used for different prediction tasks. In order to further improve the model performance, we also propose a new strategy to augment functional brain network data for contrastive learning. We evaluate this framework on different classification and regression tasks using the data from HCP and OASIS. Our results from extensive experiments demonstrate the superiority of the proposed model compared to several state-of-the-art techniques. Additionally, we use graph saliency maps, derived from these prediction tasks, to demonstrate detection and interpretation of phenotypic biomarkers.

Motivation & Objective

• Address the limitation of existing graph neural networks in handling signed brain functional networks, which are common in fMRI data.
• Overcome the scarcity of brain network data by proposing a novel data augmentation strategy for contrastive learning.
• Enhance model interpretability to enable biological insight into predicted phenotypes through saliency mapping.
• Develop a hierarchical graph pooling mechanism that captures multi-scale brain network structures for better representation learning.

Proposed method

• Propose a hierarchical signed graph pooling framework that learns low-dimensional representations of brain functional networks by aggregating node features across multiple levels of graph hierarchy.
• Incorporate signed graph convolutional networks based on balance theory to model positive and negative functional connections between brain regions.
• Design a contrastive learning objective using augmented brain network samples to improve generalization and representation quality.
• Apply data augmentation via edge perturbation and node feature masking to generate positive pairs for contrastive pre-training.
• Use graph saliency maps derived from attention mechanisms to highlight key brain regions associated with clinical phenotypes.
• Integrate global readout with hierarchical pooling to produce graph-level representations suitable for downstream classification and regression tasks.

Figure 1: Diagram of the proposed contrastive graph learning framework (in the bottom black box) with hierarchical signed graph representation learning model (in the top black box) for functional brain network embedding and downstream tasks (i.e., phenotype classification or regression).

Experimental results

Research questions

• RQ1 Can a hierarchical signed graph pooling model outperform existing unsigned graph learning methods on brain network representation learning for clinical prediction?
• RQ2 How effective is the proposed contrastive learning strategy with data augmentation in improving model performance under limited brain network data?
• RQ3 To what extent can the model’s predictions be interpreted through saliency maps to identify biologically relevant brain regions for specific phenotypes?
• RQ4 Does the integration of balance theory in signed graph learning enhance the modeling of functional brain connectivity patterns?

Key findings

• The proposed model achieves state-of-the-art performance on both classification and regression tasks using HCP and OASIS datasets, outperforming existing SOTA methods.
• Contrastive learning with data augmentation significantly improves model generalization, especially in low-data regimes common in neuroimaging.
• Graph saliency maps successfully identify biologically plausible brain regions associated with clinical phenotypes such as MMSE, Flanker, and aggressive behavior.
• The model detects known neuroanatomical regions like the precuneus, insula, and cingulate cortex as key contributors to cognitive and behavioral traits.
• Hierarchical pooling enables better capture of multi-scale brain network organization compared to flat GNNs, improving representation quality.
• The use of signed graph learning based on balance theory leads to more interpretable and biologically consistent node relationships in functional networks.

Figure 2: Visualization of the averaged adjacency matrices for original and contrastive samples. The averaged contrastive sample pair is generated by using a window size $d=10$ .