Abstract:Graph Transformers have recently been successful in various graph representation learning tasks, providing a number of advantages over message-passing Graph Neural Networks. Utilizing Graph Transformers for learning the representation of the brain functional connectivity network is also gaining interest. However, studies to date have underlooked the temporal dynamics of functional connectivity, which fluctuates over time. Here, we propose a method for learning the representation of dynamic functional connectivity with Graph Transformers. Specifically, we define the connectome embedding, which holds the position, structure, and time information of the functional connectivity graph, and use Transformers to learn its representation across time. We perform experiments with over 50,000 resting-state fMRI samples obtained from three datasets, which is the largest number of fMRI data used in studies by far. The experimental results show that our proposed method outperforms other competitive baselines in gender classification and age regression tasks based on the functional connectivity extracted from the fMRI data.

What problem does this paper attempt to address?

### Problems Addressed by the Paper This paper aims to address the issue of representation learning for Dynamic Functional Connectivity (DFC) in brain networks. Specifically, while existing methods have achieved certain results in Static Functional Connectivity (SFC), they often overlook the dynamic characteristics of functional connectivity that change over time. These dynamic characteristics are crucial for understanding brain functions. ### Main Contributions 1. **Definition of Connectome Embedding**: - A new connectome embedding \(X_t\) is proposed, which can include functional connectivity information at time \(t\), including location, structure, and temporal information. - Sliding window correlation and GRU time encoding methods are used to extract this information. 2. **Proposing the TENET Model**: - A Transformer-based model, TENET, is designed. This model learns the representation of dynamic functional connectivity through Multi-Head Self-Attention (MHSA). - The model is divided into two parts: the spatial attention module (Connectome Transformer) and the temporal attention module (Transformer Encoder), which handle spatial and temporal information respectively. 3. **Large-Scale Experimental Validation**: - Experiments are conducted using over 50,000 resting-state fMRI samples from three datasets, making it one of the largest fMRI datasets to date. - Experimental results show that TENET outperforms other baseline methods in gender classification and age regression tasks, demonstrating its effectiveness in capturing dynamic functional connectivity. ### Key Techniques - **Connectome Embedding**: Combines location, structure, and temporal information to form a comprehensive embedding vector. - **Multi-Head Self-Attention Mechanism**: Learns the relationships between nodes through self-attention, enhancing the model's ability to capture dynamic characteristics. - **Large-Scale Dataset**: Uses a large-scale fMRI dataset for training and validation, improving the model's generalization ability and reproducibility. ### Conclusion By proposing the TENET model, this paper successfully addresses the issue of representation learning for dynamic functional connectivity in brain networks. Experimental results indicate that TENET performs excellently in gender classification and age regression tasks, proving its advantages in capturing and utilizing dynamic functional connectivity. Future research will further explore the applications of TENET in attention interpretation, external validation performance, and theoretical understanding.

Large-scale Graph Representation Learning of Dynamic Brain Connectome with Transformers

Graph Transformers for Large Graphs

Brain Network Transformer

Dynamic Graph Transformer for Brain Disorder Diagnosis

Transformer3: A Pure Transformer Framework for fMRI-Based Representations of Human Brain Function

Deep Representations for Time-varying Brain Datasets

Large Transformers are Better EEG Learners

DyFormer: A Scalable Dynamic Graph Transformer with Provable Benefits on Generalization Ability

Dynamic Adaptive Spatio-temporal Graph Convolution for fMRI Modelling

BrainTGL: A dynamic graph representation learning model for brain network analysis

Disentangling Spatial-Temporal Functional Brain Networks via Twin-Transformers

DTFormer: A Transformer-Based Method for Discrete-Time Dynamic Graph Representation Learning

Contrasformer: A Brain Network Contrastive Transformer for Neurodegenerative Condition Identification

EEG-Deformer: A Dense Convolutional Transformer for Brain-computer Interfaces

Community-Aware Transformer for Autism Prediction in fMRI Connectome

Supra-Laplacian Encoding for Transformer on Dynamic Graphs

Deformable Graph Transformer

Beyond the Snapshot: Brain Tokenized Graph Transformer for Longitudinal Brain Functional Connectome Embedding

Self-Supervised Graph Representation Learning for Neuronal Morphologies

TransformerG2G: Adaptive time-stepping for learning temporal graph embeddings using transformers

Graph Transformers Dream of Electric Flow