Aref Einizade
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Aref Einizade has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Sepideh Hajipour Sardouie
Pacific Parkinson’s Research Centre, University of British Columbia, Vancouver, BC V6E 2M6, Canada
7 shared publications - Jhony Heriberto Giraldozuluaga
LTCI, Télécom Paris, Institut Polytechnique de Paris
4 shared publications Sepideh Hajipour Sardouie
3 shared publicationsGari D. Clifford
Georgia Institute of Technology, GA, USA; Emory School of Medicine, GA, USA
2 shared publicationsMohsen Mozafari
2 shared publicationsSamaneh Nasiri
Massachusetts General Hospital, Harvard Medical School, MA, USA
2 shared publications
Latest Publications
- Open Access
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
WiGNeXt: Windowed Vision Graph Neural Network with eXtended Locality
Graph Neural Networks (GNNs) have recently been explored as an alternative to Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) by representing images as graphs and modeling region interactions through message passing. However, current vision GNNs (ViGs) either rely on costly k-Nearest Neighbors (k-NN) graph construction or sacrifice global connectivity, making them inefficient for high-resolution tasks.In this work, we introduce and generalize a family of windowed ViG models for efficient and scalable image processing. Our models construct graphs within non-overlapping windows, enabling global information propagation via different cross-window connectivity strategies. In particular, we adopt a shifted-window partitioning to enable efficient interactions between windows and a dilated-window partitioning to enhance long-range connectivity. Moreover, we propose a novel graph convolutional operator to integrate spatial information in the message-passing function. Experiments on ImageNet-1K and COCO show that WiGNeXt matches state-of-the-art ViT and ViG backbones on classification while yielding better accuracy-efficiency trade-offs for detection and segmentation, achieving 45.3 APbox and 41.3 APmask on COCO while using only 3.97 GB of GPU memory. We publicly released the code and pre-trained models at https://github.com/EIDOSLAB/WiGNet/tree/wignex.
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWiGNeXt: Windowed Vision Graph Neural Network with eXtended Locality
Graph Neural Networks (GNNs) have recently been explored as an alternative to Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) by representing images as graphs and modeling region interactions through message passing. However, current vision GNNs (ViGs) either rely on costly k-Nearest Neighbors (k-NN) graph construction or sacrifice global connectivity, making them inefficient for high-resolution tasks.In this work, we introduce and generalize a family of windowed ViG models for efficient and scalable image processing. Our models construct graphs within non-overlapping windows, enabling global information propagation via different cross-window connectivity strategies. In particular, we adopt a shifted-window partitioning to enable efficient interactions between windows and a dilated-window partitioning to enhance long-range connectivity. Moreover, we propose a novel graph convolutional operator to integrate spatial information in the message-passing function. Experiments on ImageNet-1K and COCO show that WiGNeXt matches state-of-the-art ViT and ViG backbones on classification while yielding better accuracy-efficiency trade-offs for detection and segmentation, achieving 45.3 APbox and 41.3 APmask on COCO while using only 3.97 GB of GPU memory. We publicly released the code and pre-trained models at https://github.com/EIDOSLAB/WiGNet/tree/wignex.
- Open Access
- 12 August 2026
- Published by Elsevier in SSRN Electronic Journal
No Labels, No Problem: An Unsupervised Criterion for Graph Learning Evaluation and Selection
Graph topologies learned from data often lack a reliable, label-free criterion to assess their structural fidelity. When multiple graph learning methods or hyperparameter configurations are applied to the same dataset, they can generate significantly different structures. However, in the absence of labels or a reference network, it is challenging to identify which graph accurately represents the underlying relationships among the nodes. We introduce an unsupervised criterion for evaluating and selecting graph topologies, grounded in a Laplacian smoothness prior. The core idea is that a well-aligned topology should support the reconstruction of missing node features from local neighborhoods. Given a candidate structure and a node-feature matrix, we measure reconstruction quality under progressive node removal and summarize performance using the Frobenius-normalized mean squared error (NMSE). This yields a scalar score that enables direct, label-free comparison across candidate topologies without requiring downstream tasks or ground-truth structure.We evaluate the criterion in three complementary settings: a standard graph topology inference benchmark, a controlled synthetic experiment, and two real-world Passive Acoustic Monitoring (PAM) deployments in tropical landscapes. The results show broad consistency with supervised references when available and demonstrate that the criterion can support graph topology selection when labels and reference networks are unavailable. Overall, this supports the use of smoothness-driven reconstruction as a practical tool for evaluating learned graph structures in label-scarce settings.
- 12 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessNo Labels, No Problem: An Unsupervised Criterion for Graph Learning Evaluation and Selection
Graph topologies learned from data often lack a reliable, label-free criterion to assess their structural fidelity. When multiple graph learning methods or hyperparameter configurations are applied to the same dataset, they can generate significantly different structures. However, in the absence of labels or a reference network, it is challenging to identify which graph accurately represents the underlying relationships among the nodes. We introduce an unsupervised criterion for evaluating and selecting graph topologies, grounded in a Laplacian smoothness prior. The core idea is that a well-aligned topology should support the reconstruction of missing node features from local neighborhoods. Given a candidate structure and a node-feature matrix, we measure reconstruction quality under progressive node removal and summarize performance using the Frobenius-normalized mean squared error (NMSE). This yields a scalar score that enables direct, label-free comparison across candidate topologies without requiring downstream tasks or ground-truth structure.We evaluate the criterion in three complementary settings: a standard graph topology inference benchmark, a controlled synthetic experiment, and two real-world Passive Acoustic Monitoring (PAM) deployments in tropical landscapes. The results show broad consistency with supervised references when available and demonstrate that the criterion can support graph topology selection when labels and reference networks are unavailable. Overall, this supports the use of smoothness-driven reconstruction as a practical tool for evaluating learned graph structures in label-scarce settings.
- 3 October 2025
- Published by Springer Nature
- p. 430-447
Subgraph Gaussian Embedding Contrast for Self-supervised Graph Representation Learning
- 3 October 2025
- Published by Springer Nature
- p. 430-447
Subgraph Gaussian Embedding Contrast for Self-supervised Graph Representation Learning
- Open Access
Higher-Order GNNs Meet Efficiency: Sparse Sobolev Graph Neural Networks
Graph Neural Networks (GNNs) have shown great promise in modeling relationships between nodes in a graph, but capturing higher-order relationships remains a challenge for large-scale networks. Previous studies have primarily attempted to utilize the information from higher-order neighbors in the graph, involving the incorporation of powers of the shift operator, such as the graph Laplacian or adjacency matrix. This approach comes with a trade-off in terms of increased computational and memory demands. Relying on graph spectral theory, we make a fundamental observation: the regular and the Hadamard power of the Laplacian matrix behave similarly in the spectrum. This observation has significant implications for capturing higher-order information in GNNs for various tasks such as node classification and semi-supervised learning. Consequently, we propose a novel graph convolutional operator based on the sparse Sobolev norm of graph signals. Our approach, known as Sparse Sobolev GNN (S2-GNN), employs Hadamard products between matrices to maintain the sparsity level in graph representations. S2-GNN utilizes a cascade of filters with increasing Hadamard powers to generate a diverse set of functions. We theoretically analyze the stability of S2-GNN to show the robustness of the model against possible graph perturbations. We also conduct a comprehensive evaluation of S2-GNN across various graph mining, semi-supervised node classification, and computer vision tasks. In particular use cases, our algorithm demonstrates competitive performance compared to state-of-the-art GNNs in terms of performance and running time.
Open AccessHigher-Order GNNs Meet Efficiency: Sparse Sobolev Graph Neural Networks
Graph Neural Networks (GNNs) have shown great promise in modeling relationships between nodes in a graph, but capturing higher-order relationships remains a challenge for large-scale networks. Previous studies have primarily attempted to utilize the information from higher-order neighbors in the graph, involving the incorporation of powers of the shift operator, such as the graph Laplacian or adjacency matrix. This approach comes with a trade-off in terms of increased computational and memory demands. Relying on graph spectral theory, we make a fundamental observation: the regular and the Hadamard power of the Laplacian matrix behave similarly in the spectrum. This observation has significant implications for capturing higher-order information in GNNs for various tasks such as node classification and semi-supervised learning. Consequently, we propose a novel graph convolutional operator based on the sparse Sobolev norm of graph signals. Our approach, known as Sparse Sobolev GNN (S2-GNN), employs Hadamard products between matrices to maintain the sparsity level in graph representations. S2-GNN utilizes a cascade of filters with increasing Hadamard powers to generate a diverse set of functions. We theoretically analyze the stability of S2-GNN to show the robustness of the model against possible graph perturbations. We also conduct a comprehensive evaluation of S2-GNN across various graph mining, semi-supervised node classification, and computer vision tasks. In particular use cases, our algorithm demonstrates competitive performance compared to state-of-the-art GNNs in terms of performance and running time.
- 26 January 2024
- Published by Elsevier in Digital Signal Processing
- Vol. 146
Estimation of a causal directed acyclic graph process using non-gaussianity
- 26 January 2024
- Published by Elsevier in Digital Signal Processing
- Vol. 146
Estimation of a causal directed acyclic graph process using non-gaussianity
- 24 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Signal and Information Processing over Networks
- Vol. 9, 357-372
Learning Product Graphs From Spectral Templates
Graph Learning (GL) is at the core of leveraging connections in machine learning (ML). By observing a dataset of graph signals and considering specific assumptions, Graph Signal Processing (GSP) provides practical constraints in GL. Inferring a graph with desired frequency signatures, i.e., spectral templates, from stationary graph signals has gained great attention. However, a severe computational burden is a challenging barrier, especially for inference from high-dimensional product graph signals, i.e., graph signals live on the product of smaller factor graphs. Few product GL methods have been proposed for mostly inference with smoothness assumption, while they are limited to learning only two factor graphs, handle only the Cartesian products, and have not addressed GL with desired spectral templates. To bridge the mentioned gaps, we propose a method for learning product graphs from product graph signals in which the product GL problem can be broken into separated optimizations associated with each (significantly smaller) factor graph. Besides, unlike the current approaches, our method can learn from any type of product graph (possibly with more than two factor graphs) without needing to know the type of graph products beforehand, and with significantly reduced complexity than learning directly from product graph signals. In addition to devising theoretical sufficient recovery conditions and validating them numerically, experimental results on synthetic and real-world data, i.e., multi-view image and brain signal analysis, illustrate meaningful factor graphs supported by expert-related research and also superiority over the current methods for learning from spectral templates. The implementation codes are available at https://github.com/ArefEinizade2/ProdSpecTemp .
- 24 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Signal and Information Processing over Networks
- Vol. 9, 357-372
Learning Product Graphs From Spectral Templates
Graph Learning (GL) is at the core of leveraging connections in machine learning (ML). By observing a dataset of graph signals and considering specific assumptions, Graph Signal Processing (GSP) provides practical constraints in GL. Inferring a graph with desired frequency signatures, i.e., spectral templates, from stationary graph signals has gained great attention. However, a severe computational burden is a challenging barrier, especially for inference from high-dimensional product graph signals, i.e., graph signals live on the product of smaller factor graphs. Few product GL methods have been proposed for mostly inference with smoothness assumption, while they are limited to learning only two factor graphs, handle only the Cartesian products, and have not addressed GL with desired spectral templates. To bridge the mentioned gaps, we propose a method for learning product graphs from product graph signals in which the product GL problem can be broken into separated optimizations associated with each (significantly smaller) factor graph. Besides, unlike the current approaches, our method can learn from any type of product graph (possibly with more than two factor graphs) without needing to know the type of graph products beforehand, and with significantly reduced complexity than learning directly from product graph signals. In addition to devising theoretical sufficient recovery conditions and validating them numerically, experimental results on synthetic and real-world data, i.e., multi-view image and brain signal analysis, illustrate meaningful factor graphs supported by expert-related research and also superiority over the current methods for learning from spectral templates. The implementation codes are available at https://github.com/ArefEinizade2/ProdSpecTemp .
- 15 May 2023
- Published by Elsevier in Neural Networks
- Vol. 164, 667-680
ProductGraphSleepNet: Sleep staging using product spatio-temporal graph learning with attentive temporal aggregation
- 15 May 2023
- Published by Elsevier in Neural Networks
- Vol. 164, 667-680
ProductGraphSleepNet: Sleep staging using product spatio-temporal graph learning with attentive temporal aggregation
- 13 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. 16 (1) , 296-307
Iterative Pseudo-Sparse Partial Least Square and Its Higher Order Variant: Application to Inference From High-Dimensional Biosignals
Partial Least Square (PLS) regression and its (L1 or L2 norm) regularized versions have been proposed to handle the high-dimensionality aspects of the problem at hand and select relevant features. Addressing these issues improves the generalizability of decoding the unseen data, with the severe challenge of high computational complexity. In order to avoid directly solving the L1 norm optimization problem or performing matrix inversion, this paper proposes two PLS-based algorithms, Pseudo-Sparse PLS (PS-PLS) and iterative Pseudo-Sparse Higher-Order PLS (iPS-HOPLS). In these proposed methods, we add the Pseudo-Sparsity term to reduce the L1 norm of the regression coefficient vector in a selective scheme for better importance interpretation while keeping the algorithm as simple as possible. Regarding the evaluation of the proposed methods, we investigate three critical high-dimensionality regression problems of 1) the prediction of 3D trajectory from Electrocorticography (ECoG) recordings, 2) decoding continuous fluctuation of the Electromyography (EMG) powers from recorded Magnetoencephalography (MEG) signals, and 3) continuous decoding of the finger forces from the High-Density surface Electromyogram (HD-sEMG) signals. As well as providing cognitive-relevant interpretations, the experimental results show significant improvements over the generic methods and competitive performance compared to the state-of-the-art regularized PLS approaches.
- 13 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. 16 (1) , 296-307
Iterative Pseudo-Sparse Partial Least Square and Its Higher Order Variant: Application to Inference From High-Dimensional Biosignals
Partial Least Square (PLS) regression and its (L1 or L2 norm) regularized versions have been proposed to handle the high-dimensionality aspects of the problem at hand and select relevant features. Addressing these issues improves the generalizability of decoding the unseen data, with the severe challenge of high computational complexity. In order to avoid directly solving the L1 norm optimization problem or performing matrix inversion, this paper proposes two PLS-based algorithms, Pseudo-Sparse PLS (PS-PLS) and iterative Pseudo-Sparse Higher-Order PLS (iPS-HOPLS). In these proposed methods, we add the Pseudo-Sparsity term to reduce the L1 norm of the regression coefficient vector in a selective scheme for better importance interpretation while keeping the algorithm as simple as possible. Regarding the evaluation of the proposed methods, we investigate three critical high-dimensionality regression problems of 1) the prediction of 3D trajectory from Electrocorticography (ECoG) recordings, 2) decoding continuous fluctuation of the Electromyography (EMG) powers from recorded Magnetoencephalography (MEG) signals, and 3) continuous decoding of the finger forces from the High-Density surface Electromyogram (HD-sEMG) signals. As well as providing cognitive-relevant interpretations, the experimental results show significant improvements over the generic methods and competitive performance compared to the state-of-the-art regularized PLS approaches.
- 30 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Signal and Information Processing over Networks
- Vol. 9, 35-47
Joint Graph Learning and Blind Separation of Smooth Graph Signals Using Minimization of Mutual Information and Laplacian Quadratic Forms
The smoothness of graph signals has found desirable real applications for processing irregular (graph-based) signals. When the latent sources of the mixtures provided to us as observations are smooth graph signals, it is more efficient to use graph signal smoothness terms along with the classic independence criteria in Blind Source Separation (BSS) approaches. In the case of underlying graphs being known, Graph Signal Processing (GSP) provides valuable tools; however, in many real applications, these graphs can not be well-defined a priori and need to be learned from data. In this paper, a GSP-based approach for joint Graph Learning (GL) and BSS of smooth graph signal sources is proposed, which, to the best of our knowledge, has not been addressed. The minimization of Mutual Information (MI) is exploited in our work because of its desirable advantages, such as having no local minimum and exact (not approximative) modeling of statistical independence. As well as investigating the convergence analysis and identifiability conditions and deriving Cramér-Rao Lower Bound (CRLB), comprehensive experiments of generating (cyclostationary) graph signal mixtures are provided to show the superiority over classic and also GSP-based BSS approaches in the case of the latent sources being smooth on unknown underlying graphs. Based on the provided experimental results, our proposed method performs quite robustly in even high noise levels in both terms of source separation and recovery of underlying graphs.
- 30 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Signal and Information Processing over Networks
- Vol. 9, 35-47
Joint Graph Learning and Blind Separation of Smooth Graph Signals Using Minimization of Mutual Information and Laplacian Quadratic Forms
The smoothness of graph signals has found desirable real applications for processing irregular (graph-based) signals. When the latent sources of the mixtures provided to us as observations are smooth graph signals, it is more efficient to use graph signal smoothness terms along with the classic independence criteria in Blind Source Separation (BSS) approaches. In the case of underlying graphs being known, Graph Signal Processing (GSP) provides valuable tools; however, in many real applications, these graphs can not be well-defined a priori and need to be learned from data. In this paper, a GSP-based approach for joint Graph Learning (GL) and BSS of smooth graph signal sources is proposed, which, to the best of our knowledge, has not been addressed. The minimization of Mutual Information (MI) is exploited in our work because of its desirable advantages, such as having no local minimum and exact (not approximative) modeling of statistical independence. As well as investigating the convergence analysis and identifiability conditions and deriving Cramér-Rao Lower Bound (CRLB), comprehensive experiments of generating (cyclostationary) graph signal mixtures are provided to show the superiority over classic and also GSP-based BSS approaches in the case of the latent sources being smooth on unknown underlying graphs. Based on the provided experimental results, our proposed method performs quite robustly in even high noise levels in both terms of source separation and recovery of underlying graphs.
- 25 October 2022
- Published by Elsevier in Digital Signal Processing
- Vol. 132
Robust blind separation of smooth graph signals using minimization of graph regularized mutual information
- 25 October 2022
- Published by Elsevier in Digital Signal Processing
- Vol. 132
Robust blind separation of smooth graph signals using minimization of graph regularized mutual information