FoilDiff: A Hybrid Transformer Backbone for Diffusion-based Modelling of 2D Airfoil Flow Fields

InvFusion is a novel approach that integrates supervised and zero-shot diffusion methods for solving inverse problems. It addresses the limitations of existing models by providing a degradation-aware posterior sampler that enhances accuracy while maintaining flexibility. This innovation is significant as it combines the strengths of both training-based and zero-shot techniques, marking a step forward in the application of diffusion models in various fields.

ReassembleNet is a novel approach to the challenge of reassembly in various fields such as archaeology and genomics. It addresses key limitations in existing Deep Learning methods, including scalability and real-world applicability. By representing input pieces as contour keypoints and utilizing Graph Neural Networks, ReassembleNet reduces computational complexity while integrating features from multiple modalities, enhancing its effectiveness in reconstructing complex structures.

Certified Signed Graph Unlearning (CSGU) addresses the challenges of privacy in Signed Graph Neural Networks (SGNNs). Traditional graph unlearning methods fail to preserve the unique properties of signed graphs, leading to loss of critical sign information. CSGU introduces a three-stage approach to effectively remove specific data influences while maintaining privacy guarantees and the sociological principles of SGNNs. This advancement is crucial for applications where data sensitivity is paramount.

The paper introduces DualLaguerreNet, a new architecture for Graph Neural Networks (GNNs) that addresses the limitations of single-filter models like LaguerreNet. By decoupling the graph Laplacian into low and high-frequency operators, it allows for independent learning of two adaptive Laguerre polynomial filters. This innovation aims to overcome the compromise problem of averaged responses across the graph spectrum, enhancing performance on complex heterophilic graphs.

The paper titled 'Fairness-Aware Graph Representation Learning with Limited Demographic Information' addresses the challenge of ensuring fairness in Graph Neural Networks (GNNs). It highlights that many existing methods require complete demographic data, which is often unavailable due to privacy concerns. The authors propose a new framework that generates proxies for demographic information using partial data and enforces consistent node embeddings across demographic groups, along with an adaptive confidence strategy to balance fairness and utility.

Graph Neural Networks (GNNs) have shown significant success in various applications, but they face challenges such as oversmoothing and inefficiency on heterophilic graphs. To overcome these limitations, a new framework is introduced that incorporates complex weights into graph structures, allowing for a complex diffusion process. This leads to the development of the Complex-Weighted Convolutional Network (CWCN), which learns complex-weighted structures from data and enhances diffusion with learnable matrices and nonlinear activations, proving to be effective for node-classification tasks.

Circuit link prediction is essential for automating analog circuit design by identifying missing connections in incomplete netlists. Current methods struggle with three key issues: inadequate use of topological patterns, data scarcity due to complex annotations, and limited adaptability to various netlist formats. The proposed Graph Neural Networks Based Analog Circuit Link Prediction (GNN-ACLP) method introduces innovations such as the SEAL framework for improved accuracy and the Netlist Babel Fish tool for format conversion, addressing these challenges effectively.

The article introduces Convolutional Nearest Neighbors (ConvNN), a framework that unifies Convolutional Neural Networks (CNNs) and Transformers by viewing convolution and self-attention as neighbor selection and aggregation methods. ConvNN allows for a systematic exploration of the spectrum between these two architectures, serving as a drop-in replacement for convolutional and attention layers. The framework's effectiveness is validated through classification tasks on CIFAR-10 and CIFAR-100 datasets.