Explicit Multimodal Graph Modeling for Human-Object Interaction Detection

The article discusses a novel approach to polymer property prediction using a multi-view design that incorporates various representations. The system combines four families of representations: tabular RDKit/Morgan descriptors, graph neural networks, 3D-informed representations, and pretrained SMILES language models. This ensemble method achieved a public mean absolute error (MAE) of 0.057 and a private MAE of 0.082, ranking 9th out of 2241 teams in the Open Polymer Prediction Challenge at NeurIPS 2025.

Graph neural networks (GNNs) are essential for analyzing relational data, categorized into convolutional, attentional, and message-passing variants. The standard message-passing approach, while expressive, overlooks the rich contextual information from the broader local neighborhood, limiting its ability to learn complex relationships. This article introduces a new framework called neighborhood-contextualized message-passing (NCMP) to address this limitation, enhancing the expressivity and efficiency of GNNs.

Graph Neural Networks (GNNs) are crucial for analyzing graph-structured data, but current models overlook the conservation laws relevant to physical resource flows, such as electrical currents in power grids. To improve performance, a new approach called flow attention is introduced, which aligns with Kirchhoff's first law. Experiments on electronic circuits and power grids demonstrate that this method enhances the effectiveness of attention-based GNNs in classification and regression tasks.

Recent advancements in graph neural networks (GNNs) have highlighted their effectiveness in node classification tasks. However, traditional GNNs often neglect role-based characteristics that can enhance node representation learning. To overcome these limitations, a new model called the hypergraph neural network with state space model (HGMN) has been proposed, integrating role-aware representations and employing hypergraph construction techniques to capture complex relationships among nodes.

The article presents a novel approach to incomplete multi-view clustering (IMVC) through Dynamic Deep Graph Learning with Masked Graph Reconstruction Loss. It highlights the limitations of existing methods, particularly their reliance on K-Nearest Neighbors (KNN) and Mean Squared Error (MSE) loss, which can introduce noise and reduce graph robustness. The proposed method aims to enhance the effectiveness of IMVC by addressing these challenges, thereby contributing to the advancement of Graph Neural Networks (GNNs) in this field.

The article presents the Neighborhood-Aware Star Kernel (NASK), a new graph kernel for attributed graph learning. Attributed graphs, which feature irregular topologies and a mix of numerical and categorical attributes, are prevalent in areas like social networks and bioinformatics. NASK utilizes an exponential transformation of the Gower similarity coefficient to efficiently model these attributes and incorporates multi-scale neighborhood structural information through star substructures enhanced by Weisfeiler-Lehman iterations. The theoretical proof confirms that NASK is positive definite.

Graph neural networks (GNNs) are increasingly utilized in urban spatiotemporal forecasting, particularly for predicting infrastructure issues like potholes and rodent problems. Government inspection ratings provide insights into the state of incidents in various neighborhoods, but these ratings are limited to a sparse selection of areas. To enhance prediction accuracy, a new multiview, multioutput GNN model integrates both government ratings and crowdsourced reports, addressing biases in reporting behavior and improving the understanding of urban incidents.

The paper titled 'Zero-Shot Temporal Interaction Localization for Egocentric Videos' presents a novel approach called EgoLoc, aimed at improving the localization of human-object interactions in egocentric videos. Traditional methods rely heavily on annotated action and object categories, leading to domain bias and inefficiencies. EgoLoc introduces a self-adaptive sampling strategy to enhance visual prompts for vision-language model reasoning, ultimately achieving better temporal interaction localization.