Showing 9 papers for 2026-07-11
This paper proposes representing surface EMG signals as muscle-activation graphs and training a graph neural network for real-time hand gesture recognition. By encoding forearm muscle activity as graphs, the method captures spatial-temporal activation patterns that improve discrimination among gestures. Experiments demonstrate real-time capability suitable for controlling advanced prostheses and augmented reality interfaces.
PolyUQuest introduces a verifiable, structure-aware web retrieval-augmented generation framework built on a heterogeneous graph that integrates hyperlink topology, page DOM structure, and cross-page entity relations. A two-tier router dispatches each query to one of three retrieval modes—direct block retrieval, cross-page graph traversal, and multi-hop entity retrieval—to match structural needs and provide verifiable, context-rich answers. It also ensures verifiability and provenance for generated content.
This work tackles EEG-based emotion recognition by introducing a graph-regularized learning framework where emotions are nodes in a graph connected by proximities from dimensional theories. Three complementary regularization strategies, including Graph Label Smoothing and commuting-distance penalties on the graph, enforce psychological relationships and improve robustness to label noise.
path_boost is a Python package that implements PathBoost, a gradient-boosting method for interpretable graph-level prediction based on labeled path features. During training it automatically discovers predictive paths, yielding an additive model where each path contributes transparently to the prediction, and it avoids exhaustively enumerating all possible paths.
This paper explores the internal kNN graph constructed by UMAP and shows that it encodes valuable manifold structure prior to projection. Applying standard graph algorithms (e.g., PageRank) on this graph enhances sensemaking by identifying representative points and revealing connectivity patterns.
MetaHGNIE introduces meta-path induced hypergraph contrastive learning for heterogeneous knowledge graphs to capture higher-order interactions beyond pairwise messages. By disentangling topology and semantic attributes and leveraging meta-relational structures, it yields more discriminative node importance representations for tasks like recommendation and search.
This work uses knowledge graphs to construct accurate portraits of scientific and technological academic conferences, organizing entities such as papers, researchers, institutions, projects, and topics. It aims to improve efficient information discovery and decision making by modeling conference ecosystems and their relationships.
This paper proposes hippocampus-inspired multi-view hypergraph learning for web finance fraud detection, bridging cognitive neuroscience and graph intelligence. It addresses long-tailed data distributions and rare fraudulent patterns by fusing signals across multiple views through hypergraphs, enhancing detection performance.
KG-SoftMAP encodes domain knowledge as soft, weighted edge priors on a Bayesian network and optimizes a MAP objective by combining this prior with the BDeu score. In sparse discrete data settings, data-only methods struggle to recover structure, and the KG-SoftMAP approach improves BN structure learning by leveraging informative priors.