Showing 12 papers for 2026-04-16
This work tackles learning high-quality representations of medical concepts in EHR data by enriching text-attributed knowledge graphs with large language models. It addresses the problem that cross-type dependencies—such as diagnosis–medication and medication–procedure—are often missing in clinical ontologies, hindering downstream predictions, and leverages textual attributes to enhance concept representations. The goal is richer, more robust medical concept embeddings for clinical prediction tasks.
Proposes ID and graph view contrastive learning with multi-view attention fusion for sequential recommendation. It combines ID-based item representations with graph-based representations to capture both item identity and relational structure from user interaction histories. A multi-view attention fusion module integrates these views within a contrastive learning framework to address limitations of prior methods.
We present a graph-embedding-based anomaly detection system for microservice architectures that uses unsupervised node-level embeddings learned by a GCN-GAE on a directed service-call graph to identify anomalous or under-represented services during live event traffic. The approach captures structural patterns that may be missed by traditional load testing. It aims to improve detection during real streaming workloads.
CCCE introduces a continuous code calibration engine for autonomous enterprise codebase maintenance across hundreds of repositories, languages, and packages. It uses knowledge graph traversal to propagate maintenance signals across interconnected code entities and adopts adaptive decision gating to control propagation, enabling scalable, autonomous maintenance.
The Code Whisperer integrates large language models with graph-based program analysis (ASTs, CFGs, PDGs) to detect, explain, and repair maintainability issues and vulnerabilities within a unified workflow. By aligning structural graphs with token embeddings, it provides contextual explanations and actionable fixes.
Graph Propagated Projection Unlearning (GPPU) presents a unified framework for class-level unlearning that works across vision and audio models. It uses graph-based propagation to identify class-specific directions in feature space and projects representations onto orthogonal subspaces, followed by targeted fine-tuning to remove information about forgotten classes.
This study investigates young people's perceptions of conversational generative AI in youth mental health after a co-designed Mia chatbot. Thirty-two participants explored views on genAI chatbots in services and produced recommendations for reframing Mia for consumer use and integration into services.
MCPThreatHive is an open-source platform that automates the end-to-end lifecycle of MCP threat intelligence. It performs continuous multi-source data collection, AI-driven threat extraction and classification, stores structured knowledge in a graph, and provides interactive visualization, operationalizing the MCP-38 threat taxonomy.
Leveraging LLM-GNN integration for open-world question answering over knowledge graphs, the work addresses answering questions on incomplete or evolving graphs. LLMs contribute language understanding and reasoning, while GNNs model graph topology, and the integration aims to enable robust open-world QA beyond closed-world assumptions.
GraphScout empowers LLMs with intrinsic exploration for agentic graph reasoning by enabling autonomous interaction with graphs during retrieval and reasoning. It supports iterative LLM–graph collaboration (GraphRAG) with reduced manual guidance, improving factual grounding and agentic decision-making.
This work proposes a large language model-based framework for local life service recommendation that jointly predicts a user's immediate living need and recommends corresponding services. The agentic reasoning component enables end-to-end coordination between need prediction and service recommendation for more accurate, timely suggestions.
Debate to Align introduces a two-stage multi-agent debate framework to improve entity alignment across knowledge graphs. It uses LLM-based embeddings to identify uncertain entities and constructs a candidate entity set, which is then argued by agents to reach more reliable alignment decisions.