graph neural networks

Neuro-Symbolic Reasoning with Cognitive Graph Structures

This paper introduces a novel approach to neuro-symbolic reasoning utilizing cognitive graph structures. The core idea is to integrate the strengths of neural networks – their ability to learn complex patterns from data – with symbolic knowledge representations – providing explicit, interpretable reasoning. We propose a system where a neural network learns to predict relationships within a dynamically evolving cognitive graph. This dynamic structure allows the system to adapt and refine its knowledge base based on observed data, leading to more flexible and explainable reasoning compared to traditional static neural-symbolic methods. The key innovation lies in the graph's ability to self-organize and update its structure, guided by the neural network's predictions, creating a truly interactive and adaptive reasoning framework. This approach addresses limitations of existing methods by providing a mechanism for continuous learning and refinement of symbolic knowledge, ultimately leading to improved performance on complex reasoning tasks.

Dynamic Time Warping and Graph Neural Networks Combination

This paper proposes a novel approach to analyzing time series data by combining Dynamic Time Warping (DTW) with Graph Neural Networks (GNNs). The core idea is to leverage DTW for effective feature extraction from time series, followed by a GNN to learn intricate relationships within the extracted representations. Traditional GNNs often struggle with the inherent temporal dynamics of time series data. DTW addresses this by providing a flexible and robust method for aligning and representing time series with varying speeds. The resulting graph structure, informed by DTW, is then processed by the GNN to capture complex dependencies. We demonstrate the effectiveness of this combination through a theoretical framework and outline the key components of the proposed methodology. The aim is to provide a foundation for future research in applying dynamic methods to graph-based learning.

Dynamic Topology-Dependent Graph Reasoning

This paper proposes a novel approach to program optimization that leverages dynamic graph reasoning based on topology. Traditional static dependency graphs fail to adapt to runtime variations and resource constraints, leading to suboptimal performance. We introduce a system that utilizes a neural network to learn and represent program dependencies, and dynamically adjusts the topology of the dependency graph based on observed execution behavior. This is achieved through a reinforcement learning framework where the neural network learns to modify dependency weights, effectively creating a "living" graph that reflects the true runtime dependencies. The core claim is that this dynamic adaptation significantly improves compile-time resource allocation and runtime execution path selection. The key mechanism is the reinforcement learning trained neural network, which provides a continuous feedback loop for optimization. We present a detailed description of the system architecture and the reinforcement learning algorithm, emphasizing the adaptive nature of the dependency graph. Experimental results, although not presented here due to the focus on the theoretical framework, would demonstrate the potential benefits of this approach.

Neural Network Information Flow Topology Representation

This paper investigates the representation of information flow within neural networks, proposing that the transmission of information is not a random process but rather governed by a specific topological structure. The core argument centers on the idea that the network's topology dictates the pathways through which information propagates and, consequently, influences its processing capabilities. We introduce a methodology for constructing a neural network information flow topology graph based on analyzing the strength, delay, and direction of connections between neurons. Utilizing topological analysis techniques, this framework aims to decipher the operational mechanisms of neural networks by interpreting the network's structure as a representation of information flow. This approach distinguishes itself from traditional neural network research, which frequently focuses on individual neuron activation without considering the overarching network topology. The research emphasizes a holistic view, recognizing the importance of the network's interconnectedness in shaping information processing. The methodology presented offers a novel perspective for understanding complex neural systems, potentially leading to improved network design and analysis.

TOWARDS FAULT PROPAGATION MODELING IN THE WLCG USING GRAPH NEURAL NETWORKS

Reliable data movement is essential to the Worldwide LHC Computing Grid. This project asks whether FTS queues that share a storage endpoint contain useful information about one another’s future state. I process 52,037,899 raw queue records from January 2026 into 23,820,642 sparse queue states sampled every 2 minutes. At each prediction time, observed storage endpoints form graph nodes and directed FTS queues form temporal edges. A convolutional neural network (CNN) followed by a long short-term memory (LSTM) network first encodes the previous 20 minutes of every queue independently. One simple message-passing layer then averages incident edge embeddings at each endpoint and returns the source and destination context to the target edge. The resulting graph neural network (GNN), a parameter-matched multilayer perceptron (MLP), and a degree-preserving random graph are compared to isolate the effect of real WLCG endpoint assignment. On the final test period, real topology did not give a convincing advantage for throughput regression: random topology performed at least as well, and persistence retained the lowest mean absolute error. In contrast, the real GNN reached 0.5704 ± 0.0049 average precision for 20-minute bad-link onset, compared with 0.5212 ± 0.0031 for random topology. As a sanity check, the GNN was also compared with two simple rules based on the current bad states at the two endpoints. The stronger rule reached only 0.2627 AP. The graph advantage remained in direct and autoregressive bad-state forecasts up to 60 minutes. Matched seven-input regressions gave a target-dependent result: real topology improved success-rate MSE, while throughput showed no clear graph advantage when the inputs and selected queue windows were kept the same. This makes an explanation based only on classification being easier less likely. The results suggest that endpoint context is useful for degradation-related FTS controller quantities, but not clearly for workload-driven throughput. This is a first step towards fault-propagation modelling. The current model predicts only queues observed at the forecast origin; it does not predict future queue appearance or disappearance.

Relational Quantum Ground State Prediction: A Green AI Approach to the 2D Transverse Field Ising Model

Simulating quantum many-body ground states is constrained by the exponential scaling of the Hilbert space. Recent approaches rely on massive, parameter-heavy neural networks and resource-intensive Variational Monte Carlo (VMC) techniques. We compare two paradigms applied to the 2D Transverse Field Ising Model (TFIM) on an 8,100-qubit unfrustrated square lattice: Hamilton-Zero (a 547-million parameter foundation model) and ODSA 1 (Ontometric Dual Stream Architecture, a proprietary ML design pattern and hyper-compact graph-topological surrogate with ~154,000 parameters). The ODSA architecture leverages relational calculus and a 1/N fractional mass term to enable scale-invariant zero-shot extrapolation from micro-cluster training. It achieves 0.7% error against the exact thermodynamic limit (-2.03129) in 83 milliseconds on a standard CPU. In contrast, under zero-shot evaluation at 8,100 qubits on its own large-system benchmark, Hamilton-Zero produces an unphysical positive energy (+0.128, against a reference near −0.50), and no fine-tuned result at that scale is reported. On 100 disordered spin glass instances, ODSA evaluates each in 93 milliseconds — a speedup of approximately 400 hours against equivalent VMC. We conclude that mathematical reframing via relational topology dramatically outperforms brute-force parameter scaling, achieving deterministic physical convergence with a fraction of the carbon and hardware footprint.

Phase-Difference-Based Single-Anchor Visible Light Positioning Enhanced by Graph Neural Networks

Dynamic Circuit Synthesis via Generative Graph Neural Networks

This paper presents a novel approach to digital circuit synthesis leveraging Generative Graph Neural Networks (GNNs). The core claim is that designing complex digital circuits is a significantly time-consuming and iterative process. To address this, we propose a system that utilizes a GNN to automatically generate circuit designs from high-level specifications. The system learns a mapping between abstract functional requirements and the corresponding digital circuit components, optimizing for performance and resource utilization. Our approach offers the potential for substantial efficiency gains in circuit design, moving away from manual, iterative methods. The system dynamically adapts to the specification, creating a truly dynamic synthesis process. The generated circuits are represented as graphs, allowing the GNN to effectively learn and generate complex network structures. This work contributes to the automation of circuit design, a crucial step towards accelerating innovation in digital hardware.

基于图神经网络的复杂系统建模与模拟

This paper presents a novel approach to complex system modeling and simulation utilizing graph neural networks (GNNs). The core focus is to develop a method capable of automatically learning complex relationships within a system, and generating high-fidelity predictions. Traditional methods often struggle to effectively model dynamic relationships. Graph Neural Networks (GNNs) offer a powerful tool for this purpose. This paper details the architecture of our proposed GNN-based system, its training methodology, and its demonstrated performance in simulating a complex system. We provide a comprehensive analysis of the method's strengths and weaknesses, highlighting its potential for advancing the field of complex system modeling and simulation.

基于动态图的自适应图神经网络

This paper introduces a novel self-adaptive graph neural network (GNN) architecture based on dynamic graph representations. Traditional GNNs often struggle with the inherent limitations of static graphs, hindering their ability to effectively handle complex and dynamic data. Our proposed method leverages a dynamic graph structure, enabling the network to adapt to evolving data patterns and improve model generalization. We detail a mechanism for dynamically updating the graph's structure based on incoming data, facilitating a more responsive and robust learning process. The design incorporates a novel update rule that ensures a balance between adaptation and stability, resulting in enhanced performance across a range of datasets. We demonstrate the effectiveness of our approach through extensive experiments on several benchmark datasets, showcasing significant improvements in accuracy and generalization compared to existing GNN methods. This work establishes a foundation for more adaptive and robust GNN models, particularly suitable for applications involving data with significant temporal or structural changes.

Self-Supervised Learning of Graph Embeddings via Contrastive Predictive Coding

Existing self-supervised learning methods for graph embeddings often rely on heuristic approaches without a robust theoretical underpinning. This work introduces a novel framework for learning graph embeddings based on contrastive predictive coding (CPC). The core idea is to train a neural network to predict future graph states given their past states, utilizing a contrastive loss function to promote similarity between embeddings of structurally similar graphs. This approach offers a more principled and stable learning process compared to existing methods. We demonstrate the effectiveness of our approach through a theoretical analysis and outline the key components and operational details. The primary contribution lies in establishing a clear connection between predictive modeling and graph embedding learning, providing a foundation for future research in this area. The proposed method aims to generate embeddings that capture the underlying structural information of graphs effectively, ultimately improving performance in downstream graph-related tasks. ---

基于图的复杂系统建模的自适应学习

This paper introduces a novel method for complex system modeling based on graph neural networks (GNNs) and adaptive learning. Traditional approaches often require manual design of complex network structures, limiting the model's flexibility. Our method leverages GNNs to automatically learn and update graph structures, enabling more accurate representation of system-internal relationships. We propose an adaptive learning algorithm that iteratively refines the graph structure based on data observations, leading to improved model performance. The core mechanism involves utilizing GNNs to capture the intricate dependencies within a system, while the adaptive learning ensures the model continuously adjusts its representation to better reflect the underlying dynamics. The proposed approach offers a significant advancement in the modeling of complex systems by providing a flexible and automated framework for graph structure learning and dynamic representation refinement.