This paper introduces a novel approach to graph embedding that explicitly incorporates temporal dynamics. Traditional graph embedding techniques typically treat graphs as static structures, neglecting the evolving relationships between nodes over time. We propose a method, Temporal Graph Embedding via Predictive Recurrence, which leverages recurrent neural networks (RNNs) and graph convolutional networks (GCNs) to learn node embeddings that capture both the current structural information of the graph and the predictive influence of past interactions. The core idea is to iteratively update node embeddings within a temporal window, using observed connections and predicted future connections to drive the learning process. The integration of a GCN layer ensures that structural information remains a crucial component of the embedding, leading to more robust and informative representations. Our approach demonstrates the potential to significantly improve graph embedding quality by accounting for the dynamic nature of real-world graphs. We provide a detailed formulation of the method, including mathematical notations, and discuss its theoretical underpinnings. ---
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel framework for neuro-symbolic reasoning, addressing the longstanding challenge of integrating neural networks and symbolic reasoning. The core idea is to leverage the strengths of both approaches by employing a Graph Neural Network (GNN) to represent knowledge and perform initial reasoning steps, followed by a rule extraction algorithm that automatically learns logical rules from the GNN's learned representations. This approach avoids the need for manual rule engineering and allows the system to adapt and refine its reasoning capabilities over time. We demonstrate the effectiveness of this framework through a theoretical analysis and provide a detailed description of the components and their interactions. The key contributions of this work include a unified architecture for neuro-symbolic reasoning, a method for extracting logical rules from GNN representations, and a mechanism for iteratively refining both the GNN and the learned rules. The system's ability to learn complex relationships and generate logical rules represents a significant step towards more robust and explainable AI systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
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.
Pavel Khudov Yakovlev, Maria del Carmen Misa Moreira, Sofia Vallecorsa· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel approach to knowledge discovery within graphs by introducing Self-Referential Graph Neural Networks (SR-GNNs). Traditional Graph Neural Networks (GNNs) operate passively, learning representations based solely on the graph structure and node features. This limits their ability to adapt and refine their understanding, particularly in complex graph scenarios. Our core idea is to equip nodes within a GNN with the capability to directly influence the network's architecture and training process, establishing a feedback loop for iterative improvement. This allows the network to dynamically adjust its parameters, node embeddings, and even structural elements, ultimately leading to a more robust and accurate representation of the graph's underlying knowledge. The SR-GNN framework is designed to address the limitations of static GNNs by incorporating a self-reflective mechanism. We demonstrate the potential of this approach through a theoretical analysis and outline the key components needed for implementation, highlighting its potential to significantly enhance knowledge discovery tasks within graph-structured data.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
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. ---
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel dynamic topology optimization algorithm based on graph neural networks (GNNs). Traditional topology optimization methods often rely on manual design, struggling to adapt to complex systems exhibiting dynamic changes. Graph Neural Networks (GNNs) offer a powerful framework for automatically learning and optimizing topology structures, addressing this challenge. We propose a new GNN architecture, parameterized by a node and edge representation, that learns to dynamically update the topology based on feedback from neighboring nodes. The algorithm's performance is evaluated on a set of complex, multi-layered topological structures, demonstrating significant improvements in optimization speed and solution quality compared to existing approaches. We highlight the advantages of this approach in handling dynamic and high-dimensional topologies.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper proposes a novel network computation model termed "Space-Time Network Computation" (STNC). The core claim is that integrating spatial and temporal dimensions offers a more efficient approach to processing data exhibiting spatial and temporal dependencies. STNC leverages an extension of Graph Neural Networks (GNNs) by introducing the concepts of time streams and spatial streams. These streams are utilized within a graph structure to simulate and compute spatial-temporal interactions. The fundamental innovation lies in overcoming the dimensional limitations of traditional network computations, providing a new paradigm for handling complex spatial-temporal data. This approach utilizes graph structures to represent relationships, incorporating temporal evolution through time streams and spatial relationships through spatial streams. The key mathematical formulation is presented as follows: Let *G* = (*V*, *E*) be a graph, where *V* represents the set of nodes (vertices) and *E* represents the set of edges. Each node *vi* ∈ *V* is associated with a state vector *si* ∈ ℝd, where *d* is the dimension of the state vector. Each edge *eij* ∈ *E* is characterized by a weight *wij* ≥ 0. A *time stream* τ is defined as a sequence of state vectors: τ = (*si(0)*, *si(1)*, *si(2)*, ..., *si(T)*) , where *si(t)* is the state vector of node *i* at time *t*. A *spatial stream* σ is defined as a set of state vectors representing the spatial neighborhood of a node: σ(*i*) = {*sj(t)* | (*j*, *i*) ∈ *E*} for each node *i*. The core computation within STNC involves updating node states based on their time streams and spatial streams. The update rule can be expressed as: *si(t+1)* = f(*si(t)*, τi, σ(*i*)) where *f* is an update function, τi is the time stream associated with node *i*, and σ(*i*) is the spatial stream for node *i*. This update function can be a neural network or a more traditional rule-based system. The goal is to iteratively evolve the state vectors of nodes over time, capturing the influence of their neighbors and the temporal dynamics. The effectiveness of STNC is predicated on the efficient representation and manipulation of these time streams and spatial streams within the graph structure. The choice of *f* is crucial and will be explored in subsequent sections. The proposed model offers a flexible framework capable of accommodating diverse spatial-temporal data and associated dynamics. Further research will focus on optimizing the update function *f* and exploring its applicability to various domains.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel dynamic adaptive graph neural network (DAGNN) architecture designed to optimize network performance through automated structural adjustments. Traditional graph neural networks often rely on fixed, pre-defined structures, limiting their adaptability to diverse data. Our approach leverages a dynamic self-adaptation algorithm that continuously refines the network's topology based on input characteristics, thereby enhancing predictive accuracy and generalization capabilities. We present a comprehensive evaluation demonstrating the effectiveness of this method compared to existing DAGNNs.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Predicting protein-protein interactions (PPIs) is a fundamental challenge in bioinformatics with significant implications for understanding biological systems and drug discovery. Traditional methods often struggle to capture the dynamic nature of PPIs, which are not static events but rather fluctuate over time due to various cellular processes. This paper proposes a novel approach utilizing Graph Neural Networks (GNNs) to predict PPIs while explicitly incorporating temporal context. We construct a graph representation of protein interactions, where nodes represent proteins and edges denote the interactions. A GNN is then trained on this graph to learn embeddings that encode both structural and temporal features associated with each protein. The model's architecture allows it to predict future PPI events based on the current state of the network, accounting for the time-dependent dynamics. Our approach demonstrates improved accuracy compared to existing methods by leveraging the temporal dimension, offering a more realistic and nuanced model of PPIs. We detail the model architecture, training procedure, and evaluation results, highlighting the benefits of incorporating temporal dynamics into PPI prediction.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Hundreds of millions of people interact with language models (LMs) every day, using them for tasks such as writing assistance and information seeking. As their range of applications grows, it becomes increasingly essential to consider the role of social context in how these systems are designed and deployed. In particular, this thesis focuses on LMs' sociolinguistic competence—how associations between linguistic variation and social dimensions can be incorporated into LMs, how they are learned and manifested, and how they can lead to harm. In the first part of the thesis, we develop computational methods that improve LMs' sociolinguistic competence by explicitly injecting social context into the model architecture. We focus on two forms of social context—social networks and geographic location—and draw on recent advances in graph neural networks and multi-task learning to integrate them into LMs. Across a range of benchmarks, the proposed methods yield substantial gains. In the second part of the thesis, we explore whether LMs acquire sociolinguistic competence as a by-product of pretraining and posttraining, without being explicitly conditioned on social context. Experiments on dialectal variation and ideological framing suggest that LMs indeed learn associations between linguistic variation and social dimensions, albeit with varying levels of detail. Beyond these sociolinguistic associations, we also examine the question of how LMs' outputs reflect ideological leanings more generally, finding substantial evidence of instability. In the third part of the thesis, we investigate the harms that associations between linguistic variation and social dimensions can produce in LMs. Focusing on African American English, we find that LMs associate its speakers with pernicious stereotypes triggered by linguistic features alone, and that current posttraining practices do not address this covert racism. Preventing such harms is a critical goal for future research to ensure safe and equitable language technology. Finally, we release new analysis tools and datasets that facilitate broader empirical study of social context in natural language processing and computational social science, supporting subsequent work in these areas.
Valentin Hofmann· Universitätsbibliothek der L...· 0 citations
This paper presents a novel approach to complex system modeling leveraging the power of graph neural networks (GNNs). We construct a model designed to predict system behavior and states by effectively capturing relationships between nodes and edges within the system. The core mechanism of this approach is the utilization of GNNs to build a predictive architecture, offering a more efficient and insightful method compared to traditional approaches. The research demonstrates the efficacy of this model through comprehensive simulations and analysis, highlighting its potential for handling complex systems with intricate dynamics.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.