This work proposes CASTANET, which integrates spatio-temporal graph neural networks and causal treatment effect estimation to utilize incident records while mitigating selection bias, and shows that CASTANET reduces RMSE by 4.0% overall compared to the best baseline and by 10.1% on incident-conditioned evaluation.
Toshiya Kitahara, Ryu Shirakami, Koh Takeuchi et al.· 0 citations
This dimensionality dominates the data, and a supervised ESM-2 sequence baseline was benchmarked against the ESM1v zero-shot ensemble and AlphaMissense under position-based 5-fold cross-validation, together with three architectural extensions: AlphaFold structural features, multi-task learning across paired assays, and contact-graph neural networks.
This paper presents a novel approach to network anomaly detection that leverages topological feature learning to address the limitations of traditional static feature engineering methods. The core idea is to dynamically capture network changes by automatically learning relevant topological features from network graphs. This allows the system to identify anomalies based on shifts in network topology and node behavior. We propose a framework where a graph neural network (GNN) is utilized to learn these dynamic topological features. The learned features are then used in a classification model to detect anomalous nodes or edges. Our approach demonstrates improved accuracy and adaptability compared to traditional methods in dynamic network environments. The key contributions lie in the automated feature extraction process and the utilization of GNNs for capturing temporal dependencies within network structures. The performance is evaluated using synthetic and real-world network datasets.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel framework for generating graph embeddings that explicitly accounts for the temporal dynamics of relationships within a network. Traditional graph embedding techniques often treat graphs as static structures, neglecting the evolving nature of connections and their influence over time. Our approach leverages a recurrent neural network (RNN) coupled with a custom-designed loss function that measures "temporal divergence" – the difference in embeddings of nodes at consecutive time steps – weighted by the strength and type of relationships. This allows the model to learn and represent how relationships change over time, leading to more accurate and robust graph embeddings. We argue that this method represents a significant advancement over existing static embedding techniques and offers the potential to capture emergent network behaviors and predict future network modifications with improved accuracy. The core claim of this work is that incorporating temporal dynamics dramatically enhances graph embedding capabilities. ---
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Biological regulatory networks (BRNs) govern cellular processes with inherent stochasticity, reflecting the probabilistic nature of gene expression and protein interactions. Traditional Graph Neural Networks (GNNs) often treat these networks as deterministic, leading to inaccurate representations and predictions. This work proposes a novel framework for modeling BRNs that explicitly incorporates stochasticity. We introduce probabilistic layers within a GNN architecture to represent the uncertainty in network dynamics, drawing inspiration from stochastic differential equations (SDEs). These layers allow the GNN to learn and propagate probabilistic information, capturing the random fluctuations observed in biological systems. Our approach provides a more realistic and robust model of BRNs, offering potential improvements in predicting network behavior and identifying key regulatory nodes. The core claim is to develop GNN architectures that can explicitly model the stochastic nature of biological regulatory networks, accounting for the inherent randomness in gene expression and protein interactions. The core mechanism involves introducing probabilistic layers within the GNN to represent the uncertainty in the network's dynamics, incorporating elements of stochastic differential equations. This addresses a critical gap in current GNN applications to biology, acknowledging the inherent noise in biological systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper investigates the integration of topology optimization algorithms with graph neural networks (GNNs) to develop a novel framework for efficient and robust optimization of complex topological structures. Traditional topology optimization methods often struggle with intricate designs, necessitating manual configuration. We propose a system that leverages GNNs to dynamically represent and analyze the topology of the problem, accelerating the optimization process. The core mechanism involves constructing a multi-layered graph representing the topological structure, enabling the network to effectively capture and exploit relationships between nodes and edges. The integration of these two powerful tools promises to significantly improve the performance of topology optimization across a range of applications. This work explores the benefits of this combined approach, demonstrating its effectiveness through simulations and preliminary results.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper explores a novel approach to program generation and explanation by integrating the strengths of neural networks and neuro-symbolic reasoning. Current program generation techniques often fall short due to a lack of explicit logical reasoning and interpretability. We propose a framework that combines a neural network encoder to translate problem representations into neural network states and a neuro-symbolic decoder to generate program code based on these states, leveraging knowledge graphs and logical rules. Furthermore, the neuro-symbolic decoder facilitates code explanation, providing a traceable execution path. The core claim is that this hybrid approach surpasses the limitations of single-technique methods, offering a more intelligent and explainable solution for automated program generation and understanding. The methodology outlines a system architecture, detailing the components and their interactions, and highlights the key innovations within the system.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Overlapping clinical features and the absence of objective diagnostic markers frequently lead to the misdiagnosis of adolescent bipolar disorder (BD) as major depressive disorder (MDD). However, direct comparisons of functional brain network topology between adolescents with MDD and BD remain limited, particularly regarding their associations with clinical symptom dimensions. A total of 55 adolescents with MDD, 35 with BD, and 44 healthy controls (HC) were recruited. We hypothesized that adolescents with MDD and BD would exhibit distinct patterns of functional brain network organization associated with specific clinical symptoms. Graph-theoretical analyses were used to identify disorder-specific topological alterations, and support vector machine (SVM) models were constructed using significantly altered nodal metrics as classification features, with model performance evaluated using a nested cross-validation framework. Associations between altered nodal metrics and clinical measures were also examined. Compared with MDD patients, adolescents with BD exhibited higher nodal metrics in specific nodes within the default mode network (DMN) and prefrontal regions. Relative to HC, MDD patients showed reduced nodal connectivity and efficiency in visual cortical regions. Correlation analyses revealed that the clustering coefficients of the right dorsolateral superior frontal gyrus and right orbital superior frontal gyrus were positively associated with attention/vigilance performance, whereas the clustering coefficient of the right cuneus was associated with depressive and anxiety symptoms ( p < 0.05). The linear-kernel SVM achieved a mean classification accuracy of 78.5%, a balanced accuracy of 74.0%, and an AUC of 0.739 in distinguishing BD from MDD. Adolescents with MDD and BD exhibited distinct patterns of nodal functional brain network organization, particularly within the default mode, visual, and prefrontal systems. Altered network topology was associated with cognitive and affective symptom dimensions. SVM analyses further suggested that these topological features contain information relevant to differentiating adolescent MDD from BD. These findings provide further insight into the neural mechanisms underlying adolescent affective disorders. Not applicable.
Yitong Liu, Yue Zhang, Cai Li et al.· BMC Psychiatry· 0 citations
This paper introduces a novel deep learning model designed for analyzing complex systems by directly processing topological data. The core concept revolves around representing complex data as graphs and leveraging Graph Neural Networks (GNNs) for deep learning. Traditional deep learning methods often struggle with non-Euclidean data, presenting a significant limitation when dealing with systems where relationships are more important than coordinates. This research aims to bridge this gap by providing a framework capable of learning from topological features, ultimately leading to improved insights and predictive capabilities. The model utilizes persistent homology to extract topological features and then feeds them into a GNN for further processing. We demonstrate the potential of this approach through theoretical analysis and discuss its implications for various applications. This work contributes to the growing field of topological data analysis (TDA) by integrating it with the power of deep learning.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel approach to graph embedding that leverages graph neural networks to create a 'meta-graph' of data across multiple scales. The goal is to facilitate the identification of long-range dependencies and non-local relationships within complex datasets. Current methods typically focus on local analysis, restricting the scope of investigation to individual nodes or small clusters. This work proposes a technique that allows for a more comprehensive understanding of the data's structure by representing the data as a network of interconnected nodes, enabling the analysis of relationships that extend beyond immediate neighbors. We demonstrate the effectiveness of this method through a series of quantitative and qualitative analyses, showcasing its ability to uncover hidden patterns and dependencies across multiple scales. The core mechanism involves constructing a meta-graph, where nodes represent data points and edges represent relationships. This facilitates the modeling of complex network structures and the identification of long-range connections.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Existing graph embedding techniques primarily focus on static graph representations, often neglecting the crucial aspect of temporal dynamics inherent in many real-world graphs. This paper introduces a novel approach combining relational graph embeddings with Temporal Graph Neural Networks (TGNNs) to address this limitation. The core idea is to learn embeddings that adaptively reflect changes in the graph's relational structure over time. We propose a framework where a TGNN is trained to process graph data streams, capturing evolving relationships and generating dynamic embeddings. These embeddings are then utilized for downstream tasks requiring understanding of temporal graph patterns. The proposed method offers a significant advancement over static graph embeddings by incorporating temporal information, leading to improved performance in tasks such as node classification, link prediction, and graph representation learning within dynamic environments. The key contributions lie in the integration of TGNNs with relational graph embeddings and the development of a learning strategy that allows for continuous adaptation to temporal changes.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Graph Neural Networks (GNNs) have demonstrated remarkable success in various domains, including social network analysis, drug discovery, and recommendation systems. However, their inherent complexity often leads to a "black box" problem, where it is difficult to understand the reasoning behind their predictions. This paper introduces a novel approach to explain GNN predictions by learning symbolic representations of the graph structure. We propose a model that maps the GNN's activations to a symbolic representation, enabling the generation of human-readable explanations. This method addresses the interpretability challenge in GNNs, offering a pathway to trust and confidence in their predictions. The core claim is that by leveraging symbolic representation learning, we can transform opaque GNN behavior into understandable insights. The proposed mechanism provides a foundation for building more transparent and reliable GNN-based systems.
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.
