Graph representation learning has emerged as a powerful technique for analyzing complex relationships within graph-structured data. However, existing methods often struggle to effectively capture the temporal dynamics inherent in many real-world graphs, where relationships evolve over time. This research introduces a novel approach that addresses this limitation by integrating the information bottleneck principle with temporal constraints. The core idea is to learn compressed representations of nodes that minimize information loss while simultaneously enforcing consistency between representations across different time steps. This is achieved through a variational information bottleneck applied to the graph adjacency matrix, coupled with a recurrent neural network that captures temporal dependencies. The resulting model offers improved graph representation learning capabilities, particularly for datasets where temporal evolution plays a significant role. The key contributions of this work lie in the synergistic combination of information theory and recurrent neural networks within a graph learning framework. Specifically, we aim to define a loss function that balances the compression of node representations with the preservation of temporal relationships. The effectiveness of this approach will be evaluated on benchmark graph datasets exhibiting temporal characteristics.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel approach to graph theory based on neural rings – a dynamic neural network representation of graph structure. The core innovation lies in employing a neural ring that learns and adjusts its parameters to optimize the graph's inherent properties, leading to enhanced information propagation efficiency. Traditional graph theory methods often require extensive manual design. This system leverages a dynamically evolving neural ring, offering a more automated and adaptable framework for graph optimization. We detail the system's architecture, training process, and performance evaluation, demonstrating significant improvements in information flow compared to existing methods.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper investigates a novel dynamic approach to adaptive neural network weight adjustment, specifically focusing on the utilization of graph neural networks (GNNs) for dynamic node weighting. We introduce a method that intelligently adjusts node weights during training based on the network's internal state and data characteristics, addressing limitations inherent in traditional approaches. The core of this strategy lies in the inherent self-adaptability of GNNs, enabling a more robust and flexible model. This results in improved prediction accuracy, particularly in scenarios where the data distribution shifts significantly. The paper details the proposed algorithm, its implementation, and a comprehensive evaluation demonstrating its efficacy across a range of datasets.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel approach to symbolic manipulation learning using self-supervised learning techniques. We leverage Recurrent Neural Networks (RNNs) to learn manipulation rules directly from data, combined with Graph Representation Learning (GRL) to capture the structural complexities inherent in symbolic domains. The core idea is to train an RNN to predict the outcome of applying a sequence of symbolic operations on a graph representation of the input, eliminating the need for explicit, human-annotated labels. This approach enables RNNs to develop sophisticated reasoning capabilities without relying on traditional supervised learning paradigms. We argue that the combination of RNNs' sequential processing power and GRL's ability to encode structural information provides a strong foundation for learning symbolic manipulation, potentially unlocking more robust and adaptable systems for automated reasoning. The effectiveness of this strategy is demonstrated through the framework's design and theoretical justification, paving the way for future research and development in this area.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel approach to visual event detection that leverages multi-scale information fusion. The core idea is to enhance detection accuracy and robustness by integrating visual information at different scales, such as low-level textures and high-level semantic cues. We propose a framework utilizing multi-scale convolutional neural networks (CNNs) to extract features at various resolutions. Subsequently, an attention mechanism or graph neural network (GNN) is employed to effectively fuse these features. Unlike existing methods that primarily rely on single-scale visual features, our approach addresses a critical gap in the field, offering a more comprehensive and robust solution for visual event detection. The proposed method demonstrates improved performance across various benchmark datasets.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Predicting drug-target interactions (DTIs) is a crucial task in drug discovery, aiming to identify potential therapeutic candidates and minimize adverse effects. Traditional methods often rely on high-throughput screening or homology modeling, which are time-consuming and expensive. This paper proposes a novel approach utilizing Graph Neural Networks (GNNs) to predict DTIs based on the structural similarity between drugs and their target proteins. We represent both drugs and proteins as graphs, where nodes represent atoms and edges represent bonds, and employ Graph Convolutional Networks (GCNs) to learn embeddings that capture the structural features of these molecules. The learned embeddings are then used to predict the likelihood of interaction between a drug and a target protein. Our results demonstrate that GNNs can achieve competitive accuracy in DTI prediction, offering a computationally efficient and structurally informed alternative to existing methods. The core claim is that GNNs can accurately predict DTIs by leveraging the structural similarity between drugs and their target proteins. The core mechanism involves representing drugs and proteins as graphs and utilizing GCNs to learn embeddings that capture their structural features and predict interactions based on similarity.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper presents a novel approach to neural symbolic reasoning by integrating distributed neural networks with knowledge graphs. The core claim is that this hybrid architecture can achieve more efficient and reliable symbolic reasoning, ultimately boosting the generalization capabilities of neural reasoning systems. The proposed mechanism involves constructing a distributed system where neural networks perform perception and pattern learning, while knowledge graphs provide contextual background knowledge and reasoning rules. These components communicate asynchronously, facilitating information exchange and inference. We address the limitations of existing methods, which either rely on single-machine neural inference or static, pre-built knowledge graphs, by introducing a dynamic, interactive framework for knowledge fusion and reasoning. This approach allows for continuous learning and adaptation based on evolving knowledge, leading to a more robust and flexible reasoning system. The system is designed to handle complex reasoning tasks involving multiple entities and relationships, and the asynchronous communication ensures scalability and fault tolerance. The ultimate goal is to create a reasoning system that can not only perform accurate inferences but also explain its reasoning process, fostering trust and transparency.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces Neural Symbolic Time Table Reasoning (NSTI), a novel framework for understanding complex time series data by integrating the strengths of neural networks and symbolic reasoning. The core claim is that combining neural networks with symbolic time table inference provides a more interpretable and robust understanding of temporal data compared to traditional time series models. NSTI employs an encoder network to extract feature vectors from the time series data, which are then fed into a symbolic reasoning engine based on time tables. This engine utilizes knowledge graphs and rules to infer events, relationships, and patterns within the time series. A trainable bridge module facilitates interaction between the neural network and the symbolic engine, learning to map neural features to symbolic representations and guiding the reasoning process. This approach represents a significant advance over standalone neural time series prediction or rule-based symbolic time table systems, offering a synergistic combination for enhanced understanding and reasoning capabilities. The proposed method aims to address the limitations of existing approaches by providing both the fine-grained feature extraction capabilities of neural networks and the logical inference and explainability of symbolic systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Predicting drug-target interactions (DTIs) is a crucial task in drug discovery, aiming to identify potential therapeutic candidates and minimize adverse effects. Traditional methods often rely on high-throughput screening or homology modeling, which are time-consuming and expensive. This paper proposes a novel approach utilizing Graph Neural Networks (GNNs) to predict DTIs based on the structural similarity between drugs and their target proteins. We represent both drugs and proteins as graphs, where nodes represent atoms and edges represent bonds, and employ Graph Convolutional Networks (GCNs) to learn embeddings that capture the structural features of these molecules. The learned embeddings are then used to predict the likelihood of interaction between a drug and a target protein. Our results demonstrate that GNNs can achieve competitive accuracy in DTI prediction, offering a computationally efficient and structurally informed alternative to existing methods. The core claim is that GNNs can accurately predict DTIs by leveraging the structural similarity between drugs and their target proteins. The core mechanism involves representing drugs and proteins as graphs and utilizing GCNs to learn embeddings that capture their structural features and predict interactions based on similarity.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Predicting protein structure from its amino acid sequence remains a central challenge in computational biology. Traditional methods often rely on homology modeling and ab initio approaches, which can be computationally expensive and limited in their accuracy. This work explores the potential of graph embeddings to address this challenge. We propose a novel framework where proteins are represented as graphs, with amino acid residues as nodes and interactions as edges. Graph neural networks are then employed to learn embeddings for these nodes, capturing the intricate relationships between residues. These embeddings are subsequently used to predict the 3D structure of the protein. Our approach leverages the power of graph embedding techniques, which have demonstrated success in various domains, to tackle the complex problem of protein structure prediction. We demonstrate that graph embeddings can effectively capture the structural information encoded within protein sequences, leading to improved prediction accuracy. The core claim of this work is that graph embeddings can be utilized to predict protein structure accurately, capturing the complex relationships between residues. The core mechanism involves representing a protein as a graph, learning embeddings using graph neural networks, and then predicting the 3D structure. This novel approach offers a promising direction for future research in protein structure prediction.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Current graph neural networks (GNNs) predominantly operate as passive data processors, lacking the capacity for situated reasoning – the ability to understand and respond to the context of a situation. This paper introduces Cognitive Graph Networks (CGNs), a novel framework designed to overcome this limitation. CGNs integrate sensory input with prior knowledge, leveraging a cognitive architecture that mimics human-like reasoning. This architecture incorporates contextual information and probabilistic inference, enabling CGNs to actively reason about the environment. The core innovation lies in transitioning from a passive graph representation to an active, situated reasoning process within complex environments. This framework offers a significant advancement in the field of graph neural networks, paving the way for more intelligent and adaptable AI systems. The key components of CGNs are: (1) a standard Graph Neural Network (GNN) layer for feature extraction; (2) a Contextual Memory Module (CMM) for storing and retrieving relevant knowledge; and (3) a Rule-Based Inference Engine (RBI) for applying contextual rules and performing probabilistic reasoning. The integration of these modules allows CGNs to dynamically adapt to changing environments and produce more nuanced and accurate reasoning outputs. The effectiveness of CGNs is demonstrated through a series of simulations, showcasing their superior performance compared to traditional GNNs in scenarios requiring situated reasoning.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This paper introduces a novel approach to knowledge graph reasoning by fusing neural networks with symbolic reasoning techniques. Traditional knowledge graph reasoning methods often rely heavily on probabilistic models, leading to limited interpretability and potentially sacrificing accuracy. Our proposed framework addresses these limitations by integrating the strengths of both approaches. Specifically, we represent relationships and rules within a knowledge graph in a symbolic format, allowing for deductive reasoning. Simultaneously, neural networks are employed for relation prediction and learning complex patterns that may not be easily captured by traditional symbolic rules. The outputs of the neural network are then integrated back into the knowledge graph, enhancing the reasoning process. We demonstrate the effectiveness of this neuro-symbolic fusion in improving both the accuracy and interpretability of knowledge graph inferences. The core claim of this work is the construction of a system capable of merging neural networks and symbolic reasoning for knowledge graph reasoning, ultimately boosting accuracy and interpretability. The central mechanism involves translating knowledge graph relationships and rules into symbolic representations, leveraging neural networks for relation prediction and inference, and then consolidating results back into the knowledge graph. This approach differentiates itself from existing methods that predominantly utilize probabilistic models, offering a more transparent and potentially more accurate solution.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
