期刊,World wide web 2025年28卷3期_国家学术搜索

A computationally efficient and lightweight model for high-accuracy OCT image classification

Huangjie CaoXiaoyi LianLina ChenZhengjie Duan...

1.1-1.16页

查看更多>>摘要：Abstract Current automated retinal OCT classification systems encounter deployment limitations due to excessive computational complexity. This paper presents Light-AP-EfficientNet, a lightweight architecture employing adaptive pooling for enhanced feature representation. We first optimize the convolutional layers of EfficientNet to eliminate redundant structures, significantly reducing the model’s parameter count. Then, adaptive pooling layers are integrated to enable the model to capture both global and local features, improving its classification performance. Experimental results show that Light-AP-EfficientNet achieves 99.7% accuracy, 99.7% Recall, and 0.997 F1 Score on the UCSD dataset, requiring only 17% of the parameters of ShuffleNetV2 and 19% of the computational load of MobileNetV2. The model processes a single image in just 0.028 seconds on a CPU and 0.009 seconds on a GPU. Additionally, it outperforms recent models in accuracy, precision, and recall, with improvements of up to 4.5% in accuracy, 5.42% in precision, and 4.5% in F1 Score. With high accuracy and reduced hardware requirements, Light-AP-EfficientNet is ideal for computer vision tasks in resource-constrained environments.

原文链接:

NETL
NSTL

CCT-GNN: Collaborative category and time-aware graph neural networks for session-based recommendation systems

Mohammad MoosazadehMarjan Kaedi

1.1-1.26页

查看更多>>摘要：Abstract A session-based recommendation system (SBRS) focuses on the user’s interactions in the current session to provide recommendations. Recent works employ graph neural networks (GNN) to capture complex relationships between clicked items in a session. Some of these studies incorporate collaborative information from similar/neighbor sessions. To identify neighbor sessions, they calculate the similarity of sessions by counting the common items between two sessions. But this is too simplistic and the similarity depends on other features such as the order of common items. Furthermore, meta-data such as items’ categories have not been considered, while items can be categorized into a limited number of categories, which can be informative in finding similar sessions and predicting user intent. In this paper, we propose a novel method, named Collaborative Category and Time-aware Graph Neural Networks (CCT-GNN), which models users’ interactions in two levels: (i) Global-level, which identifies neighbor sessions effectively and explores collaborative information to improve model performance. (ii) Local-level, in which the current session interactions are transformed into an item category graph to model different types of relations between items and categories. Experimental results demonstrate CCT-GNN superiority over state-of-the-art methods. Source code is available at: https://github.com/Moosazadeh/CCT-GNN.

原文链接:

NETL
NSTL
Springer Nature

Geometry fusion representation for knowledge graph completion using multi-view information bottleneck

Kai ChenHan YuXin SongYe Wang...

1.1-1.25页

查看更多>>摘要：Abstract The inherent incompleteness of Knowledge Graphs (KGs) has spurred significant research efforts in the domain of knowledge graph completion (KGC), which is grounded in the premise of leveraging existing knowledge to infer the unknown and fill in the gaps. A multitude of studies have focused on representation learning across a spectrum of geometric spaces, namely Euclidean, hyperbolic, and spherical. Each of these spaces excels in modeling distinct structural and characteristic elements within KGs, thereby enhancing the ability to uncover and reason about missing knowledge. Recognizing the distinct modeling advantages of each space, there is a growing effort to integrate these disparate geometries. Despite these efforts, current fusion methods, which rely on simplistic weighted summation, fail to adequately focus on the objectives of reasoning and tend to retain a significant amount of redundant information. To address this, we propose the Multi-View Information Bottleneck based Geometry Fusion (MVIBGF) method for KGC. We utilize a multi-view learning approach, aligning each view with a unique geometric space, and apply the Information Bottleneck principle to enhance mutual information with the inference goal while minimizing it within individual views. This process ensures that only the most salient and reasoning-relevant information is retained, thus enhancing the precision and efficiency of the knowledge integration. Experiments show that MVIBGF outperforms existing KGC methods, proving its robustness and effectiveness in KG modeling.

原文链接:

NETL
NSTL
Springer Nature

A computationally efficient and lightweight model for high-accuracy OCT image classification

Huangjie CaoXiaoyi LianLina ChenZhengjie Duan...

1.1-1.16页

查看更多>>摘要：Abstract Current automated retinal OCT classification systems encounter deployment limitations due to excessive computational complexity. This paper presents Light-AP-EfficientNet, a lightweight architecture employing adaptive pooling for enhanced feature representation. We first optimize the convolutional layers of EfficientNet to eliminate redundant structures, significantly reducing the model’s parameter count. Then, adaptive pooling layers are integrated to enable the model to capture both global and local features, improving its classification performance. Experimental results show that Light-AP-EfficientNet achieves 99.7% accuracy, 99.7% Recall, and 0.997 F1 Score on the UCSD dataset, requiring only 17% of the parameters of ShuffleNetV2 and 19% of the computational load of MobileNetV2. The model processes a single image in just 0.028 seconds on a CPU and 0.009 seconds on a GPU. Additionally, it outperforms recent models in accuracy, precision, and recall, with improvements of up to 4.5% in accuracy, 5.42% in precision, and 4.5% in F1 Score. With high accuracy and reduced hardware requirements, Light-AP-EfficientNet is ideal for computer vision tasks in resource-constrained environments.

原文链接:

NETL
NSTL

CCT-GNN: Collaborative category and time-aware graph neural networks for session-based recommendation systems

Mohammad MoosazadehMarjan Kaedi

1.1-1.26页

查看更多>>摘要：Abstract A session-based recommendation system (SBRS) focuses on the user’s interactions in the current session to provide recommendations. Recent works employ graph neural networks (GNN) to capture complex relationships between clicked items in a session. Some of these studies incorporate collaborative information from similar/neighbor sessions. To identify neighbor sessions, they calculate the similarity of sessions by counting the common items between two sessions. But this is too simplistic and the similarity depends on other features such as the order of common items. Furthermore, meta-data such as items’ categories have not been considered, while items can be categorized into a limited number of categories, which can be informative in finding similar sessions and predicting user intent. In this paper, we propose a novel method, named Collaborative Category and Time-aware Graph Neural Networks (CCT-GNN), which models users’ interactions in two levels: (i) Global-level, which identifies neighbor sessions effectively and explores collaborative information to improve model performance. (ii) Local-level, in which the current session interactions are transformed into an item category graph to model different types of relations between items and categories. Experimental results demonstrate CCT-GNN superiority over state-of-the-art methods. Source code is available at: https://github.com/Moosazadeh/CCT-GNN.

原文链接:

NETL
NSTL
Springer Nature

Geometry fusion representation for knowledge graph completion using multi-view information bottleneck

Kai ChenHan YuXin SongYe Wang...

1.1-1.25页

查看更多>>摘要：Abstract The inherent incompleteness of Knowledge Graphs (KGs) has spurred significant research efforts in the domain of knowledge graph completion (KGC), which is grounded in the premise of leveraging existing knowledge to infer the unknown and fill in the gaps. A multitude of studies have focused on representation learning across a spectrum of geometric spaces, namely Euclidean, hyperbolic, and spherical. Each of these spaces excels in modeling distinct structural and characteristic elements within KGs, thereby enhancing the ability to uncover and reason about missing knowledge. Recognizing the distinct modeling advantages of each space, there is a growing effort to integrate these disparate geometries. Despite these efforts, current fusion methods, which rely on simplistic weighted summation, fail to adequately focus on the objectives of reasoning and tend to retain a significant amount of redundant information. To address this, we propose the Multi-View Information Bottleneck based Geometry Fusion (MVIBGF) method for KGC. We utilize a multi-view learning approach, aligning each view with a unique geometric space, and apply the Information Bottleneck principle to enhance mutual information with the inference goal while minimizing it within individual views. This process ensures that only the most salient and reasoning-relevant information is retained, thus enhancing the precision and efficiency of the knowledge integration. Experiments show that MVIBGF outperforms existing KGC methods, proving its robustness and effectiveness in KG modeling.

原文链接:

NETL
NSTL
Springer Nature

JAL: an algebra for JSON query optimization

Anne Jasmijn LangerakFlavius FrasincarJasmijn Klinkhamer

1.1-1.42页

查看更多>>摘要：Abstract As databases become larger and less structured, the JavaScript Object Notation (JSON) data format has risen in usage compared to other data formats like XML. At the same time, while extracting data from these large datasets efficiently is of obvious importance, there has been far less research regarding the optimization of JSON queries than there has relating to the querying of XML data. Thus a JSON Data Model and JSON Algebra (JAL) are proposed, as well as a heuristic optimization algorithm, for the purpose of improving the efficiency of queries of JSON data. We implement the proposed algorithm and compare the efficiency gain that it provides in terms of both the theoretical and physical cost of executing queries. We find that the algorithm significantly reduces query costs compared to an unoptimized baseline. Additionally, we find that the efficiency gain is considerably larger when querying databases with many documents than those with relatively fewer documents.

原文链接:

NETL
NSTL
Springer Nature

JAL: an algebra for JSON query optimization

Anne Jasmijn LangerakFlavius FrasincarJasmijn Klinkhamer

1.1-1.42页

查看更多>>摘要：Abstract As databases become larger and less structured, the JavaScript Object Notation (JSON) data format has risen in usage compared to other data formats like XML. At the same time, while extracting data from these large datasets efficiently is of obvious importance, there has been far less research regarding the optimization of JSON queries than there has relating to the querying of XML data. Thus a JSON Data Model and JSON Algebra (JAL) are proposed, as well as a heuristic optimization algorithm, for the purpose of improving the efficiency of queries of JSON data. We implement the proposed algorithm and compare the efficiency gain that it provides in terms of both the theoretical and physical cost of executing queries. We find that the algorithm significantly reduces query costs compared to an unoptimized baseline. Additionally, we find that the efficiency gain is considerably larger when querying databases with many documents than those with relatively fewer documents.

原文链接:

NETL
NSTL
Springer Nature

A temporal dependency preserving approach for anomaly detection on multivariate time series

Seif-Eddine BenkabouKhalid BenabdeslemDou El Kefel MansouriSouleyman Chaib...

1.1-1.37页

查看更多>>摘要：Abstract Multivariate time series present significant methodological challenges for anomaly detection, primarily due to the intricate nature of their temporal dependencies and the dynamic interplay among variables. These complexities render traditional methods inadequate for precise and reliable anomaly detection. This paper confronts these challenges by introducing an innovative, unsupervised framework that concurrently integrates data encoding, preservation of temporal structure, and residual analysis. By modeling temporal dependencies through sinusoidal exponential decay functions, our approach identifies deviations from this model in the latent space as anomalies. We validate the effectiveness of our framework through extensive experiments on real-world datasets, benchmarking it against state-of-the-art approaches.

原文链接:

NETL
NSTL
Springer Nature

DySpec: Faster speculative decoding with dynamic token tree structure

Yunfan XiongRuoyu ZhangYanzeng LiLei Zou...

1.1-1.26页

查看更多>>摘要：Abstract While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate. Prevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fail to generalize to diverse query distributions. In this paper, we propose DySpec, a faster speculative decoding algorithm with a novel dynamic token tree structure. We begin by bridging the draft distribution and acceptance rate from intuitive and empirical clues and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run-time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, DySpec yields a higher acceptance rate and acceleration than fixed trees. DySpec can drastically improve throughput and reduce latency of token generation across various data distribution and model sizes, which outperforms strong competitors significantly, including Specinfer and Sequoia. Under low temperature setting, DySpec can improve throughput up to 9.1 ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} and reduce latency up to 9.4 ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} on Llama2-70B. Under high temperature setting, DySpec can also improve throughput up to 6.21 ×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document}, despite the increasing difficulty of speculating more than one token per step for the draft model.

原文链接:

NETL
NSTL
Springer Nature