Zhen Xie (519905), Wenzhe Hou (22486438), Feiyang Wu (11301268) et al. · Figshare · 2025
Readers may be interested in how the hyperparameters and are selected. To provide insight, we conducted a simple analysis on the Yeast dataset, testing various combinations of these hyperparameters. We found that (0.3,0.7) yields the best performance. In the table, the reported average q-error performance for each combination is obtained by normalizing its mean q-error with respect to the mean q-error of the (0.3,0.7) setting. As shown in Table and in the experiments reported in the main text, no significant gradient explosion or vanishing was observed. This is attributed to the use of ReLU activations in the network, the application of tanh and normalization in the term, and the logarithmic transformation of predicted and true values in the term, as in LearnSC. Together, these measures keep both errors and gradients within a controlled range. (PDF)
Cited by 0Open access
Thomas Bartz–Beielstein · Machine learning · 2024
Cited by 1
Matthias Feurer, Frank Hutter · The Springer series on challenges in machine learning · 2019
Cited by 1,147
Jinghong Xu (106352), Yuqian Zhou (6928082), Qian Liu (135614) et al. · Figshare · 2025
<div> When the solution domain, internal parameters, and initial and boundary conditions of partial differential equation (PDE) are changed, many potential characteristics of the equation’s solutions are still similar. This provides the possibility to reduce the cost of PDE operator learning through transfer learning methods. Based on Fourier neural operator (FNO), we propose a novel sparse neural operator network named λ-FNO. By introducing the λ parameter matrix and using a new pruning method to make the network sparse, the operator learning ability of λ-FNO is greatly improved. Using λ-FNO can efficiently learn the operator from the discrete initial function space on the uniform grid to the discrete equation’s solution space on the unstructured grid, which is not available in FNO. Finally, we apply λ-FNO to several specific transfer tasks of partial differential equations under conditional distributions to demonstrate its excellent transferability. The experimental results show that when the shape of the solution domain of the equation or its internal parameters change, our framework can capture the potential invariant information of its solution and complete related transfer learning tasks with less cost, higher accuracy, and faster speed. In addition, the sparse framework has excellent extension and can be easily extended to other network architectures to enhance its performance. Our model and data generation code can get through <a href="https://github.com/Xumouren12/TL-FNO" target="_blank">https://github.com/Xumouren12/TL-FNO</a>. </div>
Cited by 0Open access
Ann-Kristin Becker (9099912), Till Ittermann (528482), Markus Dörr (13132537) et al. · Figshare · 2022
<p>Prediction results from the grid-based hyperparameter optimization of the random forest model using 10-fold nested cross-validation. Rows show results for varying values of the parameter nodesize (tested values: 15, 40, 65), and columns show results for varying values of the maximal number of terminal nodes (maxnodes, tested values: 15, 40, 65). The parameter mtry was optimized separately and set to 5.</p> <p>(TIF)</p>
Cited by 0Open access
Wei Ma (299083), Yunyun Hou (18251316), Mingyu Jin (4219984) et al. · Figshare · 2024
<div><p>Multi-stage attacks are one of the most critical security threats in the current cyberspace. To accurately identify multi-stage attacks, this paper proposes an anomaly-based multi-stage attack detection method. It constructs a Multi-Stage Profile (MSP) by modeling the stable system’s normal state to detect attack behaviors. Initially, the method employs Doc2Vec to vectorize alert messages generated by the intrusion detection systems (IDS), extracting profound inter-message correlations. Subsequently, Hidden Markov Models (HMM) are employed to model the normal system state, constructing an MSP, with relevant HMM parameters dynamically acquired via clustering algorithms. Finally, the detection of attacks is achieved by determining the anomaly threshold through the generation probability (GP). To evaluate the performance of the proposed method, experiments were conducted using three public datasets and compared with three advanced multi-stage attack detection methods. The experimental results demonstrate that our method achieves an accuracy of over 99% and precision of 100% in multi-stage attack detection. This confirms the effectiveness of our method in adapting to different attack scenarios and ultimately completing attack detection.</p></div>
Cited by 0Open access
Margherita Rosnati (10765739), Vincent Fortuin (10765742) · Figshare · 2021
Cited by 0Open access
Hai Huang (140371), Jinming Wu (7432031), Xinling Tang (2247616) et al. · Figshare · 2025
<div> Deep learning, as a high-performance data analysis method, has demonstrated superior efficiency and accuracy in side-channel attacks compared to traditional methods. However, many existing models enhance accuracy by stacking network layers, leading to increased algorithmic and computational complexity, overfitting, low training efficiency, and limited feature extraction capabilities. Moreover, deep learning methods rely on data correlation, and the presence of noise tends to reduce this correlation, increasing the difficulty of attacks. To address these challenges, this paper proposes the application of an InceptionNet-based network structure for side-channel attacks. This network utilizes fewer training parameters. achieves faster convergence and demonstrates improved attack efficiency through parallel processing of input data. Additionally, a LU-Net-based network structure is proposed for denoising side-channel datasets. This network captures the characteristics of input signals through an encoder, reconstructs denoised signals using a decoder, and utilizes LSTM layers and skip connections to preserve the temporal coherence and spatial details of the signals, thereby achi-eving the purpose of denoising. Experimental evaluations were conducted on the ASCAD dataset and the DPA Contest v4 dataset for comparative studies. The results indicate that the deep learning attack model proposed in this paper effectively enhances side-channel attack performance. On the ASCAD dataset, the recovery of keys requires only 30 traces, and on the DPA Contest v4 dataset, only 1 trace is needed for key recovery. Furthermore, the proposed deep learning denoising model significantly reduces the impact of noise on side-channel attack performance, thereby improving efficiency. </div>
Cited by 0Open access
Yicheng Long (366418) · Figshare · 2025
<div> Amid substantial capital influx and the rapid evolution of online user groups, the increasing complexity of user behavior poses significant challenges to cybersecurity, particularly in the domain of vulnerability prediction. This study aims to enhance the accuracy and practical applicability of cyberspace vulnerability prediction. By incorporating the dynamics of user behavioral changes and the logic of platform scaling driven by investment, two representative cybersecurity datasets are selected for analysis: the Canadian Institute for Cybersecurity Intrusion Detection System 2017 and the Network-Based Intrusion Detection Evaluation Dataset 2015. A standardized data preprocessing pipeline is constructed, including redundancy elimination, feature selection, and sample balancing, to ensure data representativeness and compatibility. To address the limited adaptability of traditional support vector machine (SVM) models in identifying nonlinear attacks, this study introduces a distribution-driven, dynamically adaptive kernel optimization approach. This method adjusts kernel parameters or switches kernel functions in real time according to the statistical characteristics of input data, thereby improving the model’s generalization capability and responsiveness in complex attack scenarios. Performance evaluations are conducted on both datasets using cross-validation. The results show that, compared to traditional models, the improved SVM achieves an 11.2% increase in prediction accuracy. Furthermore, the model demonstrates a 22.2% improvement in computational efficiency, measured as the ratio of prediction count to processing time. It also exhibits lower false positive rates and greater stability in detecting common cyberattacks such as distributed denial of service, phishing, and malware. In addition, this study analyzes user behavioral variations under different levels of attack pressure based on network access activity. Findings indicate that during periods of high platform load, attack frequency is positively correlated with users’ defensive behavior, confirming a potential causal sequence of “capital influx—user expansion—increased attack exposure.” This study offers a practical modeling framework and empirical foundation for improving predictive performance and enhancing users’ sense of cybersecurity. </div>
Cited by 0Open access
Thomas Bartz–Beielstein · 2024
Cited by 0
2002
Cited by 0
Yi Wei Tye (22208735), XinYing Chew (8690334), Umi Kalsom Yusof (22208738) et al. · Figshare · 2025
<div> Advances in data collection have resulted in an exponential growth of high-dimensional microarray datasets for binary classification in bioinformatics and medical diagnostics. These datasets generally possess many features but relatively few samples, resulting in challenges associated with the “curse of dimensionality”, such as feature redundancy and an elevated risk of overfitting. While traditional feature selection approaches, such as filter-based and wrapper-based approaches, can help to reduce dimensionality, they often struggle to capture feature interactions while adequately preserving model generalization. Therefore, this paper introduces the Adaptive Cluster-Guided Simple, Fast, and Efficient (ACG-SFE) feature selection, a hybrid approach designed to address the challenges of high-dimensional microarray data in binary classification. ACG-SFE enhances the Simple, Fast, and Efficient (SFE) evolutionary feature selection model by integrating hierarchical clustering to dynamically group correlated features based on the optimal number of clusters determined by the Silhouette index, Davies-Bouldin score, and the feature-to-observation ratio while adaptively selecting representative features within clusters using mutual information and adjusting the selection threshold through a progress factor. This hybrid filter-wrapper approach improves feature interactions, effectively minimizing redundancy and overfitting while enhancing classification performance. The proposed model is assessed against four state-of-the-art evolutionary feature selection models on 11 high-dimensional microarray datasets. Experimental results indicate that ACG-SFE effectively selects a small yet pertinent feature subset, minimizing redundancy while attaining enhanced classification accuracy and F-measure. Furthermore, its reduced RMSE between train and test accuracy substantiates its capability to reduce overfitting, outperforming comparative models. These findings establish ACG-SFE as an effective feature selection model for handling high-dimensional microarray data in binary classification, enhancing classification accuracy while selecting minimal relevant features to reduce unnecessary complexity and the risk of overfitting. </div>
Cited by 0Open access
Divyagna Bavikadi (21526214), Ayushi Agarwal (12633909), Shashank Ganta (21526217) et al. · Figshare · 2025
Effect of <i>K</i> for threshold 1.4 for GBT model with univariate selection.
Cited by 0Open access
Xiaohua Zeng (793632), Changzhou Liang (13262354), Qian Yang (300330) et al. · Figshare · 2025
<div><p>Stock price prediction is a challenging research domain. The long short-term memory neural network (LSTM) widely employed in stock price prediction due to its ability to address long-term dependence and transmission of historical time signals in time series data. However, manual tuning of LSTM parameters significantly impacts model performance. PSO-LSTM model leveraging PSO’s efficient swarm intelligence and strong optimization capabilities is proposed in this article. The experimental results on six global stock indices demonstrate that PSO-LSTM effectively fits real data, achieving high prediction accuracy. Moreover, increasing PSO iterations lead to gradual loss reduction, which indicates PSO-LSTM’s good convergence. Comparative analysis with seven other machine learning algorithms confirms the superior performance of PSO-LSTM. Furthermore, the impact of different retrospective periods on prediction accuracy and finding consistent results across varying time spans are. Conducted in the experiments.</p></div>
Cited by 0Open access
Franck Dernoncourt, Shamim Nemati, Elias Baedorf Kassis et al. · 2016
Algorithms and features in medical studies contain many “knobs” that govern the learning process from a high-level perspective: they are called hyperparameters, and investigators typically tune them by hand. In this case study, we present three mathematically grounded techniques to automatically optimize hyperparameters, and demonstrate their use in the problem of outcome prediction for ICU patients who suffer from sepsis.
Cited by 9Open access
Divyagna Bavikadi (21526214), Ayushi Agarwal (12633909), Shashank Ganta (21526217) et al. · Figshare · 2025
Frequency of Selected Biomarkers, where <i>K</i> = 10 for multiple thresholds().
Cited by 0Open access
Congyuan Xu (3338283), Yong Zhan (523196), Guanghui Chen (316677) et al. · Figshare · 2025
<div> The network intrusion detection system (NIDS) plays a critical role in maintaining network security. However, traditional NIDS relies on a large volume of samples for training, which exhibits insufficient adaptability in rapidly changing network environments and complex attack methods, especially when facing novel and rare attacks. As attack strategies evolve, there is often a lack of sufficient samples to train models, making it difficult for traditional methods to respond quickly and effectively to new threats. Although existing few-shot network intrusion detection systems have begun to address sample scarcity, these systems often fail to effectively capture long-range dependencies within the network environment due to limited observational scope. To overcome these challenges, this paper proposes a novel elevated few-shot network intrusion detection method based on self-attention mechanisms and iterative refinement. This approach leverages the advantages of self-attention to effectively extract key features from network traffic and capture long-range dependencies. Additionally, the introduction of positional encoding ensures the temporal sequence of traffic is preserved during processing, enhancing the model’s ability to capture temporal dynamics. By combining multiple update strategies in meta-learning, the model is initially trained on a general foundation during the training phase, followed by fine-tuning with few-shot data during the testing phase, significantly reducing sample dependency while improving the model’s adaptability and prediction accuracy. Experimental results indicate that this method achieved detection rates of 99.90% and 98.23% on the CICIDS2017 and CICIDS2018 datasets, respectively, using only 10 samples. </div>
Cited by 0Open access
Derek Reiman (6293798), Godhev Kumar Manakkat Vijay (10739706), Heping Xu (95297) et al. · Figshare · 2021
Cited by 0Open access
Paul Schmidt-Barbo (18954994), Gabriel Kalweit (11047566), Mehdi Naouar (18954997) et al. · Figshare · 2024
<div><p>The classification of B cell lymphomas—mainly based on light microscopy evaluation by a pathologist—requires many years of training. Since the B cell receptor (BCR) of the lymphoma clonotype and the microenvironmental immune architecture are important features discriminating different lymphoma subsets, we asked whether BCR repertoire next-generation sequencing (NGS) of lymphoma-infiltrated tissues in conjunction with machine learning algorithms could have diagnostic utility in the subclassification of these cancers. We trained a random forest and a linear classifier via logistic regression based on patterns of clonal distribution, VDJ gene usage and physico-chemical properties of the top-n most frequently represented clonotypes in the BCR repertoires of 620 paradigmatic lymphoma samples—nodular lymphocyte predominant B cell lymphoma (NLPBL), diffuse large B cell lymphoma (DLBCL) and chronic lymphocytic leukemia (CLL)—alongside with 291 control samples. With regard to DLBCL and CLL, the models demonstrated optimal performance when utilizing only the most prevalent clonotype for classification, while in NLPBL—that has a dominant background of non-malignant bystander cells—a broader array of clonotypes enhanced model accuracy. Surprisingly, the straightforward logistic regression model performed best in this seemingly complex classification problem, suggesting linear separability in our chosen dimensions. It achieved a weighted F1-score of 0.84 on a test cohort including 125 samples from all three lymphoma entities and 58 samples from healthy individuals. Together, we provide proof-of-concept that at least the 3 studied lymphoma entities can be differentiated from each other using BCR repertoire NGS on lymphoma-infiltrated tissues by a trained machine learning model.</p></div>
Cited by 0Open access
Broti, Nawara Mahmood · Figshare · 2024
This file contains hyperparameter tuning experiments.
Cited by 0Open access