Standard Metrics Research Articles

Hybrid anomaly detection: Leveraging autoencoder for feature learning and random forest neural network for discriminative classification

Anomaly detection, a critical aspect of data analysis and cybersecurity, aims to identify unusual patterns that deviate from the expected norm. In this study, we propose a hybrid approach that combines the strengths of Autoencoder neural networks and Multiclass Support Vector Machines (SVM) for robust anomaly detection. The Autoencoder is utilized for feature learning and extraction, capturing intricate patterns in the data, while the Multiclass SVM provides a discriminative classification mechanism to distinguish anomalies from normal patterns. Specifically, the Autoencoder is trained on normal data to acquire a compact and efficient representation of the underlying patterns, with the reconstruction errors serving as indicative measures of anomalies. Concurrently, a Multiclass SVM is trained to classify instances into multiple classes, including an anomaly class. The anomaly scores from the Autoencoder and the decision function of the Multiclass SVM, along with that of the Random Forest Neural Network (AE-RFNN), are combined, leveraging their complementary strengths. A thresholding mechanism is then employed to classify instances as normal or anomalous based on the combined scores. The performance of the hybrid model is evaluated using standard metrics such as precision, recall, F1-score, and the area under the Receiver Operating Characteristic (ROC) curve. The proposed hybrid anomaly detection approach demonstrates effectiveness in capturing complex patterns and discerning anomalies across diverse datasets. Additionally, the model offers flexibility for adaptation to evolving data distributions. This study contributes to the advancement of anomaly detection methodologies by presenting a hybrid solution that combines feature learning and discriminative classification for improved accuracy and generalization.

Machine learning based detection of T–wave alternans in real ambulatory conditions

Background and objectiveT-wave alternans (TWA) is a fluctuation in the repolarization morphology of the ECG. It is associated with cardiac instability and sudden cardiac death risk. Diverse methods have been proposed for TWA analysis. However, TWA detection in ambulatory settings remains a challenge due to the absence of standardized evaluation metrics and detection thresholds. MethodsIn this work we use traditional TWA analysis signal processing-based methods for feature extraction, and two machine learning (ML) methods, namely, K–nearest–neighbor (KNN) and random forest (RF), for TWA detection, addressing hyper–parameter tuning and feature selection. The final goal is the detection in ambulatory recordings of short, non-sustained and sparse TWA events. ResultsWe train ML methods to detect a wide variety of alternant voltage from 20 to 100 μV, i.e., ranging from non–visible micro–alternans to TWA of higher amplitudes, to recognize a wide range in concordance to risk stratification. In classification, RF outperforms significantly the recall in comparison with the signal processing methods, at the expense of a small lost in precision. Despite ambulatory detection stands for an imbalanced category context, the trained ML systems always outperform signal processing methods. ConclusionsWe propose a comprehensive integration of multiple variables inspired by TWA signal processing methods to fed learning-based methods. ML models consistently outperform the best signal processing methods, yielding superior recall scores.

Fast and stable pedicel detection for robust visual servoing to harvest shaking fruits

This study introduces a fast and stable pedicel detection method for robust visual servoing (RVS), supplemented with video stabilization (ViS) and fast pedicel detection, to realize automated harvesting of shaking fruits. By incorporating ViS, the system effectively mitigates the effects of motion blur, thereby ensuring consistent and precise object detection. In addition, the Fourier spectrum-based band-stop filter (FSBF) is used to improve clarity. The proposed approach also leverages the fast point feature histogram (FPFH) for fast pedicel detection, achieving real-time detection rates of 15–37 fps. Furthermore, it incorporates 6D pose estimation, culminating in the implementation of a 6D pose-based robust visual servoing (6DRVS) system. The performance of this system is evaluated using standard metrics such as perception accuracy, approach accuracy, precision, recall, accuracy, and F1-score in both preliminary tests and on-site experiments at two cucumber farms in Korea. The 6DRVS, supplemented with fast pedicel detection and ViS, exhibited improvements across all evaluation metrics. It recorded 90.00% perception accuracy, 82.22% approach accuracy, 0.957 precision, 0.938 recall, 0.900 accuracy, and 0.947 F1-score, highlighting its essential role in ensuring precise and efficient harvesting.

Towards Biologically Plausible and Private Gene Expression Data Generation

Generative models trained with Differential Privacy (DP) are becoming increasingly prominent in the creation of synthetic data for downstream applications. Existing literature, however, primarily focuses on basic benchmarking datasets and tends to report promising results only for elementary metrics and relatively simple data distributions. In this paper, we initiate a systematic analysis of how DP generative models perform in their natural application scenarios, specifically focusing on real-world gene expression data. We conduct a comprehensive analysis of five representative DP generation methods, examining them from various angles, such as downstream utility, statistical properties, and biological plausibility. Our extensive evaluation illuminates the unique characteristics of each DP generation method, offering critical insights into the strengths and weaknesses of each approach, and uncovering intriguing possibilities for future developments. Perhaps surprisingly, our analysis reveals that most methods are capable of achieving seemingly reasonable downstream utility, according to the standard evaluation metrics considered in existing literature. Nevertheless, we find that none of the DP methods are able to accurately capture the biological characteristics of the real dataset. This observation suggests a potential over-optimistic assessment of current methodologies in this field and underscores a pressing need for future enhancements in model design.

Silicate Mineralogy from Vis–NIR Reflectance Spectra

The asteroid composition is the key to understanding the origin and evolution of the solar system. The composition is imprinted at specific wavelengths of the asteroid reflectance spectra. We wish to find the optimal wavelength range and step of reflectance spectra that contain sufficient information about S-complex asteroids while keeping the data volume as low as possible. We especially aim for the ASPECT instrument on board the Milani/Hera CubeSat that will observe the S-complex binary asteroid (65803) Didymos–Dimorphos. We use labeled reflectance spectra of the most common silicate found in meteorites, namely olivine, orthopyroxene, clinopyroxene, and their mixtures. The spectra are interpolated to various wavelength grids. We use convolutional neural networks and train them with the labeled interpolated reflectance spectra. The reliability of the network outputs is evaluated using standard regression metrics. We do not find any significant dependence between the error of the model predictions and normalization position, fineness of coverage within the 1 μm band, and wavelength step up to 50 nm. High-precision predictions of the olivine and orthopyroxene modal abundances are obtained using spectra that cover wavelengths from 750 to 1050 nm and from 750 to 1250 nm, respectively. For high-precision predictions of the olivine chemical composition, the spectra should cover wavelengths from 750 to 1550 nm. The orthopyroxene chemical composition can be estimated from spectra that cover wavelengths from 750 to 1350 nm. We design a simple web interface through which everybody can use the pretrained models.

Reconstructing brain functional networks through identifiability and deep learning.

We propose a novel approach for the reconstruction of functional networks representing brain dynamics based on the idea that the coparticipation of two brain regions in a common cognitive task should result in a drop in their identifiability, or in the uniqueness of their dynamics. This identifiability is estimated through the score obtained by deep learning models in supervised classification tasks and therefore requires no a priori assumptions about the nature of such coparticipation. The method is tested on EEG recordings obtained from Alzheimer's and Parkinson's disease patients, and matched healthy volunteers, for eyes-open and eyes-closed resting-state conditions, and the resulting functional networks are analysed through standard topological metrics. Both groups of patients are characterised by a reduction in the identifiability of the corresponding EEG signals, and by differences in the patterns that support such identifiability. Resulting functional networks are similar, but not identical to those reconstructed by using a correlation metric. Differences between control subjects and patients can be observed in network metrics like the clustering coefficient and the assortativity in different frequency bands. Differences are also observed between eyes open and closed conditions, especially for Parkinson's disease patients.

Keratoconus disease classification with multimodel fusion and vision transformer: a pretrained model approach

ObjectiveOur objective is to develop a novel keratoconus image classification system that leverages multiple pretrained models and a transformer architecture to achieve state-of-the-art performance in detecting keratoconus.Methods and analysisThree pretrained...

Sensitivity of HAFS-B Tropical Cyclone Forecasts to Planetary Boundary Layer and Microphysics Parameterizations

Abstract Understanding how model physics impact tropical cyclone (TC) structure, motion, and evolution is critical for the development of TC forecast models. This study examines the impacts of microphysics and planetary boundary layer (PBL) physics on forecasts using the Hurricane Analysis and Forecast System (HAFS), which is newly operational in 2023. The “HAFS-B” version is specifically evaluated, and three sensitivity tests (for over 400 cases in 15 Atlantic TCs) are compared with retrospective HAFS-B runs. Sensitivity tests are generated by 1) changing the microphysics in HAFS-B from Thompson to GFDL, 2) turning off the TC-specific PBL modifications that have been implemented in operational HAFS-B, and 3) combining the PBL and microphysics modifications. The forecasts are compared through standard verification metrics, and also examination of composite structure. Verification results show that Thompson microphysics slightly degrades the days 3–4 forecast track in HAFS-B, but improves forecasts of long-term intensity. The TC-specific PBL changes lead to a reduction in a negative intensity bias and improvement in RI skill, but cause some degradation in prediction of 34-kt (1 kt ≈ 0.51 m s−1) wind radii. Composites illustrate slightly deeper vortices in runs with the Thompson microphysics, and stronger PBL inflow with the TC-specific PBL modifications. These combined results demonstrate the critical role of model physics in regulating TC structure and intensity, and point to the need to continue to develop improvements to HAFS physics. The study also shows that the combination of both PBL and microphysics modifications (which are both included in one of the two versions of HAFS in the first operational implementation) leads to the best overall results. Significance Statement A new hurricane model, the Hurricane Analysis and Forecast System (HAFS), is being introduced for operational prediction during the 2023 hurricane season. One of the most important parts of any forecast model are the “physics parameterizations,” or approximations of physical processes that govern things like turbulence, cloud formation, etc. In this study, we tested these approximations in one configuration of HAFS, HAFS-B. Specifically, we looked at two different versions of the microphysics (modeling the growth of water and ice in clouds) and boundary layer physics (the approximations for turbulence in the lowest level of the atmosphere). We found that both of these sets of model physics had important effects on the forecasts from HAFS. The microphysics had notable impacts on the track forecasts, and also changed the vertical depth of the model hurricanes. The boundary layer physics, including some of our changes based on observed hurricanes and turbulence-resolving models, helped the model better predict rapid intensification (periods where the wind speed increases quickly). Work is ongoing to improve the model physics for better forecasts of rapid intensification and overall storm structure, including storm size. The study also shows the combination of both PBL and microphysics modifications overall leads to the best results and thus was used as one of the two first operational implementations of HAFS.

DLCT LUNG Detect Net: Leveraging Deep Learning for Lung Tumor Detection in CT scans

Lung cancer is a critical global health concern, necessitating precise early diagnosis and intervention for better patient outcomes. Computed Tomography (CT) scans are pivotal in lung cancer detection, and leveraging advanced technology is crucial. This study introduces "DLCTLungDetectNet," a Convolutional Neural Network (CNN) based deep learning framework, with a focus on early lung cancer detection using CT scan images. The core innovation lies in the integration of the robust "FusionNet," a hybrid model amalgamating feature from ResNet50 and InceptionV3. We conduct a comprehensive comparative analysis, showcasing the superior performance of DLCTLungDetectNet over established architectures such as VGG16, VGG19, and Inception v3. Rigorous evaluation based on standard metrics substantiates DLCTLungDetectNet's high accuracy, precision, Area Under Curve (AUC), and F1 score. This research not only highlights the potential of deep learning in enhancing lung cancer diagnosis but also establishes a benchmark, showcasing the efficacy of the FusionNet hybrid model for achieving superior accuracy in automated lung tumor detection.

Calculating Flanking Transmission for Improved Airborne Sound Insulation in Buildings

The complexities of airborne noise insulation in tropical buildings provide challenges for predicting the value of airborne noise in buildings. The analysis employs standardized metrics, the Weighted Sound Reduction Index (DnT,w), to evaluate the impact of wall, floor, and ceiling combinations on acoustic performance. Notably, the study addresses the persistent challenge of flanking transmission, contributing to disparities between predicted and measured values. This study focuses on common construction materials in Indonesia. The outcomes reveal that lightweight construction materials, such as lightweight brick walls, consistently exhibit superior sound reduction, emphasizing their potential benefits in tropical regions. Additionally, the choice of ceiling materials, particularly gypsum board over wooden ceilings, significantly influences acoustic outcomes

An Architectural Solution to a Biological Problem: A Systematic Review of Lighting Designs in Healthcare Environments

Neuroscience and biological evidence emphasizes the profound influence of natural light on human health, offering benefits such as reducing fatigue, heightened alertness in healthcare providers, and improving patient outcomes. The objective of this review is to identify scientific studies and research to evaluate and report evidence of indoor lighting conditions’ influence on health outcomes, which can be used to develop lighting designs that align circadian rhythms in healthcare settings. A comprehensive search was conducted to identify rigorous empirical studies focused on the link between interior lighting conditions with health outcomes in the healthcare environment. For a comprehensive review of the existing literature, a four-phased methodology was employed including literature search, screening, and selection. Literature appraisals were conducted to determine the relevance and quality of evidence for each study identified. In addition, using a thematic analysis, patterns were identified, analyzed, and interpreted within the literature review. Accordingly, the results were organized into two main groups interventional human subjects and simulation-based studies. Despite evidence that natural light influences human health and happiness, a synthesis of reviewed studies suggests that the evidence for the benefits of artificial lighting in healthcare settings is less conclusive, with potential factors including variations in lighting design, inconsistent implementation of lighting interventions, and differing sample populations across studies. We conclude with an executive summary suggesting that future research should use standardized metrics and methods to focus on bridging the gap between theoretical understanding and practical application in lighting design for healthcare environments. Collaboration among architects, designers, lighting experts, and healthcare professionals can address these factors contributing to building a stronger evidence-based design for the benefits of artificial lighting in healthcare settings.

Efficient Skin Lesion Detection using YOLOv9 Network

Skin lesion detection plays a crucial role in dermatological diagnosis and treatment. In this study, we propose an efficient approach for skin lesion detection using the YOLOv9 network. Leveraging state-of-the-art deep learning techniques, our model demonstrates robust performance in accurately identifying various skin lesion types, including acne, atopic dermatitis, keratosis pilaris, leprosy, psoriasis, and wart. We conducted comprehensive experiments using a curated dataset comprising 2721 training images, 288 validation images, and 145 test images. The model was trained and evaluated based on standard metrics such as Precision, Recall, and mean Average Precision (mAP). Our results indicate promising detection accuracy, with an overall Precision of 60.5%, Recall of 86.0%, and an mAP of 81.4%. Class-wise analysis reveals varying levels of performance across different disease classes, highlighting the model's proficiency in detecting common dermatological conditions such as acne and wart lesions. Furthermore, we provide insights into potential challenges and limitations, including dataset size and class imbalance, and discuss avenues for future research to address these issues. Our study contributes to the advancement of AI-driven solutions for dermatological diagnosis and underscores the efficacy of the YOLOv9 network in skin lesion detection

Diabetes Prediction Using Machine Learning Algorithm

bstract: Diabetes, a chronic metabolic disorder affecting millions worldwide, requires early detection and management to mitigate its complications. Machine learning (ML) techniques have emerged as promising tools for predictive analytics in healthcare, offering the potential to improve diagnostic accuracy and patient outcomes. This paper presents a comprehensive review of ML algorithms applied to diabetes prediction, encompassing diverse methodologies and datasets. The study evaluates the performance of various ML algorithms, including but not limited to logistic regression, decision trees, support vector machines, random forests, and deep learning approaches, in predicting the onset or progression of diabetes. Additionally, feature selection techniques and data pre-processing methods are explored to enhance model robustness and interpretability. Furthermore, this review highlights the significance of dataset characteristics such as size, imbalance, and feature diversity in influencing model performance. Challenges associated with model interpretability, scalability, and deployment in clinical settings are also discussed, alongside potential strategies to address these issues. The findings suggest that ML algorithms demonstrate promising capabilities in diabetes prediction, with many studies reporting high accuracy, sensitivity, and specificity. However, there remains a need for standardized evaluation metrics and benchmark datasets to facilitate comparisons across studies. Moreover, efforts to enhance model interpretability and address data privacy concerns are crucial for promoting the adoption of ML-based predictive models in healthcare practice. In conclusion, this review underscores the potential of ML techniques in diabetes prediction and emphasizes the importance of interdisciplinary collaboration between data scientists, clinicians, and healthcare stakeholders to leverage these advancements for improved patient care and disease management

On data set tensions and signatures of new cosmological physics

ABSTRACT Can new cosmic physics be uncovered through tensions amongst data sets? Tensions in parameter determinations amongst different types of cosmological observation, especially the ‘Hubble tension’ between probes of the expansion rate, have been invoked as possible indicators of new physics, requiring extension of the ΛCDM paradigm to resolve. Within a fully Bayesian framework, we show that the standard tension metric gives only part of the updating of model probabilities, supplying a data co-dependence term that must be combined with the Bayes factors of individual data sets. This shows that, on its own, a reduction of data set tension under an extension to ΛCDM is insufficient to demonstrate that the extended model is favoured. Any analysis that claims evidence for new physics solely on the basis of alleviating data set tensions should be considered incomplete and suspect. We describe the implications of our results for the interpretation of the Hubble tension.

ASSESSING THE POTENTIAL OF ENERGY STORAGE SOLUTIONS FOR GRID EFFICIENCY: A REVIEW

Energy storage solutions play a pivotal role in enhancing grid efficiency and reliability, offering a multitude of benefits for grid operators, utilities, and consumers alike. This comprehensive review examines the potential of energy storage technologies in optimizing grid performance. The review begins by outlining the significance of energy storage in addressing the challenges of grid modernization and the integration of renewable energy sources. It explores various types of energy storage technologies, including batteries, pumped hydro storage, compressed air energy storage, and thermal energy storage, assessing their capabilities, limitations, and suitability for grid applications. Furthermore, the review delves into the benefits that energy storage solutions bring to the grid, such as peak shaving, load leveling, frequency regulation, and grid stabilization. It analyzes real-world case studies and examples of energy storage deployments, highlighting their effectiveness in improving grid reliability and resilience while reducing operational costs. Despite the promising potential of energy storage, the review also addresses the challenges and barriers hindering widespread adoption. These include cost considerations, regulatory frameworks, technical constraints, and the need for standardized performance metrics and interoperability standards. Looking ahead, the review discusses emerging trends and opportunities in energy storage technologies, such as advancements in battery technology, grid-scale storage projects, and the integration of storage with renewable energy systems and smart grid infrastructure. It emphasizes the importance of continued research, innovation, and collaboration to unlock the full potential of energy storage solutions for grid efficiency. In conclusion, the review underscores the critical role of energy storage in optimizing grid efficiency, enhancing renewable energy integration, and ensuring grid stability and resilience. It calls for concerted efforts from policymakers, industry stakeholders, and researchers to overcome existing challenges and accelerate the deployment of energy storage technologies for a more sustainable and resilient energy future.
 Keywords: Efficiency, Grid, Solutions, Energy Storage, Potential.

MicroBundleCompute: Automated segmentation, tracking, and analysis of subdomain deformation in cardiac microbundles.

Advancing human induced pluripotent stem cell derived cardiomyocyte (hiPSC-CM) technology will lead to significant progress ranging from disease modeling, to drug discovery, to regenerative tissue engineering. Yet, alongside these potential opportunities comes a critical challenge: attaining mature hiPSC-CM tissues. At present, there are multiple techniques to promote maturity of hiPSC-CMs including physical platforms and cell culture protocols. However, when it comes to making quantitative comparisons of functional behavior, there are limited options for reliably and reproducibly computing functional metrics that are suitable for direct cross-system comparison. In addition, the current standard functional metrics obtained from time-lapse images of cardiac microbundle contraction reported in the field (i.e., post forces, average tissue stress) do not take full advantage of the available information present in these data (i.e., full-field tissue displacements and strains). Thus, we present "MicroBundleCompute," a computational framework for automatic quantification of morphology-based mechanical metrics from movies of cardiac microbundles. Briefly, this computational framework offers tools for automatic tissue segmentation, tracking, and analysis of brightfield and phase contrast movies of beating cardiac microbundles. It is straightforward to implement, runs without user intervention, requires minimal input parameter setting selection, and is computationally inexpensive. In this paper, we describe the methods underlying this computational framework, show the results of our extensive validation studies, and demonstrate the utility of exploring heterogeneous tissue deformations and strains as functional metrics. With this manuscript, we disseminate "MicroBundleCompute" as an open-source computational tool with the aim of making automated quantitative analysis of beating cardiac microbundles more accessible to the community.

Green Data Center : Reviews And Challenges.

Data is undoubtedly the currency of the 21st century. The data that we are providing to or retrieving from the users has a very important role to play in the evolution of services in both the computing and non-computing spheres. Various internet services, such as web hosting, e-commerce, and Social Networking, find their roots in the Data centers. With the ever-increasing concerns about carbon emissions, reliance on non-renewable energy resources, and energy consumption by the cooling systems in conventional Data centers, it is imperative to shift our focus to newer, greener, and cleaner technologies that will aim to reduce the overall carbon footprint, technologies that will reduce energy consumption, and the integration of renewable energy resources into the existing setup. There are several studies that have highlighted challenges regarding the adoption of technologies in the hardware and software domains and the standardization of performance metrics, which are generally used to study trends in energy consumption and net emissions and, therefore, assess the sustainability and efficiency of Data center operations. This shift to Green Data centers will certainly involve both standardized performance metrics, effective strategies to reduce energy consumption, and the integration of renewable energy resources to reduce the carbon footprint. Furthermore, we are going to discuss data center network architecture that will make our future Green Data centers scalable, fault-tolerant, agile, and energy efficient. This paper reviews various strategies for designing an efficient data center, ways to incorporate renewable energy resources, standardization and benchmarking of performance metrics, and optimizing the cooling systems to improve energy efficiency. Finally, we are going to outline the future directions and opportunities for further research and innovation.

ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision

Mirror detection is an active research topic in computer vision. However, all existing mirror detectors learn mirror representations from large-scale pixel-wise datasets, which are tedious and expensive to obtain. Although weakly-supervised learning has been widely explored in related topics, we note that popular weak supervision signals (e.g., bounding boxes, scribbles, points) still require some efforts from the user to locate the target objects, with a strong assumption that the images to annotate always contain the target objects. Such an assumption may result in the over-segmentation of mirrors. Our key idea of this work is that the existence of mirrors over a time period may serve as a weak supervision to train a mirror detector, for two reasons. First, if a network can predict the existence of mirrors, it can essentially locate the mirrors. Second, we observe that the reflected contents of a mirror tend to be similar to those in adjacent frames, but exhibit considerable contrast to regions in far-away frames (e.g., non-mirror frames). To this end, in this paper, we propose ZOOM, the first method to learn robust mirror representations from extremely-weak annotations of per-frame ZerO-One Mirror indicators in videos. The key insight of ZOOM is to model the similarity and contrast (between mirror and non-mirror regions) in temporal variations to locate and segment the mirrors. To this end, we propose a novel fusion strategy to leverage temporal consistency information for mirror localization, and a novel temporal similarity-contrast modeling module for mirror segmentation. We construct a new video mirror dataset for training and evaluation. Experimental results under new and standard metrics show that ZOOM performs favorably against existing fully-supervised mirror detection methods.

Fairness without Demographics through Shared Latent Space-Based Debiasing

Ensuring fairness in machine learning (ML) is crucial, particularly in applications that impact diverse populations. The majority of existing works heavily rely on the availability of protected features like race and gender. However, practical challenges such as privacy concerns and regulatory restrictions often prohibit the use of this data, limiting the scope of traditional fairness research. To address this, we introduce a Shared Latent Space-based Debiasing (SLSD) method that transforms data from both the target domain, which lacks protected features, and a separate source domain, which contains these features, into correlated latent representations. This allows for joint training of a cross-domain protected group estimator on the representations. We then debias the downstream ML model with an adversarial learning technique that leverages the group estimator. We also present a relaxed variant of SLSD, the R-SLSD, that occasionally accesses a small subset of protected features from the target domain during its training phase. Our extensive experiments on benchmark datasets demonstrate that our methods consistently outperform existing state-of-the-art models in standard group fairness metrics.

Stratified GNN Explanations through Sufficient Expansion

Explaining the decisions made by Graph Neural Networks (GNNs) is vital for establishing trust and ensuring fairness in critical applications such as medicine and science. The prevalence of hierarchical structure in real-world graphs/networks raises an important question on GNN interpretability: "On each level of the graph structure, which specific fraction imposes the highest influence over the prediction?" Currently, the prevailing two categories of methods are incapable of achieving multi-level GNN explanation due to their flat or motif-centric nature. In this work, we formulate the problem of learning multi-level explanations out of GNN models and introduce a stratified explainer module, namely STFExplainer, that utilizes the concept of sufficient expansion to generate explanations on each stratum. Specifically, we learn a higher-level subgraph generator by leveraging both hierarchical structure and GNN-encoded input features. Experiment results on both synthetic and real-world datasets demonstrate the superiority of our stratified explainer on standard interpretability tasks and metrics such as fidelity and explanation recall, with an average improvement of 11% and 8% over the best alternative on each data type. The case study on material domains also confirms the value of our approach through detected multi-level graph patterns accurately reconstructing the knowledge-based ground truth.