Characteristic Attributes Research Articles

Diverse Explanations From Data-Driven and Domain-Driven Perspectives in the Physical Sciences

Abstract Machine learning methods have been remarkably successful in material science, providing novel scientific insights, guiding future laboratory experiments, and accelerating materials discovery. Despite the promising performance of these models, understanding the decisions they make is also essential to ensure the scientific value of their outcomes. However, there is a recent and ongoing debate about the diversity of explanations, which potentially leads to scientific inconsistency. This Perspective explores the sources and implications of these diverse explanations in ML applications for physical sciences. Through three case studies in materials science and molecular property prediction, we examine how different models, explanation methods, levels of feature attribution, and stakeholder needs can result in varying interpretations of ML outputs. 
Our analysis underscores the importance of considering multiple perspectives when interpreting ML models in scientific contexts and highlights the critical need for scientists to maintain control over the interpretation process, balancing data-driven insights with domain expertise to meet specific scientific needs. By fostering a comprehensive understanding of these inconsistencies, we aim to contribute to the responsible integration of eXplainable Artificial Intelligence (XAI) into physical sciences and improve the trustworthiness of ML applications in scientific discovery.

Limitations of Feature Attribution in Long Text Classification of Standards

Managing complex AI systems requires insight into a model's decision-making processes. Understanding how these systems arrive at their conclusions is essential for ensuring reliability. In the field of explainable natural language processing, many approaches have been developed and evaluated. However, experimental analysis of explainability for text classification has been largely constrained to short text and binary classification. In this applied work, we study explainability for a real-world task where the goal is to assess the technological suitability of standards. This prototypical use case is characterized by large documents, technical language, and a multi-label setting, making it a complex modeling challenge. We provide an analysis of approx. 1000 documents with human-annotated evidence. We then present experimental results with two explanation methods evaluating plausibility and runtime of explanations. We find that the average runtime for explanation generation is at least 5 minutes and that the model explanations do not overlap with the ground truth. These findings reveal limitations of current explanation methods. In a detailed discussion, we identify possible reasons and how to address them on three different dimensions: task, model and explanation method. We conclude with risks and recommendations for the use of feature attribution methods in similar settings.

Feature attributions for water–solubility predictions obtained via artificial intelligence methods and chemists

Abstract The field of explainable artificial intelligence (AI) has garnered significant research interest. In particular, “feature attribution” in the field of chemistry has been focused upon. However, studies on comparisons of the relationship between AI- and human-based feature attributions when predicting the same outcome are scarce. Thus, this study aimed to investigate this relationship by comparing machine-learning-based feature attributions (graph neural networks and integrated gradients) with those of chemists (Hansch–Fujita method) considering the case of predicting water–solubility. The AI-based attributions were found to be similar to those of chemists despite their distinct origins.

Exploring a Data‐Driven Approach to Identify Regions of Change Associated With Future Climate Scenarios

AbstractA key consideration for evaluating climate projections is uncertainty in future radiative forcing scenarios. Although it is straightforward to monitor greenhouse gas concentrations and compare observations with specified climate scenarios, it remains less obvious how to detect and attribute regional pattern changes with plausible future mitigation scenarios. Here we introduce a machine learning approach for linking patterns of climate change with radiative forcing scenarios and use a feature attribution method to understand how these linkages are made. We train a neural network using output from the SPEAR Large Ensemble to classify whether temperature or precipitation maps are most likely to originate from one of several potential radiative forcing scenarios. Despite substantial atmospheric internal variability, the neural network learns to identify “fingerprint” patterns, including significant localized regions of change, that associate specific patterns of climate change with radiative forcing scenarios in each year of the simulations. We illustrate this using output from additional ensembles with sharp reductions in future greenhouse gases and highlight specific regions (in this example, the subpolar North Atlantic and Central Africa) that are critical for associating the new simulations with changes in radiative forcing scenarios. Overall, this framework suggests that explainable machine learning could provide one strategy for detecting a regional climate response to future mitigation efforts.

Research on similarity bias in dual objective visual search based on nuclear power human-machine interface icons

Using icons from nuclear power interface as the research object, this study explored how icon similarity affected the performance of dual objective visual search. Firstly, the process of generating similarity bias was described from the perspective of human cognitive processing. The feature attributes of nuclear power icons were extracted, and then associated and mapped with similarity bias attributes. Secondly, a total of 16 instruction icons, device icons, and component icons were selected to propose icon coding logic for different similarity dimensions, and similarity experimental materials were designed. Finally, a dual objective search experiment with a 4 × 4 matrix was conducted to explore the impact of graph similarity on search performance and to determine the priority of perceptual similarity, semantic similarity, and memory similarity. High-level (H) similarity between the two targets resulted in superior visual search performance (p = 0.01 between the response times of similarity high and medium/low). Improving experiential familiarity enhanced search performance in cases of low-level (L) (p = 0.021) and medium-level (M) (p ≤ 0.009) icon similarity, but had no significant impact on search performance in cases of high-level (H) similarity (p ≥ 0.269). Compared to semantic similarity, enhancing perceptual similarity was more likely to improve search performance(p = 0.024).

Clarity in complexity: how aggregating explanations resolves the disagreement problem

The Rashômon Effect, applied in Explainable Machine Learning, refers to the disagreement between the explanations provided by various attribution explainers and to the dissimilarity across multiple explanations generated by a particular explainer for a single instance from the dataset (differences between feature importances and their associated signs and ranks), an undesirable outcome especially in sensitive domains such as healthcare or finance. We propose a method inspired from textual-case based reasoning for aligning explanations from various explainers in order to resolve the disagreement and dissimilarity problems. We iteratively generated a number of 100 explanations for each instance from six popular datasets, using three prevalent feature attribution explainers: LIME, Anchors and SHAP (with the variations Tree SHAP and Kernel SHAP) and consequently applied a global cluster-based aggregation strategy that quantifies alignment and reveals similarities and associations between explanations. We evaluated our method by weighting the \\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\:k$$\\end{document}-NN algorithm with agreed feature overlap explanation weights and compared it to a non-weighted \\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\:k$$\\end{document}-NN predictor, having as task binary classification. Also, we compared the results of the weighted \\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\:k$$\\end{document}-NN algorithm using aggregated feature overlap explanation weights to the weighted \\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\:k$$\\end{document}-NN algorithm using weights produced by a single explanation method (either LIME, SHAP or Anchors). Our global alignment method benefited the most from a hybridization with feature importance scores (information gain), that was essential for acquiring a more accurate estimate of disagreement, for enabling explainers to reach a consensus across multiple explanations and for supporting effective model learning through improved classification performance.

Residual behaviour and damage assessment of UHPC-filled double-skin steel tubular columns after lateral impact

Ultra-high performance concrete (UHPC)-filled double-skin steel tubular (DST) column has great potential to be used in the protective structures. Although its lateral impact behaviour has been well understood, the residual behaviour after lateral impact remains unexplored. As a result, this study extensively investigated the residual behaviour and damage assessment of UHPC-filled DST columns after lateral impact. Firstly, a set of six DST columns were designed and tested under lateral impact, followed by static axial compression. In addition, two intact columns were subjected to static axial compression for comparative analysis. Secondly, the refined finite element models were developed and validated using the current test data, and the impact resistant mechanism of UHPC-filled DST columns with different impact locations was analysed. Thirdly, the suitability of different damage indexes for the damage assessment of impacted UHPC-filled DST columns was evaluated. Two damage indexes, the ratio of mid-height deflection to column height (R1), and the ratio of local deflection to the column diameter (R2), were proposed for the DST columns. Finally, two types of machine learning-based models were developed to predict the impact damage of UHPC-filled DST columns. The prediction models were interpreted locally and globally using the additive feature attribution method Shapley Additive Explanation (SHAP). The machine learning-based prediction models can rapidly evaluate the damage extent of impacted UHPC-filled DST column, which hold great significance for the selection of strengthening and retrofitting schemes.

A New Computer-Aided Diagnosis System for Breast Cancer Detection from Thermograms Using Metaheuristic Algorithms and Explainable AI

Advances in the early detection of breast cancer and treatment improvements have significantly increased survival rates. Traditional screening methods, including mammography, MRI, ultrasound, and biopsies, while effective, often come with high costs and risks. Recently, thermal imaging has gained attention due to its minimal risks compared to mammography, although it is not widely adopted as a primary detection tool since it depends on identifying skin temperature changes and lesions. The advent of machine learning (ML) and deep learning (DL) has enhanced the effectiveness of breast cancer detection and diagnosis using this technology. In this study, a novel interpretable computer aided diagnosis (CAD) system for breast cancer detection is proposed, leveraging Explainable Artificial Intelligence (XAI) throughout its various phases. To achieve these goals, we proposed a new multi-objective optimization approach named the Hybrid Particle Swarm Optimization algorithm (HPSO) and Hybrid Spider Monkey Optimization algorithm (HSMO). These algorithms simultaneously combined the continuous and binary representations of PSO and SMO to effectively manage trade-offs between accuracy, feature selection, and hyperparameter tuning. We evaluated several CAD models and investigated the impact of handcrafted methods such as Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), Gabor Filters, and Edge Detection. We further shed light on the effect of feature selection and optimization on feature attribution and model decision-making processes using the SHapley Additive exPlanations (SHAP) framework, with a particular emphasis on cancer classification using the DMR-IR dataset. The results of our experiments demonstrate in all trials that the performance of the model is improved. With HSMO, our models achieved an accuracy of 98.27% and F1-score of 98.15% while selecting only 25.78% of the HOG features. This approach not only boosts the performance of CAD models but also ensures comprehensive interpretability. This method emerges as a promising and transparent tool for early breast cancer diagnosis.

Improved transfer learning two-branch convolutional neural network image dehazing

In view of the existing image dehazing algorithms, there are problems of incomplete dehazing and image color distortion. A dehazing network combining Transfer learning sub-net and Residual attention sub-net is proposed. First, the pretrained model of the transfer learning subnet is adopted to enhance the feature attributes of the samples. Second, the struct ure of the dual-branch network is constructed, and the residual attention sub-network is used to assist the transfer learning subnetwork to train the parameters of the network model. Finally, the method of tail ensemble learning is used to fuse the features of the dual network to obtain the model parameters of the dehazed image, so as to complete the image restoration task. The experimental results show that the algorithm proposed in the paper improves the PSNR index by 1.87 dB and 4.22 dB on the RESIDE data set and the O-HAZE data set respectively, and the SSIM index on the O-HAZE data set by 6.7%.

How can geostatistics help us understand deep learning? An exploratory study in SAR-based aircraft detection

Deep Neural Networks (DNNs) have garnered significant attention across various research domains due to their impressive performance, particularly Convolutional Neural Networks (CNNs), known for their exceptional accuracy in image processing tasks. However, the opaque nature of DNNs has raised concerns about their trustworthiness, as users often cannot understand how the model arrives at its predictions or decisions. This lack of transparency is particularly problematic in critical fields such as healthcare, finance, and law, where the stakes are high. Consequently, there has been a surge in the development of explanation methods for DNNs. Typically, the effectiveness of these methods is assessed subjectively via human observation on the heatmaps or attribution maps generated by eXplanation AI (XAI) methods. In this paper, a novel GeoStatistics Explainable Artificial Intelligence (GSEAI) framework is proposed, which integrates spatial pattern analysis from Geostatistics with XAI algorithms to assess and compare XAI understandability. Global and local Moran’s I indices, commonly used to assess the spatial autocorrelation of geographic data, assist in comprehending the spatial distribution patterns of attribution maps produced by the XAI method, through measuring the levels of aggregation or dispersion. Interpreting and analyzing attribution maps by Moran’s I scattergram and LISA clustering maps provide an accurate global objective quantitative assessment of the spatial distribution of feature attribution and achieves a more understandable local interpretation. In this paper, we conduct experiments on aircraft detection in SAR images based on the widely used YOLOv5 network, and evaluate four mainstream XAI methods quantitatively and qualitatively. By using GSEAI to perform explanation analysis of the given DNN, we could gain more insights about the behavior of the network, to enhance the trustworthiness of DNN applications. To the best of our knowledge, this is the first time XAI has been integrated with geostatistical algorithms in SAR domain knowledge, which expands the analytical approaches of XAI and also promotes the development of XAI within SAR image analytics.

Dynamic Multi-Factor Correlation Analysis for Prediction of Provincial Carbon Emissions in China’s Bohai Rim Region

This study presents a dynamic multi-factor correlation analysis method designed to predict provincial carbon dioxide emissions (CDE) within China’s Bohai Rim region, including Tianjin, Hebei, Shandong, and Liaoning. By employing the sliding window technique, dynamic correlation curves are computed between various influencing factors and CDE at different time intervals, thereby facilitating the identification of key feature attributes. A novel metric, the Consistency Index of Influencing Factors (CIIF), is introduced to evaluate the consistency of these factors across regions. Furthermore, the Accurate Predictive Capability Indicator (APCI) is defined to measure the impact of different feature categories on the prediction accuracy. The findings reveal that models relying on a single influencing factor exhibit limited accuracy, whereas combining multiple factors with diverse correlation features significantly improves the prediction accuracy. This study introduces a refined analytical framework and a comprehensive indicator system for CDE prediction. It enhances the understanding of the complex factors that influence CDE and provides a scientific rationale for implementing effective emission reduction strategies.

Filter transfer learning algorithm for nonlinear systems modeling with heterogeneous features

Transfer learning can handle the domain adaptation of different feature spaces in nonlinear systems. Most existing studies only focus on common features between heterogeneous scenes rather than specific features that account for latent similarities between them. To deal with this problem, a filter transfer learning algorithm for nonlinear systems modeling with heterogeneous features is proposed. First, nonlinear mapping is constructed to learn the potential relationships between different feature attributes, including common features and specific features. Then, source knowledge from different domains can be obtained in the form of process parameters according to the mapping relationship. Second, a hierarchical filter framework is presented to reconstruct source knowledge in different transfer phases. In the pre-transfer phase, a knowledge filter is designed to increase the diversity of knowledge through selection, crossover, and mutation operations. In the post-transfer phase, a guided filter is established to achieve a coupling balance between source knowledge and target domain by using target samples as guidance information. Third, a dynamic parameter learning strategy is given to promote the learning performance of heterogeneous tasks. Finally, the effectiveness of innovations and the superiority of this proposed algorithm are verified by the experimental results in nonlinear systems with heterogeneous features.

A review and benchmark of feature importance methods for neural networks

Feature attribution methods (AMs) are a simple means to provide explanations for the predictions of black-box models such as neural networks. Due to their conceptual differences, the numerous different methods, however, yield ambiguous explanations. While this allows for obtaining different insights into the model, it also complicates the decision regarding which method to adopt. This article summarizes the current state of the art regarding AMs, which includes the requirements and desiderata of the methods themselves as well as the properties of their explanations. Based on a survey of existing methods, a representative subset consisting of the δ-sensitivity index, permutation feature importance, variance-based feature importance in artificial neural networks and DeepSHAP, is described in greater detail and, for the first time, benchmarked in a regression context. Specifically for this purpose, a new verification strategy for model-specific AMs is proposed. As expected, the explanations’ agreement with the intuition and among each other clearly depends on the AMs’ properties. This has two implications. First, careful reasoning about the selection of an AM is required. Second, it is recommended to apply multiple AMs and combine their insights in order to reduce the model’s opacity even further.

Radar-Derived Convective Features and Atmospheric Environments of Hail-Producing Storms over the Contiguous United States

Abstract Hail smaller than 0.75 in. is known to cause economic impacts yet remains understudied due to report biases towards recording larger hail sizes (≤1 in.). In this study, we assembled ground hail reports during 2017–22 from the National Centers for Environmental Information (NCEI) Storm Data, Community Collaborative Rain, Hail and Snow Network (CoCoRaHS), and Meteorological Phenomena Identification Near the Ground (mPING) databases. Then, these reports are collocated with the attributes of radar-derived convective features from the Multi-Radar Multi-Sensor (MRMS) system and fifth major global reanalysis produced by European Centre for Medium-Range Weather Forecasts (ERA5) atmospheric vertical profiles to construct a dataset describing properties of a full spectrum of hailstorms. The characteristics of radar reflectivity and atmospheric profiles are examined for hail of different sizes reported within selected regions over the contiguous United States (CONUS). In addition to the seasonal and diurnal variations, the morphology of convective features shows apparent regional differences from west to east in CONUS. The maximum expected size of hail (MESH) performance against reported hail sizes shows underestimation of hail with significant sizes and overestimation of small hail sizes. ERA5 vertical atmospheric profiles are explored to form relationships between storm environment and hail sizes. In addition to the relationships between wind shear and hail sizes, the roles of low-level relative humidity and freezing level height in regard to hail melting are discussed.

MolAnchor method for explaining compound predictions based on substructures

In medicinal chemistry, the impact of machine learning remains limited if predictions are not understood, which often precludes experimental follow-up. Therefore, chemically intuitive approaches that aid in model understanding and interpretation at the molecular level of detail are sought after. While feature attribution methods quantifying feature importance for model decisions are widely used in many areas, they must typically be combined with visualization techniques, if possible, to render the results accessible from a chemical viewpoint. On the other hand, there are approaches such as counterfactuals that yield closely related chemical structures with different prediction outcomes, providing direct access to structural features that critically influence model decisions. Herein, we introduce another approach designed to rationalize chemical predictions based on molecular structure. Therefore, we adapt principles underlying the anchor concept from explainable artificial intelligence (XAI) and alter them for molecular machine learning. The resulting method, termed MolAnchor, systematically identifies substructures in test compounds that determine property predictions, thus ensuring chemical interpretability. The MolAnchor methodology is made freely available to the medicinal chemistry community as a part of our study.

An interpretable tinnitus prediction framework using gap-prepulse inhibition in auditory late response and electroencephalogram.

Tinnitus is a neuropathological condition that results in mild buzzing or ringing of the ears without an external sound source. Current tinnitus diagnostic methods often rely on subjective assessment and require intricate medical examinations. This study aimed to propose an interpretable tinnitus diagnostic framework using auditory late response (ALR) and electroencephalogram (EEG), inspired by the gap-prepulse inhibition (GPI) paradigm. We collected spontaneous EEG and ALR data from 44 patients with tinnitus and 47 hearing loss-matched controls using specialized hardware to capture responses to sound stimuli with embedded gaps. In this cohort study of tinnitus and control groups, we examined EEG spectral and ALR features of N-P complexes, comparing the responses to gap durations of 50 and 20 ms alongside no-gap conditions. To this end, we developed an interpretable tinnitus diagnostic model using ALR and EEG metrics, boosting machine learning architecture, and explainable feature attribution approaches. Our proposed model achieved 90 % accuracy in identifying tinnitus, with an area under the performance curve of 0.89. The explainable artificial intelligence approaches have revealed gap-embedded ALR features such as the GPI ratio of N1-P2 and EEG spectral ratio, which can serve as diagnostic metrics for tinnitus. Our method successfully provides personalized prediction explanations for tinnitus diagnosis using gap-embedded auditory and neurological features. Deficits in GPI alongside activity in the EEG alpha-beta ratio offer a promising screening tool for assessing tinnitus risk, aligning with current clinical insights from hearing research.

Visual interpretation of deep learning model in ECG classification: A comprehensive evaluation of feature attribution methods

Feature attribution methods can visually highlight specific input regions containing influential aspects affecting a deep learning model's prediction. Recently, the use of feature attribution methods in electrocardiogram (ECG) classification has been sharply increasing, as they assist clinicians in understanding the model's decision-making process and assessing the model's reliability. However, a careful study to identify suitable methods for ECG datasets has been lacking, leading researchers to select methods without a thorough understanding of their appropriateness. In this work, we conduct a large-scale assessment by considering eleven popular feature attribution methods across five large ECG datasets using a model based on the ResNet-18 architecture. Our experiments include both automatic evaluations and human evaluations. Annotated datasets were utilized for automatic evaluations and three cardiac experts were involved for human evaluations. We found that Guided Grad-CAM, particularly when its absolute values are utilized, achieves the best performance. When Guided Grad-CAM was utilized as the feature attribution method, cardiac experts confirmed that it can identify diagnostically relevant electrophysiological characteristics, although its effectiveness varied across the 17 different diagnoses that we have investigated.

New feature attribution method for explainable aspect-based sentiment classification

Recently, the use of feature attribution methods in explainable artificial intelligence has attracted significant attention. While many proposed methods in this domain involve analyzing the impact on results by filtering out specific features. Although these methods can provide explanations, they often overlook a relevant factor: the deletion of features alters the structural information of the text, resulting in the loss of relations between features, thus leading to explanations that are not comprehensive and appropriate. To address these challenges, this study introduces a novel feature attribution method known as General Explainable Sentiment Classification (GESC). Unlike traditional approaches that delete features, GESC focuses on preserving the inherent relations by substituting the features with synonyms. Furthermore, the GESC method employs the binary tumbleweed algorithm (BTA) to identify the optimal feature combination. Moreover, designers have created an adaptive iterative technique to accommodate texts of varying lengths. The application of GESC to aspect-based sentiment classification demonstrates its practicability. The experimental results indicate that the BTA primarily improves the efficiency of finding the optimal combination of features compared with other algorithms. The GESC method provides more compelling explanations than conventional feature attribution approaches.

Energy bubble entropy guided symplectic geometry mode decomposition for rotating machinery incipient fault feature extraction

Abstract Extracting incipient fault features is a critical aspect of monitoring the rotating machinery operation condition. However, existing methods based on symplectic geometry mode decomposition (SGMD) suffer from limited parameter adaptability and noise robustness. Therefore, this paper proposes an Energy bubble entropy guided SGMD method to extract incipient fault feature. Firstly, the SGMD method is employed to initially separate fault characteristic components from noisy signal. Furthermore, the energy bubble entropy(EbEn) is constructed to evaluate the attributes of incipient feature within the signal, which requires almost no parameter setting with good robustness and computational efficiency. Thirdly, the Empirical Bayes shrinkage method effectively mitigates irrelevant noise and enhances the significance of incipient fault feature. Simulated and experimental signals are employed to substantiate the efficacy of the EbEn guided SGMD method. The comparison analysis with relevant methods exhibits that this method has greater robustness and adaptivity.

Feature radiance fields (FeRF): A multi-level feature fusion method with deep neural network for image synthesis

Neural Radiance Field (NeRF) has brought revolutionary changes to the field of image synthesis with its unique ability to generate highly realistic multi-view consistent images from a neural scene representation. However, current NeRF-based methods still largely depend on multiple, precisely posed images, especially for complex or dynamic scenes, limiting their versatility. Furthermore, some recent strategies attempt to integrate simple feature extraction networks with volume rendering techniques to reduce multi-view dependence but create blurry outputs, highlighting the need for more sophisticated feature handling to unlock NeRF's full potential. In this paper, we propose an image synthesis method named FeRF, distinguished by its capacity to perform comprehensive feature extraction on individual unposed images and facilitate feature fusion at any stage. Additionally, we present an "elaborated-feature generation network" (EGN) composed of four modules, which is configured with two advanced feature extraction modules aimed at precisely refining and processing subtle, complex visual features from a single image. Given that the core objective of FeRF is the precise capture and processing of intricate features from the input images, we innovatively incorporated precisely designed attention mechanisms into the network architecture to optimize and highlight the importance of key feature attributes, thereby effectively enhancing their contribution to subsequent volume rendering processes. Extensive experimentation validates the outstanding qualitative and quantitative performance of our proposed network structure. In comparison to current image feature-based generalized image synthesis methods, it achieves superior reconstruction quality and level of detail.