Interpretable Machine Learning Methods Research Articles

Deducing neighborhoods of classes from a fitted model

AbstractIn this article, a new kind of interpretable machine learning method is presented, which can help to understand the partition of the feature space into predicted classes in a classification model using quantile shifts, and this way make the underlying statistical or machine learning model more trustworthy. Basically, real data points (or specific points of interest) are used and the changes of the prediction after slightly raising or decreasing specific features are observed. By comparing the predictions before and after the shifts, under certain conditions the observed changes in the predictions can be interpreted as neighborhoods of the classes with regard to the shifted features. Chord diagrams are used to visualize the observed changes. For illustration, this quantile shift method (QSM) is applied to an artificial example with medical labels and a real data example.

Unraveling the relationship between coastal landscapes and sentiments: An integrated approach based on social media data and interpretable machine learning methods

AbstractCoastal landscapes exert a significant impact on the human sentimental perceptions and physical and mental well‐being of people. However, little is known about explicitly linking between the landscape characteristics and people's sentimental preferences expressed in social media data. The main objective of this study was to explore the nonlinear and interaction effects of key factors that influenced sentiments in the coastal areas of Hong Kong, considering both subjective landscape preferences and objective landscape patterns. We quantified users' sentiment polarity based on the crowdsourcing textual data of Flickr. To study users' subjective landscape preferences, we computed various visual landscape objects' proportion in images. Meanwhile, eight user clusters and nine image clusters were detected by the identified visual object labels. We quantified objective landscape patterns considering the land use pattens and the availability of public service facilities. Finally, we utilized an interpretable classification model to analyze the factors that may affect sentiments and their interplay interactions. We found that ecotourism‐related clusters exhibited the most positive sentiment. The proportion of floor and sky pixels in images exhibits the highest global relative importance when predicting sentiments. This study extends a new insight on the relationship between landscape characteristics and sentiments from both subjective and objective perspectives based on social media data and interpretable machine learning methods. This research may help decision‐makers in designing landscapes that aptly satisfy to the needs of the public and promote sustainable management of the coastal environment.

Bayesian network models identify cooperative GPCR:G protein interactions that contribute to G protein coupling

Cooperative interactions in protein-protein interfaces demonstrate the interdependency or the linked network- like behavior and their effect on the coupling of proteins. Cooperative interactions also could cause ripple or allosteric effects at a distance in protein-protein interfaces. Although they are critically important in protein-protein interfaces, it is challenging to determine which amino acid pair interactions are cooperative. In this work we have used Bayesian network modeling, an interpretable machine learning method, combined with molecular dynamics trajectories to identify the residue pairs that show high cooperativity and their allosteric effect in the interface of G protein coupled receptor (GPCR) complexes with Gα subunits. Our results reveal six GPCR:Gα contacts that are common to the different Gα subtypes and show strong cooperativity in the formation of interface. Both the C-terminus helix5 and the core of the G protein are codependent entities and play an important role in GPCR coupling. We show that a promiscuous GPCR coupling to different Gα subtypes, makes all the GPCR:Gα contacts that are specific to each Gα subtype (Gαs, Gαi and Gαq). This work underscores the potential of data-driven Bayesian network modeling in elucidating the intricate dependencies and selectivity determinants in GPCR:G protein complexes, offering valuable insights into the dynamic nature of these essential cellular signaling components.

Machine-learning analysis reveals an important role for negative selection in shaping cancer aneuploidy landscapes

BackgroundAneuploidy, an abnormal number of chromosomes within a cell, is a hallmark of cancer. Patterns of aneuploidy differ across cancers, yet are similar in cancers affecting closely related tissues. The selection pressures underlying aneuploidy patterns are not fully understood, hindering our understanding of cancer development and progression.ResultsHere, we apply interpretable machine learning methods to study tissue-selective aneuploidy patterns. We define 20 types of features corresponding to genomic attributes of chromosome-arms, normal tissues, primary tumors, and cancer cell lines (CCLs), and use them to model gains and losses of chromosome arms in 24 cancer types. To reveal the factors that shape the tissue-specific cancer aneuploidy landscapes, we interpret the machine learning models by estimating the relative contribution of each feature to the models. While confirming known drivers of positive selection, our quantitative analysis highlights the importance of negative selection for shaping aneuploidy landscapes. This is exemplified by tumor suppressor gene density being a better predictor of gain patterns than oncogene density, and vice versa for loss patterns. We also identify the importance of tissue-selective features and demonstrate them experimentally, revealing KLF5 as an important driver for chr13q gain in colon cancer. Further supporting an important role for negative selection in shaping the aneuploidy landscapes, we find compensation by paralogs to be among the top predictors of chromosome arm loss prevalence and demonstrate this relationship for one paralog interaction. Similar factors shape aneuploidy patterns in human CCLs, demonstrating their relevance for aneuploidy research.ConclusionsOur quantitative, interpretable machine learning models improve the understanding of the genomic properties that shape cancer aneuploidy landscapes.

Diagnosing ozone–NOx–VOC–aerosol sensitivity and uncovering causes of urban–nonurban discrepancies in Shandong, China, using transformer-based estimations

Abstract. Narrowing surface ozone disparities between urban and nonurban areas escalate health risks in densely populated urban zones. A comprehensive understanding of the impact of ozone photochemistry on this transition remains constrained by current knowledge of aerosol effects and the availability of surface monitoring. Here we reconstructed spatiotemporal gapless air quality concentrations using a novel transformer deep learning (DL) framework capable of perceiving spatiotemporal dynamics to analyze ozone urban–nonurban differences. Subsequently, the photochemical effect on these discrepancies was analyzed by elucidating shifts in ozone regimes inferred from an interpretable machine learning method. The evaluations of the model exhibited an average out-of-sample cross-validation coefficient of determination of 0.96, 0.92, and 0.95 for ozone, nitrogen dioxide, and fine particulate matter (PM2.5), respectively. The ozone sensitivity in nonurban areas, dominated by a nitrogen-oxide-limited (NOx-limited) regime, was observed to shift towards increased sensitivity to volatile organic compounds (VOCs) when extended to urban areas. A third “aerosol-inhibited” regime was identified in the Jiaodong Peninsula, where the uptake of hydroperoxyl radicals onto aerosols suppressed ozone production under low NOx levels during summertime. The reduction of PM2.5 could increase the sensitivity of ozone to VOCs, necessitating more stringent VOC emission abatement for urban ozone mitigation. In 2020, urban ozone levels in Shandong surpassed those in nonurban areas, primarily due to a more pronounced decrease in the latter resulting from stronger aerosol suppression effects and less reduction in PM2.5. This case study demonstrates the critical need for advanced spatially resolved models and interpretable analysis in tackling ozone pollution challenges.

The travel pattern difference in dockless micro-mobility: Shared e-bikes versus shared bikes

To facilitate the tailoring of dockless bike-sharing and electric bike (e-bike) sharing services and assist in formulating effective regulations, this study aims to unravel the spatio-temporal travel patterns specific to e-bike-sharing and bike-sharing systems, utilising interpretable machine learning methods and a large-scale trip-level dataset in Kunming, China. The results show that shared bikes and e-bikes exhibit overall similarities and subtle differences in many aspects, such as trip attributes and spatial distribution. Additionally, both shared bikes and shared e-bikes have three basic temporal patterns for commuting and recreational purposes. Regarding the differences, e-bike sharing networks are more dispersed and bigger, and bike sharing tends to form densely connected clusters of flow, exhibiting a local concentration of activity. Besides, the commuting activities within e-bike sharing systems exhibit two patterns: direct travel to the destination and integration with public transit. In contrast, shared bikes predominantly rely on public transit transfers for commuting purposes.

Decoding the writing styles of disciplines: A large-scale quantitative analysis

Disciplinary writing style stems from the practice of science, reflecting the scientific culture. This study aims to explore the differences and evolution of scientific writing styles from the perspective of disciplines. A large-scale quantitative analysis was conducted over 14 million abstracts from the Microsoft Academic Graph (MAG) database across eight soft and hard disciplines. Represented by a comprehensive set of 14 symbolic, lexical, syntactic, structural, and readability features, the evolution of disciplinary writing styles was analyzed over 30 years. Interpretable machine learning methods were performed to test the discernibility of writing styles across disciplines and disclose their linguistic differences. Our findings reveal the linguistic features of soft disciplines (Art, Philosophy, and Sociology) and Mathematics generally keep stabilized, and a general trend of increasing linguistic complexity was observed for Biology, Chemistry, Computer Science, and Psychology. The good performance of the pairwise writing style classifiers indicates a well discriminability of the writing styles between disciplines. A correlation between the performance of classifiers and the distance between disciplines was identified. The feature contribution analysis using SHapley Additive exPlanations (SHAP) and Kendall's Tau rank correlation revealed the detailed commonalities and disparities in disciplines’ linguistic features. This study provides profound insights into the understanding of scientific writing and norms, which further helps develop useful tools for academic text analysis, foster interdisciplinary communication, and assist educators to construct discipline-specific writing guidance.

Teaching freight mode choice models new tricks using interpretable machine learning methods

Teaching freight mode choice models new tricks using interpretable machine learning methods

Physics-informed interpretable machine learning method for DOC monitoring in peripheral milling

Physics-informed interpretable machine learning method for DOC monitoring in peripheral milling

Predicting and understanding the ductility of BCC high entropy alloys via knowledge-integrated machine learning

Despite the fact that body-centered cubic high entropy alloys (BCC HEAs) exhibit exceptional strengthening mechanisms over a wide temperature range, their application is often hindered by poor room temperature ductility. Consequently, the accurate prediction of BCC HEAs′ ductility becomes critical in alloy design but is challenged by the intricate nature of these alloy systems. In this study, a comprehensive machine learning (ML) strategy with integrated material knowledge is proposed to address this problem. By referring to plasticity theories, thermodynamics as well as atomic properties, we develop a set of cost-effective material descriptors with various mathematical forms such as averaging, mismatching, maximum-minimum ranging and averaging as binary pairs for HEAs. These descriptors are proved to effectively incorporate material knowledge into ML modelling. They also offer potential alternatives to those obtained from costly first principles calculations. Classification and regression tasks are conducted to predict the ductility of BCC HEAs, and Gradient boosting algorithm exhibiting the best performance for both tasks is eventually selected. The established optimized models achieve high predictive accuracies exceeding 0.85. Through feature engineering, five key features related to valence electron concentration (VEC) and atomic size are screened out. These features lay the foundation for explicit formulae developed via Symbolic Classification, enabling the straightforward numerical prediction of BCC HEAs′ ductility. These key features are further analyzed by an interpretable ML method and compared against experimental and computational plasticity factors, elucidating their influences on various plasticity mechanisms. An in-depth exploration of why the proposed ML models outperform conventional plasticity criteria for HEAs is also presented. This study highlights an effective knowledge-integrated ML approach in predicting material properties resulting from complex mechanisms, with a particular focus on HEAs that often exhibit sparse data.

Learning from machine learning: the case of band-gap directness in semiconductors

Having a direct or indirect band gap can influence the potential applications of a semiconductor, for indirect band gap materials are usually not suitable for optoelectronic devices. Even though this is a fundamental property of semiconducting materials, discussed in textbooks, no unified theory exists to explain why a material has a direct or indirect band gap. Here we used an interpretable machine learning model, the multiVariate dAta eXplanation (VAX) method, to gather information from a dataset of materials extracted from the Materials Project. The dataset contains more than 10000 entries, and atomic properties such as the number of electrons, electronic affinity and orbital energies were used as features to build random forest models that successfully explain the directness of the band gaps. Our results indicate that symmetry is an important feature that dictates the target property, which is the reason why our analysis is made based on sub-groups with similar structures. These sub-groups include materials with zincblende, rocksalt, wurtzite, and perovskite structures. Besides the symmetry of the materials, the existence or not of d bands and the relative energy of atomic orbitals were found to be important in defining whether a material’s band gap is direct or indirect. In conclusion, interpretable machine learning methods such as VAX can be useful in obtaining physical interpretation from materials databases.

Exploring the primary magnetic parameters affecting chemical fractions of heavy metal(loid)s in lake sediment through an interpretable workflow

The magnetic properties of lake sediments account for close relationships with heavy metal(loid)s (HMs), but little is known about their relationships with chemical fractions (CFs) of HMs. Establishing an effective workflow to predict HMs risk among various machine learning (ML) methods in conjunction with magnetic measurement remains challenging. This study evaluated the simulation efficiency of nine ML methods in predicting the risk assessment code (RAC) and ratio of the secondary and primary phases (RSP) of HMs with magnetic parameters in sediment cores of a shallow lake. The sediment cores were collected and sliced, and the total amount and CFs of HMs, as well as magnetic parameters, were determined. Support vector machine (SVM) outperformed other models, as evidenced by coefficient of determination (R2) > 0.8. Interpretable machine learning (IML) methods were employed to identify key indicators of RAC and RSP among the magnetic parameters. Values of χARM, HIRM, χARM/χ, and χARM/SIRM of sediments ranging in 220–500 × 10−8 m3/kg, 30–40 × 10−5Am2/kg, 15–25, and 0.5–1, respectively, indicated the potential ecological risks of Cd, Hg, and Sb. This study offers new perspectives on the risk assessment of HMs in lake sediments by combining magnetic measurement with IML workflow.

Using an Interpretable Amino Acid-Based Machine Learning Method to Enhance the Diagnosis of Major Depressive Disorder.

Background: Major depressive disorder (MDD) is a leading cause of disability worldwide. At present, however, there are no established biomarkers that have been validated for diagnosing and treating MDD. This study sought to assess the diagnostic and predictive potential of the differences in serum amino acid concentration levels between MDD patients and healthy controls (HCs), integrating them into interpretable machine learning models. Methods: In total, 70 MDD patients and 70 HCs matched in age, gender, and ethnicity were recruited for the study. Serum amino acid profiling was conducted by means of chromatography-mass spectrometry. A total of 21 metabolites were analysed, with 17 from a preset amino acid panel and the remaining 4 from a preset kynurenine panel. Logistic regression was applied to differentiate MDD patients from HCs. Results: The best-performing model utilised both feature selection and hyperparameter optimisation and yielded a moderate area under the receiver operating curve (AUC) classification value of 0.76 on the testing data. The top five metabolites identified as potential biomarkers for MDD were 3-hydroxy-kynurenine, valine, kynurenine, glutamic acid, and xanthurenic acid. Conclusions: Our study highlights the potential of using an interpretable machine learning analysis model based on amino acids to aid and increase the diagnostic accuracy of MDD in clinical practice.

Responsible machine learning for United States Air Force pilot candidate selection

The United States Air Force (USAF) continues to be plagued by a chronic pilot shortage, one that could be exacerbated by an accompanying shortfall in the commercial airlines. As a result, efforts have increased to alleviate this shortage by finding methods to reduce pilot training attrition. We contribute to these efforts by setting forth a decision support system (DSS) for pilot candidate selection using modern machine learning techniques. In view of the recent Responsible Artificial Intelligence Strategy published by the United States Department of Defense, this research leverages interpretable and explainable machine learning methods to create traceable and equitable models that may be responsibly and reliably governed. These models are used to regress candidates’ average merit assignment selection system scores based on information available for selection and prior to training. More specifically, using data provided by the USAF from 2010 to 2018, this paper develops and analyzes multiple interpretable models based on Gaussian Bayesian networks, as well as multiple black-box models rendered explainable by SHAP values and conformal prediction. A preferred pair of interpretable and explainable models is selected and embedded within a DSS for USAF pilot candidate selection boards: the Air Force Pilot Applicant Selection System. The utilization of this DSS is explored, the analyses it enables are discussed, and relevant USAF policymaking issues are examined.

Interpretable Machine Learning To Accelerate the Analysis of Doping Effect on Li/Ni Exchange in Ni-Rich Layered Oxide Cathodes.

In Ni-rich layered oxide cathodes, one effective way to adjust the performance is by introducing dopants to change the degree of Li/Ni exchange. We calculated the formation energy of Li/Ni exchange defects in LiNi0.8Mn0.1X0.1O2 with different doping elements X, using first-principles calculations. We then proposed an interpretable machine learning method combining the Random Forest (RF) model and the Shapley Additive Explanation (SHAP) analysis to accelerate identification of the key factors influencing the formation energy among the complex variables introduced by doping. The valence state of the doping element effectively regulates Li/Ni exchange defects through changing the valence state of Ni and the strength of the superexchange interaction, and COOPSU-SD and MagO were proposed as two indicators to assess superexchange interaction. The volume change also affects the Li/Ni exchange defects, with a larger volume reduction corresponding to fewer Li/Ni exchange defects.

Predicting Hurricane Evacuation Decisions with Interpretable Machine Learning Methods

Facing the escalating effects of climate change, it is critical to improve the prediction and understanding of the hurricane evacuation decisions made by households in order to enhance emergency management. Current studies in this area often have relied on psychology-driven linear models, which frequently exhibited limitations in practice. The present study proposed a novel interpretable machine learning approach to predict household-level evacuation decisions by leveraging easily accessible demographic and resource-related predictors, compared to existing models that mainly rely on psychological factors. An enhanced logistic regression model (that is, an interpretable machine learning approach) was developed for accurate predictions by automatically accounting for nonlinearities and interactions (that is, univariate and bivariate threshold effects). Specifically, nonlinearity and interaction detection were enabled by low-depth decision trees, which offer transparent model structure and robustness. A survey dataset collected in the aftermath of Hurricanes Katrina and Rita, two of the most intense tropical storms of the last two decades, was employed to test the new methodology. The findings show that, when predicting the households’ evacuation decisions, the enhanced logistic regression model outperformed previous linear models in terms of both model fit and predictive capability. This outcome suggests that our proposed methodology could provide a new tool and framework for emergency management authorities to improve the prediction of evacuation traffic demands in a timely and accurate manner.

Built Environment Renewal Strategies Aimed at Improving Metro Station Vitality via the Interpretable Machine Learning Method: A Case Study of Beijing

Understanding the built environment’s impact on metro ridership is essential for developing targeted strategies for built environment renewal. Taking into consideration the limitations of existing studies, such as not proposing targeted strategies, using unified pedestrian catchment areas (PCA), and not determining the model’s accuracy, Beijing was divided into three zones from inside to outside by the distribution pattern of metro stations. Three PCAs were assumed for each zone and a total of 27 PCA combinations. The study compared the accuracy of the Ordinary Least Square (OLS) and several machine learning models under each PCA combination to determine the model to be used in this study and the recommended PCA combination for the three zones. Under the recommended PCA combinations for the three zones, the model with the highest accuracy was used to explore the built environment’s impact on metro ridership. Finally, prioritized stations for renewal were identified based on ridership and the built environment’s impact on metro ridership. The results are as follows: (1) The eXtreme Gradient Boosting (XGBoost) model has a higher accuracy and was appropriate for this study. The recommended PCA combination for the three zones in Beijing was 1000 m_1200 m_1800 m. (2) During the morning peak hours, the density of office and apartment facilities greatly influenced the ridership, with a strong threshold effect and spatial heterogeneity. Our research framework also provides a new way for other cities to determine the scope of Transit-Oriented Development (TOD) and proposes a new decision-making method for improving the vibrancy of metro stations.

Interpretable machine learning for predicting the fate and transport of pentachlorophenol in groundwater

Pentachlorophenol (PCP) is a commonly found recalcitrant and toxic groundwater contaminant that resists degradation, bioaccumulates, and has a potential for long-range environmental transport. Taking proper actions to deal with the pollutant accounting for the life cycle consequences requires a better understanding of its behavior in the subsurface. We recognize the huge potential for enhancing decision-making at contaminated groundwater sites with the arrival of machine learning (ML) techniques in environmental applications. We used ML to enhance the understanding of the dynamics of PCP transport properties in the subsurface, and to determine key hydrochemical and hydrogeological drivers affecting its transport and fate. We demonstrate how this complementary knowledge, provided by data-driven methods, may enable a more targeted planning of monitoring and remediation at two highly contaminated Swedish groundwater sites, where the method was validated. We evaluated 6 interpretable ML methods, 3 linear regressors and 3 non-linear (i.e., tree-based) regressors, to predict PCP concentration in the groundwater. The modeling results indicate that simple linear ML models were found to be useful in the prediction of observations for datasets without any missing values, while tree-based regressors were more suitable for datasets containing missing values. Considering that missing values are common in datasets collected during contaminated site investigations, this could be of significant importance for contaminated site planners and managers, ultimately reducing site investigation and monitoring costs. Furthermore, we interpreted the proposed models using the SHAP (SHapley Additive exPlanations) approach to decipher the importance of different drivers in the prediction and simulation of critical hydrogeochemical variables. Among these, sum of chlorophenols is of highest significance in the analyses. Setting that aside from the model, tetra chlorophenols, dissolved organic carbon, and conductivity found to be of highest importance. Accordingly, ML methods could potentially be used to improve the understanding of groundwater contamination transport dynamics, filling gaps in knowledge that remain when using more sophisticated deterministic modeling approaches.

Machine Learning Interpretability of Outer Radiation Belt Enhancement and Depletion Events

AbstractWe investigate the response of outer radiation belt electron fluxes to different solar wind and geomagnetic indices using an interpretable machine learning method. We reconstruct the electron flux variation during 19 enhancement and 7 depletion events and demonstrate the feature attribution analysis called SHAP (SHapley Additive exPlanations) on the superposed epoch results for the first time. We find that the intensity and duration of the substorm sequence following an initial dropout determine the overall enhancement or depletion of electron fluxes, while the solar wind pressure drives the initial dropout in both types of events. Further statistical results from a data set with 71 events confirm this and show a significant correlation between the resulting flux levels and the average AL index, indicating that the observed “depletion” event can be more accurately described as a “non‐enhancement” event. Our novel SHAP‐Enhanced Superposed Epoch Analysis (SHESEA) method can offer insight in various physical systems.

Multi-feature concatenation and multi-classifier stacking: An interpretable and generalizable machine learning method for MDD discrimination with rsfMRI

Major depressive disorder (MDD) is a serious and heterogeneous psychiatric disorder that needs accurate diagnosis. Resting-state functional MRI (rsfMRI), which captures multiple perspectives on brain structure, function, and connectivity, is increasingly applied in the diagnosis and pathological research of MDD. Different machine learning algorithms are then developed to exploit the rich information in rsfMRI and discriminate MDD patients from normal controls. Despite recent advances reported, the MDD discrimination accuracy has room for further improvement. The generalizability and interpretability of the discrimination method are not sufficiently addressed either. Here, we propose a machine learning method (MFMC) for MDD discrimination by concatenating multiple features and stacking multiple classifiers. MFMC is tested on the REST-meta-MDD data set that contains 2428 subjects collected from 25 different sites. MFMC yields 96.9% MDD discrimination accuracy, demonstrating a significant improvement over existing methods. In addition, the generalizability of MFMC is validated by the good performance when the training and testing subjects are from independent sites. The use of XGBoost as the meta classifier allows us to probe the decision process of MFMC. We identify 13 feature values related to 9 brain regions including the posterior cingulate gyrus, superior frontal gyrus orbital part, and angular gyrus, which contribute most to the classification and also demonstrate significant differences at the group level. The use of these 13 feature values alone can reach 87% of MFMC’s full performance when taking all feature values. These features may serve as clinically useful diagnostic and prognostic biomarkers for MDD in the future.