Partial Dependence Research Articles

Natural TCRs targeting KRASG12V display fine specificity and sensitivity to human solid tumors.

Neoantigens derived from KRASMUT have been described, but the fine antigen specificity of T cell responses directed against these epitopes are poorly understood. Here, we explore KRASMUT immunogenicity and the properties of 4 TCRs specific for KRASG12V restricted to HLA-A3 superfamily of class I alleles. A phase I clinical vaccine trial targeting KRASMUT was conducted. TCRs targeting KRASG12V restricted to HLA-A*03:01 or HLA-A*11:01 were isolated from vaccinated patients or healthy individuals. A comprehensive analysis of TCR antigen specificity, affinity, cross-reactivity, and CD8 coreceptor dependence was performed. TCR lytic activity was evaluated, and target antigen density was determined by quantitative immunopeptidomics. Vaccination against KRASMUT resulted in the priming of CD8+ and CD4+ T cell responses. KRASG12V -specific natural (not affinity-enhanced) TCRs exhibited exquisite specificity to mutated protein with no discernable reactivity against KRASWT. TCR-recognition motifs were determined and used to identify and exclude cross-reactivity to non-cognate peptides derived from the human proteome. Both HLA-A*03:01 and HLA-A*11:01 restricted TCR-redirected CD8+ T cells exhibited potent lytic activity against KRASG12V cancers, while only HLA-A*11:01 restricted TCR-T CD4+ T cells exhibited anti-tumor effector functions consistent with partial co-receptor dependence. All KRASG12V-specific TCRs displayed high sensitivity for antigen as demonstrated by their ability to eliminate tumor cell lines expressing low levels of of peptide/HLA (4.4 to 242) complexes per cell. This study identifies KRASG12V-specific TCRs with high therapeutic potential for the development of TCR-T cell therapies. gov NCT03592888. AACR SU2C / Lustgarten Foundation, Parker Institute for Cancer Immunotherapy, and NIH (R01 CA204261, P01 CA217805, P30 CA016520).

Deciphering the cytotoxicity of micro- and nanoplastics in Caco-2 cells through meta-analysis and machine learning

Plastic pollution, driven by micro- and nanoplastics (MNPs), poses a major environmental threat, exposing humans through various routes. Despite human colorectal adenocarcinoma Caco-2 cells being used as an in vitro model for studying the intestinal epithelium, uncertainties linger about MNPs harming these cells and the factors influencing adverse effects. Addressing this lacuna, our study aimed to elucidate the pivotal MNP parameters influencing cytotoxicity in Caco-2 cells, employing meta-analysis and machine learning techniques for quantitative assessment. Initial scrutiny of 95 publications yielded 17 that met the inclusion criteria, generating a dataset of 320 data points. This dataset underwent meticulous stratification based on polymer type, exposure time, polymer size, MNP concentration, and biological assays utilised. Subsequent dose-response curve analysis revealed moderate correlations for selected subgroups, such as the (3-[4,5-dimethylthiazol-2-yl]-2,5 diphenyl tetrazolium bromide) MTT biological assay and exposure time exceeding 24 h, with coefficient of determination (R2) values of 0.50 (p-value: 0.0065) and 0.60 (p-value: 0.0018) respectively. For the aforementioned two subgroups, the MNP concentrations surpassing 10 μg/mL led to diminished viability of Caco-2 cells. Notably, we observed challenges in employing meta-analysis to navigate this multidimensional MNP dataset. Leveraging a random forest model, we achieved improved predictive performance, with R2 values of 0.79 and a root mean square error (RMSE) of 0.14 for the prediction of the Log Response Ratio on the test set. Model interpretation indicated that size and concentration are the principal drivers influencing Caco-2 cell cytotoxicity. Additionally, the partial dependence plot illustrating the relationship between the size of MNPs and predicted cytotoxicity reveals a complex pattern. Our study provides crucial insights into the health impacts of plastic pollution, informing policymakers for targeted interventions, thus contributing to a comprehensive understanding of its human health consequences.

River algal blooms can be estimated by remote sensing reflectance

Abstract River eutrophication is difficult to diagnose and estimate quantitatively because of its complex degradation mechanism in large river systems. Conventional monitoring and modeling methods are limited to accurately revealing the evolution process and trends of river aquatic organisms. In the present study, based on HJ-1A/1B CCD sensor, combined with genetic algorithm (GA) and regression tree (GART), a remote sensing inversion prediction model was established; the model can estimate algal blooms in the Han River affected by China’s Middle Route of the South-to-North Water Diversion Project (SNWTP). During the outbreak of algal blooms, the near-infrared band reflectance evidently increased between 2009 and 2015, with increasing algal density. The algal density in the downstream of the Han River has a nearly synchronous positive change with the reflectance in the B4 (near-infrared) band and a nearly synchronous reverse change with the B1 (blue) band. B1 and B4 screened by GA reduced redundancy by 14%, leading to a good prediction performance (R 2 = 0.88). According to GART and partial dependence analysis, the B4 band is a crucial characterization factor of algal blooms in the Han River. When the remote sensing band was in the range of B1 ⩾ 0.085 and B4 ⩽ 0.101, the algal density was lower than 0.15 × 107 cells l−1, indicating no algal bloom in the downstream of the Han River. When B4 was >0.103 and B1 ⩽ 0.076, algal density was higher than 1 × 107 cells l−1 and algal blooms were very likely to occur. These findings could provide a scientific reference for diagnosing and predicting large-scale water ecological degradation in similar watersheds.

Analyzing shared e-scooter trip frequency on urban road segments in Austin, TX

The expansion of e-scooter sharing system presents a mix of advantages and challenges to the urban transportation system. This research delves into the frequency of shared e-scooter trips on urban road segments in Austin, TX, leveraging a Random Forest model to dissect the influence of built environment and demographic variables on e-scooter trip frequencies. The model was then interpreted using Shapley Additive Explanations and Partial Dependence Plots. Results indicated that presence of bike lanes, distance to city center, violent crime, walkability, and land use are the most important variables. Notably, high shared e-scooter trip frequency often coincides with high incidence of violent crimes. The study further explores the non-linear relationships between e-scooter trip frequency and these key variables, revealing threshold effects and significant shifts in usage patterns. These insights offer valuable guidance for cities in the strategic development and regulation of shared e-scooter services.

Development of dual GPBAR1 agonist and RORγt inverse agonist for the treatment of inflammatory bowel diseases

Inflammatory bowel diseases (IBD), including Crohn's disease and ulcerative colitis, are chronic disorders characterized by dysregulated immune response and persistent inflammation. Recent studies suggest that bile acid receptors, particularly GPBAR1, and the transcription factor RORγt play critical roles in modulating intestinal inflammation. This study evaluates the therapeutic potential of PBT002, a dual GPBAR1 agonist and RORγt inverse agonist, in IBD models. The effects of PBT002 were assessed through in vitro and in vivo experiments. Macrophages and T lymphocytes obtained from the buffy coat were exposed to PBT002 to evaluate its immunomodulatory activity. The beneficial effects in vivo were evaluated in mouse models of colitis induced by TNBS, DSS or DSS + IL-23 using also a Gpbar1 knock-out male mice. PBT002 exhibited an EC50 of 1.2 µM for GPBAR1 and an IC50 of 2.8 µM for RORγt. In in vitro, PBT002 modulated macrophage polarization towards an anti-inflammatory M2 phenotype and reduced Th17 cell markers while increasing Treg markers. In the TNBS-induced colitis model, PBT002 reduced weight loss, CDAI, and colon damage, while it modulated cytokine gene expression towards an anti-inflammatory profile. In GPBAR1-/-, the anti-inflammatory effects of PBT002 were attenuated, indicating partial GPBAR1 dependence. RNA sequencing revealed significant modulation of inflammatory pathways by PBT002. In DSS+IL-23 induced colitis, PBT002 mitigated disease exacerbation, reducing pro-inflammatory cytokine levels and immune cell infiltration. In conclusion, PBT002, a GPBAR1 agonist and RORγt inverse agonist, modulates both the innate and adaptive immune responses to reduce inflammation and disease severity in models of IBD.

Harnessing Machine Learning to Predict Methadone Overdose Risk: Insights from Illinois's SUDORS Data

Knowledge of decedent characteristics associated with methadone involvement in unintentional drug overdose deaths can help to inform efforts to intervene in these deaths. ML models with existing data, offer a cost-effective means to help create effective interventions. In this study, we aim to develop an optimized ML algorithm for predicting overdose death patterns due to methadone using data from Illinois's State Unintentional Drug Overdose Reporting System (SUDORS). We utilized IL SUDORS 2019-2022 data for training (n = 11931) with selected indicators (n = 28). To address the imbalance in non-methadone-related deaths, we applied the Synthetic Minority Oversampling Technique. Next, we assessed various machine learning models, including logistic regression, support vector machine, random forest, neural network, and ridge regression. Model performance was evaluated using metrics such as accuracy, precision, area under the precision-recall curve (AUPRC), etc. Utilizing separate training (n = 15813), validation (n = 3388), and test (n = 3389) datasets, the Random Forest Model outperforms all other models with 92% accuracy, 91% precision, 93% recall, and 0.97 AUPRC. Notably, these outcomes were achieved using primarily demographic indicators. Through analyzing partial dependence plots, we were able to see how changes in each indicator dynamically influence methadone-related fatalities. This study demonstrates the potential of ML models in identification of methadone involvement in unintentional drug overdose deaths and contributes to the knowledge base for the prevention of these deaths. Leveraging SUDORS data, the Random Forest Model exhibited exceptional performance, highlighting its value for healthcare professionals, prevention and harm reduction specialists, and policymakers.

An interpretable machine learning approach for predicting the capacity and failure mode of reinforced concrete columns

During seismic events, reinforced concrete (RC) columns play a crucial role in maintaining buildings’ structural integrity. This motivated engineers and practitioners to search for key parameters that influence the load-carrying capacity and failure mechanisms of such columns. However, the complexity and nonlinearity of seismic effects along with the intricate nature of RC columns as a composite system challenge the capabilities of analytical and empirical approaches to accurately capture the response of RC columns. Subsequently, the present study utilizes Machine Learning (ML) techniques to identify the failure modes and predict the corresponding capacities of RC columns based on both their geometrical and material properties. Decision trees and different ensemble methods were employed to predict both the columns’ failure mode and ultimate capacity. A multivariate dataset consisting of 486 cyclically loaded rectangular and circular columns was used to develop and validate the models. In addition, different embedded variable selection techniques were employed to evaluate the significance of input parameters in predicting the performance of columns. Moreover, partial dependence plots and accumulated local effects were employed to uncover the interrelationships between the input features and the modelled outputs. The developed models yielded an average accuracy of 90% and 95% for predicting the failure mode and ultimate capacity of RC columns, respectively. Given such high accuracy, it can be inferred that, ML techniques have the potential to provide efficient and reliable prediction tools to support seismic design and assessment decisions - mitigating seismic risks and empowering resilience planning in the face of extreme events.

Machine learning-driven prediction of phosphorus adsorption capacity of biochar: Insights for adsorbent design and process optimization

Phosphorus (P) pollution in aquatic environments poses significant environmental challenges, necessitating the development of effective remediation strategies, and biochar has emerged as a promising adsorbent for P removal at the cost of extensive research resources worldwide. In this study, a machine learning approach was proposed to simulate and predict the performance of biochar in removing P from water. A dataset consisting of 190 types of biochar was compiled from literature, encompassing various variables including biochar characteristics, water quality parameters, and operating conditions. Subsequently, the random forest and CatBoost algorithms were fine-tuned to establish a predictive model for P adsorption capacity. The results demonstrated that the optimized CatBoost model exhibited high prediction accuracy with an R2 value of 0.9573, and biochar dosage, initial P concentration in water, and C content in biochar were identified as the predominant factors. Furthermore, partial dependence analysis was employed to examine the impact of individual variables and interactions between two features, providing valuable insights for adsorbent design and operating condition optimization. This work presented a comprehensive framework for applying a machine learning approach to address environmental issues and provided a valuable tool for advancing the design and implementation of biochar-based water treatment systems.

Synergistic assessment of multi-scenario urban waterlogging through data-driven decoupling analysis in high-density urban areas: A case study in Shenzhen, China

Extreme meteorological events and rapid urbanization have led to serious urban flooding problems. Characterizing spatial variations in flooding susceptibility and elucidating its driving factors are essential for preventing damages from urban pluvial flooding. However, conventional methods, limited by spatial heterogeneity and the intricate mechanisms of urban flooding, frequently demonstrated a deficiency in precision when assessing flooding susceptibility in dense urban areas. Therefore, this study proposed a novel framework for an integrated assessment of urban flood susceptibility, based on a comprehensive cascade modeling chain consisting of XGBoost, SHapley Additive exPlanations (SHAP), and Partial Dependence Plots (PDP) in combination with K-means. It aimed to recognize the specific influence of urban morphology and the spatial patterns of flooding risk agglomeration under different rainfall scenarios in high-density urban areas. The XGBoost model demonstrated enhanced accuracy and robustness relative to other three benchmark models: RF, SVR, and BPDNN. This superiority was effectively validated during both training and independent testing in Shenzhen. The results indicated that urban 3D morphology characteristics were the dominant factors for waterlogging magnitude, which occupied 46.02 % of relative contribution. Through PDP analysis, multi-staged trends highlighted critical thresholds and interactions between significant indicators like building congestion degree (BCD) and floor area ratio (FAR). Specifically, optimal intervals like BCD between 0 and 0.075 coupled with FAR values between 0.5 and 1 have the potential to substantially mitigate flooding risks. These findings emphasize the need for strategic building configuration within urban planning frameworks. In terms of the spatial-temporal assessment, a significant aggregation effect of high-risk areas that prone to prolonged duration or high-intensity rainfall scenarios emerged in the old urban districts. The approach in the present study provides quantitative insights into waterlogging adaptation strategies for sustainable urban planning and design.

Machine learning analysis for detecting late recurrence and loss to follow‐up after renal cell carcinoma surgery

AbstractObjectivesRenal cell carcinoma (RCC) is shown to have a tendency for late recurrence, occurring 5 or more years after curative surgery. Imaging diagnosis is required for follow‐up, and there is no definitive answer as to how long this should continue. Some patients discontinue follow‐up visits at their own discretion. How best to predict late recurrence and loss to follow‐up (LF) remains unclear.Patients and methodsThis study targeted patients diagnosed with non‐metastatic RCC who underwent either radical or partial nephrectomy at Chiba University Hospital between 1988 and 2021. Follow‐up for patients with RCC is typically lifelong. We used random survival forests (RSFs), a machine learning‐based survival analysis method, to predict late recurrence and LF. For verification of prediction accuracy, we applied the time‐dependent area under the receiver operating characteristic curve (t‐AUC). To analyse the risks of late recurrence and LF, SurvSHAP(t) and partial dependence plots were used.ResultsWe analysed 1051 cases in this study. Median follow‐up was 58.5 (range: 0–376) months. The predictive accuracy of recurrence using RSF was t‐AUC 0.806, 0.761, 0.674 and 0.566 at 60, 120, 180 and 240 months postoperatively, respectively. The recurrence risk impact showed a time‐dependent increase up to approximately 50 months postoperatively. Beyond 50 months, there were no distinct risk factors characteristic of late recurrence. The predictive accuracy of LF using RSF was t‐AUC 0.542, 0.699, 0.685, 0.628 and 0.674 at 60, 120, 180, 240 and 300 months postoperatively, respectively. The risk of LF increased with advancing age beyond 70 years.ConclusionIt is difficult to identify factors that predict late recurrence. For long‐term follow‐up observation, it is essential to pay particular attention to patients with RCC aged 70 years and above. Establishing frameworks to facilitate collaboration with local hospitals near patients' residences and providing care within the community is necessary.

Explaining Machine Learning by Bootstrapping Partial Marginal Effects and Shapley Values

Machine learning and artificial intelligence are often described as “black boxes.” Traditional linear regression is interpreted through its marginal relationships as captured by regression coefficients. We show that the same marginal relationship can be described rigorously for any machine learning model by calculating the slope of the partial dependence functions, which we call the partial marginal effect (PME). We prove that the PME of OLS is analytically equivalent to the OLS regression coefficient. Bootstrapping provides standard errors and confidence intervals around the point estimates of the PMEs. We apply the PME to a hedonic house pricing example and demonstrate that the PMEs of neural networks, support vector machines, random forests, and gradient boosting models reveal the non-linear relationships discovered by the machine learning models and allow direct comparison between those models and a traditional linear regression. Finally we extend PME to a Shapley value decomposition and explore how it can be used to further explain model outputs.

Improving Lameness Detection in Cows: A Machine Learning Algorithm Application

The deployment of diverse data-generating technologies in livestock farming holds the promise of early disease detection and improved animal well-being. In this paper, we combine routinely collected dairy farm and herd data with weather and high frequency sensor data from 6 farms to predict new lameness events in various future periods, spanning from the following day to 3 weeks. A Random Forest classifier, using input features selected by the Boruta Algorithm, was used for the prediction task; effects of individual features were further assessed using partial dependence plots. We achieve precision scores of up to 93% when predicting lameness for the next 3 weeks and when using information from the last 3 weeks, combined with a balanced accuracy of 79%. Removing sensor data results have tendency to reduce the precision for predictions, especially when using information from the last one,2 or 3 weeks. Moving to a larger data set (without sensor data) of 44 farms keeps the similar balanced accuracy but reduces precision by more than 30%, revealing a substantial a trade-off in model quality between false positives (false lameness alerts) and false negatives (missed lameness events). Sensor data holds promise to further improve the precision of these models, but can be partially compensated by high resolution data from other systems, such as automated milking systems.

Data-driven prediction of electrospun nanofiber diameter using machine learning: A comprehensive study and web-based tool development

Electrospinning has emerged as a versatile technique for fabricating polymer nanofibers with diameters ranging from tens to hundreds of nanometers. Precisely controlling and predicting the nanofiber diameter is crucial for various applications. The complex interplay of multiple electrospinning parameters presents a significant challenge in this endeavor. This study introduces a novel data-driven approach using machine learning (ML) to predict and optimize the diameter of electrospun nanofibers. A comprehensive dataset of approximately 430 data points was compiled from literature sources, including polymer properties and electrospinning process parameters. Six ML algorithms were evaluated: Decision Tree (DT), Extra Trees Regression (ETR), Kernel Ridge Regression (KRR), Random Forest (RF), Support Vector Regression (SVR), and eXtreme Gradient Boosting (XGB). A 75/25 train-test split was selected for optimal model evaluation, with 10-fold cross-validation employed for hyperparameter tuning and estimating the variability of performance metrics. RF and ETR demonstrated the best predictive accuracy, with R2 values of 0.9468 and 0.9421, and RMSE values of 92.3 nm and 96.1 nm, respectively, on the test set. Feature importance analysis and SHapley Additive exPlanations (SHAP) values revealed that polymer concentration, applied voltage, and feed rate significantly influence the nanofiber diameter, which was further elucidated by partial dependence plots. To facilitate accessibility and collaboration, an interactive web server called ENDP was developed, allowing users to input polymer properties and process parameters to predict the nanofiber diameter. This study showcases the potential of ML in guiding the design and optimization of electrospun nanofibers, providing valuable insights for tailoring their properties into specific applications.

Carbon dioxide storage and cumulative oil production predictions in unconventional reservoirs applying optimized machine-learning models

To achieve carbon dioxide (CO2) storage through enhanced oil recovery, accurate forecasting of CO2 subsurface storage and cumulative oil production is essential. This study develops hybrid predictive models for the determination of CO2 storage mass and cumulative oil production in unconventional reservoirs. It does so with two multi-layer perceptron neural networks (MLPNN) and a least-squares support vector machine (LSSVM), hybridized with grey wolf optimization (GWO) and/or particle swarm optimization (PSO). Large, simulated datasets were divided into training (70%) and testing (30%) groups, with normalization applied to both groups. Mahalanobis distance identifies/eliminates outliers in the training subset only. A non-dominated sorting genetic algorithm (NSGA-II) combined with LSSVM selected seven influential features from the nine available input parameters: reservoir depth, porosity, permeability, thickness, bottom-hole pressure, area, CO2 injection rate, residual oil saturation to gas flooding, and residual oil saturation to water flooding. Predictive models were developed and tested, with performance evaluated with an overfitting index (OFI), scoring analysis, and partial dependence plots (PDP), during training and independent testing to enhance model focus and effectiveness. The LSSVM-GWO model generated the lowest root mean square error (RMSE) values (0.4052 MMT for CO2 storage and 9.7392 MMbbl for cumulative oil production) in the training group. That trained model also exhibited excellent generalization and minimal overfitting when applied to the testing group (RMSE of 0.6224 MMT for CO2 storage and 12.5143 MMbbl for cumulative oil production). PDP analysis revealed that the input features “area” and “porosity” had the most influence on the LSSVM-GWO model's prediction performance. This paper presents a new hybrid modeling approach that achieves accurate forecasting of CO2 subsurface storage and cumulative oil production. It also establishes a new standard for such forecasting, which can lead to the development of more effective and sustainable solutions for oil recovery.

Nonlinear impacts of landscape and climatological interactions on urban thermal environment during a hot and rainy summer

Investigating the nonlinear impacts of urban landscape and climatic parameters on urban temperatures, a critical issue within urban climatology. Chengdu, characterized by its hot, rainy summers and rapid urban development, serves as an ideal model to illustrate these dynamics. Our investigation utilizes advanced analytical methods such as Random Forests (RF), SHapley additive explanation (SHAP), and Partial Dependence Plots (PDP) to analyze how landscape and climatic factors influence air temperature (AT) and land surface temperature (LST). Significant findings reveal profound thermal heterogeneity across Chengdu’s urban fabric, underscored by spatially distinct phenomena where some regions exhibit strong contrasts in temperature impacts due to varying climatic and landscape factors. The study identifies relative humidity and rainfall as key drivers of temperature variations during the summer months, reflecting Chengdu’s specific climatic idiosyncrasies. These insights are critical, as they highlight how urban planning and green infrastructure can be strategically used to mitigate adverse thermal effects. The research not only enhances understanding of the complex interplays within urban microclimates but also offers new perspectives on urban heat management. It contributes to the scientific community by providing evidence-based strategies for urban planners to counter the urban heat island effect and enhance urban resilience against climate change. This comprehensive analysis underscores the importance of incorporating multiple variables into urban climate models, lays the groundwork for more refined urban environmental policies and practices.

Machine learning-based prediction of indoor thermal comfort in traditional Chinese dwellings: A case study of Hankou Lifen

Traditional residential dwellings as architectural heritage holds substantial research value and significance for regional studies. This study investigates the indoor thermal comfort of traditional “Hankou Lifen” dwellings compared to new residences during summer through field measurements and questionnaires. The results demonstrate that the indoor thermal environment of “Lifen” houses is inferior, leading to lower thermal comfort for the residents. A linear regression model assesses summer residents' thermal comfort, considering psychological perspectives to calculate natural and acceptable temperature ranges. However, these models primarily rely solely on temperature and overlook other environmental and personal factors, due to their lack of accuracy in capturing complex non-linear relationships. To address these limitations, six machine learning models with high predictive performance are selected for analyses. Data preprocessing techniques, such as upsampling, and hyper-parameter optimization are employed to enhance prediction performance. The selection of algorithms is based on metrics such as accuracy, precision, recall, and F1 score, with further performance evaluation conducted using ROC curves and confusion matrices. Through evaluation, XGBoost and LightGBM show superior performance in predicting residential thermal comfort.Finally, SHAP(SHapley Additive exPlanations) values and PDP(Partial Dependence Plot) are used to interpret the models, revealing various factors influencing residents' thermal sensations in both new and old dwellings. These findings emphasize the significance of thermal comfort in the preservation and revitalization of historic dwellings and propose a novel framework for advancing the sustainable utilization of traditional dwellings.

Collaborative and professional accommodations on Airbnb: Exploring patterns for sustainable tourism management in Spain

This study analyses the characteristics of accommodation types (collaborative and professional) in Madrid, Barcelona, Seville, and Valencia. It examines location factors, including proximity to top tourist spots, concentration of Airbnb accommodations, and population density, to identify which accommodation type contributes to negative impacts in these cities. Additionally, it delves into pricing features, reviews, length of stay, accommodation services, and host characteristics to distinguish between collaborative and professional accommodations.Logistic regression analysis and Random Forest are used, employing 500 decision trees to explore non-linear and complex relationships. Partial dependence plots offer insights into how predictions evolve across different ranges, providing valuable information about the locations, population density, tenure, and reviews associated with the two accommodation types.The findings reveal that professional accommodations tend to be concentrated in areas with high population density and near major tourist attractions, exacerbating tourist overcrowding issues. Conversely, collaborative accommodations are situated more than 4 km away from the main tourist spots, potentially contributing to a more balanced distribution of tourists and avoiding excessive concentration.These findings have significant implications for urban tourism dynamics, offering valuable insights that can inform urban planning, policy development, and sustainable tourism practices.

Model-agnostic variable importance for predictive uncertainty: an entropy-based approach

In order to trust the predictions of a machine learning algorithm, it is necessary to understand the factors that contribute to those predictions. In the case of probabilistic and uncertainty-aware models, it is necessary to understand not only the reasons for the predictions themselves, but also the reasons for the model’s level of confidence in those predictions. In this paper, we show how existing methods in explainability can be extended to uncertainty-aware models and how such extensions can be used to understand the sources of uncertainty in a model’s predictive distribution. In particular, by adapting permutation feature importance, partial dependence plots, and individual conditional expectation plots, we demonstrate that novel insights into model behaviour may be obtained and that these methods can be used to measure the impact of features on both the entropy of the predictive distribution and the log-likelihood of the ground truth labels under that distribution. With experiments using both synthetic and real-world data, we demonstrate the utility of these approaches to understand both the sources of uncertainty and their impact on model performance.

Interactive effects analysis of road, traffic, and weather characteristics on shared e-bike speeding risk: A data-driven approach

As electric bikes (e-bikes) rapidly develop in China, their traffic safety issues are becoming increasingly prominent. Accurately detecting risky riding behaviors and conducting mechanism analysis on the multiple risk factors are crucial in formulating and implementing precise management policies. The emergence of shared e-bikes and the advancements in interpretable machine learning present new opportunities for accurately analyzing the determinants of risky riding behaviors. The primary objective of this study is to examine and analyze the risk factors related to speeding behavior to aid urban management agencies in crafting necessary management policies. This study utilizes a large-scale dataset of shared e-bike trajectory data to establish a framework for detecting speeding behavior. Subsequently, the extreme gradient boosting (XGBoost) model is employed to identify the level of speeding risk by leveraging its excellent identification ability. Moreover, based on measuring the degree of interaction among road, traffic, and weather characteristics, the investigation of the complex interactive effects of these risk factors on high-risk speeding is conducted using bivariate partial dependence plots (PDP) by its superior parsing ability. Feature importance analysis results indicate that the top five ranked variables that significantly affect the identified results of speed risk levels are land use density, rainfall, road level, curbside parking density, and bike lane width. The interaction analysis results indicate that higher levels of road and bike lane width correspond to an increased possibility of high-risk speeding among riders. Land use density, curbside parking density, and rainfall display a nonlinear effect on high-risk speeding. Introducing road level, bike lane width, and time interval could change the patterns of nonlinear effects in land use density, curbside parking density, and rainfall. Finally, several policy recommendations are proposed to improve e-bike traffic safety by utilizing the extracted feature values associated with a higher probability of high-risk speeding.

Test and Measure for Partial Mean Dependence Based on Machine Learning Methods

It is of importance to investigate the significance of a subset of covariates W for the response Y given covariates Z in regression modeling. To this end, we propose a significance test for the partial mean independence problem based on machine learning methods and data splitting. The test statistic converges to the standard Chi-squared distribution under the null hypothesis while it converges to a normal distribution under the fixed alternative hypothesis. Power enhancement and algorithm stability are also discussed. If the null hypothesis is rejected, we propose a partial Generalized Measure of Correlation (pGMC) to measure the partial mean dependence of Y given W after controlling for the nonlinear effect of Z. We present the appealing theoretical properties of the pGMC and establish the asymptotic normality of its estimator with the optimal root-N convergence rate. Furthermore, the valid confidence interval for the pGMC is also derived. As an important special case when there are no conditional covariates Z, we introduce a new test of overall significance of covariates for the response in a model-free setting. Numerical studies and real data analysis are also conducted to compare with existing approaches and to demonstrate the validity and flexibility of our proposed procedures. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work.