Permutation Importance Research Articles

Defining broad scale habitat associates for an imperiled crayfish, Cambarus veteranus

The Guyandotte River Crayfish (Cambarus veteranus) is a federally protected species of crayfish known from two streams within the Upper Guyandotte river basin in West Virginia. The entire known range of C. veteranus is represented by two disjunct populations and as such an assessment of broad and fine-scale habitat associates is warranted. The goal of this study was to utilize maximum entropy modelling (Maxent) to predict broad scale habitat associates for C. veteranus and to create a map of detection probability for every stream reach within the basin. Occurrence records from a 2015 survey of the Upper Guyandotte were used to create the models. Environmental variables used in the models included stream order, elevation, landcover, and presence of mining activities. The mean area under the receiver operating characteristic curve (AUC) was 0.975 showing the model had high predictive accuracy. Stream order had the highest variable contribution and permutation importance. The model predicted that C. veteranus prefers 4th and 5th order streams. Ground validation will be completed to further test the accuracy of the models.

Environmental filtering improves ecological niche models across multiple scales

AbstractA clear challenge for ecological niche modelling is determining how to best mitigate the effects of sampling bias from commonly collected biodiversity data. Recent approaches have focused on filtering occurrences in overrepresented regions based on geographic or environmental proximity.We tested the efficacy of filtering in geographic and environmental space using occurrence data from four species. Our evaluation strategies examined 14 distance measures in geographic and environmental spaces and eight combinations of environmental variables and their ordinations. This resulted in 78 datasets for each species, which we evaluated using area under the curve (AUC), the difference between training and testingAUC, omission rate, the true skill statistic, and Schoener's D to examine the effects of different filtering schemes.The degree of change produced by filtering on predicted suitability and evaluation statistics increased with increasing range size. Environmental filtering resulted in higher model fit at larger extents and retained more occurrences than geographic filtering.Our results indicate that models should be evaluated using multiple evaluation statistics at multiple thresholds. The use of bin sizes when filtering in environmental space allows for simple comparison between species and filter types and makes for an easily reportable and repeatable distance metric. We specifically recommend that ecological niche models using natural history collection data filter in environmental space with variables derived from permutation importance or the first few axes of a principal components ordination.

A species distribution model for Paracentrotus lividus: predicted projections of habitat suitability

A species distribution model for Paracentrotus lividus: predicted projections of habitat suitability

Deep Granular Feature-Label Distribution Learning for Neuroimaging-based Infant Age Prediction.

Neuroimaging-based infant age prediction is important for brain development analysis but often suffers insufficient data. To address this challenge, we introduce label distribution learning (LDL), a popular machine learning paradigm focusing on the small sample problem, for infant age prediction. As directly applying LDL yields dramatically increased number of day-to-day age labels and also extremely scarce data describing each label, we propose a new strategy, called granular label distribution (GLD). Particularly, by assembling the adjacent labels to granules and designing granular distributions, GLD makes each brain MRI contribute to not only its own age but also its neighboring ages at a granule scale, which effectively keeps the information augmentation superiority of LDL and reduces the number of labels. Furthermore, to extremely augment the information supplied by the small data, we propose a novel method named granular feature distribution (GFD). GFD leverages the variability of the brain images at the same age, thus significantly increasing the learning effectiveness. Moreover, deep neural network is exploited to approximate the GLD. These strategies constitute a new model: deep granular feature-label distribution learning (DGFLDL). By taking 8 types of cortical morphometric features from structural MRI as predictors, the proposed DGFLDL is validated on infant age prediction using 384 brain MRI scans from 35 to 848 days after birth. Our proposed method, approaching the mean absolute error as 36.1 days, significantly outperforms the baseline methods. Besides, the permutation importance analysis of features based on our method reveals important biomarkers of infant brain development.

Incorporating local-scale variables into distribution models enhances predictability for rare plant species with biological dependencies

The conservation of rare species is typically challenging because of incomplete knowledge about their biology and distributions. Species distribution models (SDMs) have emerged as an important tool for improving the efficiency of rare species conservation. However, these models must include biologically relevant predictor variables at scales appropriate for discriminating suitable and unsuitable habitat. We used a species distribution modelling tool, maximum entropy (Maxent), to assess the relative influence of biologically relevant topographic characteristics, land cover features, geological formations, and edaphic factors on the occurrence of the endangered endemic orchid Spiranthes parksii (Navasota ladies’ tresses). Our final model produced an excellent AUC value (0.984), with the permutation importance to model fit of predictor variables representing topographic characteristics, land cover features, geological formations, and edaphic factors summing to 8.17, 35.12, 10.43, and 46.28%, respectively. Local-scale edaphic variables were the most informative, with soil taxonomic units explaining the highest amount of variance (36.40%) of all variables included in the model. These results document the importance of local edaphic characteristics in discriminating between suitable and unsuitable habitat for S. parksii, and emphasize the importance of including local-scale edaphic factors in SDMs for species such as S. parksii with specialized habitat requirements and close relationships with other organisms.

Postoperative bleeding risk prediction for patients undergoing colorectal surgery

BackgroundThere is limited consensus regarding risk factors for postoperative bleeding. The objective of this work was to investigate the capability of machine learning techniques in combination with practice-based longitudinal electronic medical record data for identifying potential new risk factors for postoperative bleeding and predicting patients at high risk of postoperative bleeding. MethodsA retrospective study was conducted for patients who underwent colorectal surgery 1998–2015 at a single tertiary referral center. Various predictors were extracted from electronic medical record. The outcome of interest was the occurrence of postoperative bleeding within 7 days of surgery. Logistic regression and gradient boosting machine models were trained. Area under the receiver operating curve and area under the precision recall curve were used to evaluate the performance to different models. ResultsOf 13,399 cases undergoing colorectal resection, 1,680 (12.5%) experienced postoperative bleeding. A total of 299 variables were evaluated. Logistic regression and gradient boosting machine models returned an area under the receiver operating curve of 0.735 and 0.822 and area under the precision recall curve of 0.287 and 0.423, respectively. In addition to well-known risk factors for postoperative bleeding, nutrition (ranked third), weakness (ranked fifth), patient mobility (ranked sixth), and activity level (ranked eighth) were found to be novel predictors in the gradient boosting machine model based on permutation importance. ConclusionThe study identified measures of functional capacity of patient as novel predictors of postoperative bleeding. The study found that risk of postoperative bleeding can be assessed, allowing for better use of human resources in addressing this important adverse event after surgery.

Socioecological system and potential deforestation in Western Amazon forest landscapes

The ecosystem services provided by tropical forests are affected by deforestation. Territorial management strategies aim to prevent and mitigate forest loss. Therefore, modeling potential land use changes is important for forest management, monitoring, and evaluation. This study determined whether there are relationships between forest vulnerability to deforestation (potential deforestation distribution) and the forest management policies applied in the Ecuadorian Amazon.Proxy and underlying variables were used to construct a statistical model, based on the principle of maximum entropy that could predict potential land use changes. Entropy can be seen as a measure of uncertainty for a density function. Receiver operating characteristics (ROC) analysis and the Jackknife Test were used to validate the model. The importance of input variables in the model was determined through: Percent Contribution (PC) and Permutation Importance (PI). The results were compared with prevailing regional forest management strategies.The socioeconomic variables that provided the largest amount of information in the overall model (AUC = 0.81) and that showed most of the information not present in other variables were: “Protected areas-Intangible zone” (PC = 24%, PI = 12.4%), “timber harvesting programs” (PC = 21.7%, PI = 4.7%), “road network” (PC = 18.9%, PI = 7.7%), and “poverty rate” (PC = 3.7%, PI = 6.1%). Also, the biophysical variable “temperature” (PC = 7,9%, PI = 22.3%) provided information in the overall model.The results suggested the need for changes in forest management strategies. Forest policies and management plans should consider integrating and strengthening protected areas and intangible zones, as well as restricting timber harvesting in native forest and establishing forest areas under permanent management. Furthermore, the results also suggested that financial incentive programs to reduce deforestation have to be evaluated because their present distribution is inefficient. In this context, conservation incentive plans need to be revised so that they focus on areas at deforestation risk.

Development of the first georeferenced map of Rhipicephalus (Boophilus) spp. in Mexico from 1970 to date and prediction of its spatial distribution.

The tick genus Ripicephalus (Boophilus), particularly R. microplus, is one of the most important ectoparasites that affects livestock health and considered an epidemiological risk because it causes significant economic losses due, mainly, to restrictions in the export of infested animals to several countries. Its spatial distribution has been tied to environmental factors, mainly warm temperatures and high relative humidity. In this work, we integrated a dataset consisting of 5843 records of Rhipicephalus spp., in Mexico covering close to 50 years to know which environmental variables mostly influence this ticks' distribution. Occurrences were georeferenced using the software DIVA-GIS and the potential current distribution was modelled using the maximum entropy method (Maxent). The algorithm generated a map of high predictive capability (Area under the curve = 0.942), providing the various contribution and permutation importance of the tested variables. Precipitation seasonality, particularly in March, and isothermality were found to be the most significant climate variables in determining the probability of spatial distribution of Rhipicephalus spp. in Mexico (15.7%, 36.0% and 11.1%, respectively). Our findings demonstrate that Rhipicephalus has colonized Mexico widely, including areas characterized by different types of climate. We conclude that the Maxent distribution model using Rhipicephalus records and a set of environmental variables can predict the extent of the tick range in this country, information that should support the development of integrated control strategies.

Hydrological Regime Monitoring and Mapping of the Zhalong Wetland through Integrating Time Series Radarsat-2 and Landsat Imagery

Zhalong wetland is a globally important breeding habitat for many rare migratory bird species. Prompted by the high demand for temporal and spatial information about the wetland’s hydrological regimes and landscape patterns, eight time series Radarsat-2 images were utilized to detect the flooding characteristics of the Zhalong wetland. Subsequently, a random forest model was built to discriminate wetlands from other land cover types, combining with optical, radar, and hydrological regime data derived from multitemporal synthetic aperture radar (SAR) images. The results showed that hydrological regimes variables, including flooding extent and flooding frequency, derived from multitemporal SAR images, improve the land cover classification accuracy in the natural wetlands distribution area. The permutation importance scores derived from the random forest classifier indicate that normalized difference vegetation index (NDVI) calculated from optical imagery and the flooding frequency derived from multitemporal SAR imagery were found to be the most important variables for land cover mapping. Accuracy testing indicate that the addition of hydrological regime features effectively depressed the omission error rates (from 52.14% to 2.88%) of marsh and the commission error (from 77.34% to 51.27%) of meadow, thereby improving the overall classification accuracy (from 76.49% to 91.73%). The hydrological regimes and land cover monitoring in the typical wetlands are important for eco-hydrological modeling, biodiversity conservation, and regional ecology and water security.

MP04-01 LONGITUDINAL SUBPHENOTYPES OF URINARY SYMPTOMS IN TYPE I DIABETES

You have accessJournal of UrologyBenign Prostatic Hyperplasia: Epidemiology & Evaluation1 Apr 2018MP04-01 LONGITUDINAL SUBPHENOTYPES OF URINARY SYMPTOMS IN TYPE I DIABETES Karandeep Singh, Hunter Wessells, J. Quentin Clemens, Rodney L. Dunn, Sarah K. Holt, James Hotaling, Rodica Pop-Busui, William Herman, Alan M. Jacobson, Barbara H. Braffett, Aruna V. Sarma, and for the DCCT/EDIC Research Group Karandeep SinghKarandeep Singh More articles by this author , Hunter WessellsHunter Wessells More articles by this author , J. Quentin ClemensJ. Quentin Clemens More articles by this author , Rodney L. DunnRodney L. Dunn More articles by this author , Sarah K. HoltSarah K. Holt More articles by this author , James HotalingJames Hotaling More articles by this author , Rodica Pop-BusuiRodica Pop-Busui More articles by this author , William HermanWilliam Herman More articles by this author , Alan M. JacobsonAlan M. Jacobson More articles by this author , Barbara H. BraffettBarbara H. Braffett More articles by this author , Aruna V. SarmaAruna V. Sarma More articles by this author , and for the DCCT/EDIC Research Groupfor the DCCT/EDIC Research Group More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.151AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES Lower urinary tract symptoms (LUTS) in men and urinary incontinence (UI) in women are dynamic conditions with numerous factors contributing to risk and progression. Understanding of natural history of urinary symptoms in diabetes is limited. We describe factors influencing the long-term trajectory of LUTS in men and UI in women with type 1 diabetes (T1DM) using non-parametric statistical methods. METHODS Longitudinal data from 565 men and 523 women in the Epidemiology of Diabetes Interventions and Complications (EDIC) study were used to define LUTS/UI annually over 5 years. Trajectories were identified using a k-means clustering approach on the American Urological Association Symptom Index (AUASI) in men and Sandvik Severity Index in women. Subphenotypes were selected based on cluster quality metrics. A multinomial random forest model was fit to predict subphenotype assignment using information from baseline. Factors associated with LUTS/UI trajectory were identified by permuting each predictor and assessing change in accuracy as compared to the non-permuted data using the fitted model. RESULTS At baseline, mean age was 51.0 (SD 6.4) and 50.1 (SD 7.0) in men and women, respectively. 27.5% of men reported moderate/severe LUTS and 28.7% of women moderate/severe UI. We identified 3 subphenotypes for UI among women and 4 for LUTS among men (Figure 1A, 1B). Subphenotypes D in men and C in women demonstrate a deterioration of function over time in contradistinction to stable patterns and low/moderate severity in other clusters. The classification model achieved a multiclass area-under-the-curve (AUC) of 0.65 for AUASI and 0.61 for Sandvik using 10-fold cross-validation. Based on permutation importance (Figure 1C, 1D), key predictors were age (p = 0.002) and autonomic dysfunction (p < 0.001) for LUTS trajectory and BMI (p = 0.02) for UI. CONCLUSIONS LUTS/UI are prevalent complications in men and women with T1DM. A limited number of factors were identified in delineating subphenotypes of LUTS/UI. These data may be relevant to future investigations of mechanisms for distinct endotypes of LUTS/UI, as well as candidates for interventions targeting metabolic factors contributing to variation in symptoms. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e31-e32 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Karandeep Singh More articles by this author Hunter Wessells More articles by this author J. Quentin Clemens More articles by this author Rodney L. Dunn More articles by this author Sarah K. Holt More articles by this author James Hotaling More articles by this author Rodica Pop-Busui More articles by this author William Herman More articles by this author Alan M. Jacobson More articles by this author Barbara H. Braffett More articles by this author Aruna V. Sarma More articles by this author for the DCCT/EDIC Research Group More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...

PREDICTION OF PIVOTAL RESPONSE TREATMENT OUTCOME WITH TASK FMRI USING RANDOM FOREST AND VARIABLE SELECTION.

Behavior intervention has shown promise for treatment for young children with autism spectrum disorder (ASD). However, current therapeutic decisions are based on trial and error, often leading to suboptimal outcomes. We propose an approach that employs task-based fMRI for early outcome prediction. Our strategy is based on the general linear model (GLM) and a random forest, combined with feature selection techniques. GLM analysis is performed on each voxel to get t-statistic of contrast between two tasks. Due to the high dimensionality of predictor variables, feature selection is crucial for accurate prediction. Thus we propose a two-step feature selection method: a "shadow" method to select all-relevant variables, followed by a stepwise method to select minimal-optimal set of variables for prediction. A few columns of random noise are generated and added as shadow variables. Regression based on the random forest is performed, and permutation importance of each variable is estimated. Candidate voxels with higher importance than the shadow are kept. Surviving voxels are fed into stepwise variable selection methods. We test both forward and backward stepwise selection. Our method was validated on a dataset of 20 children with ASD using leave-one-out cross-validation, and compared to other standard regression methods. The proposed pipeline generated highest accuracy.

What are the best predictors for invasive potential of weeds? Transferability evaluations of model predictions based on diverse environmental data sets for Flaveria bidentis

SummaryEcological niche models are widely used in the study of weed invasions, yet best approaches for selecting ecologically relevant environmental predictors for weeds remain unresolved. Here, we evaluate niche model transferability based on diverse environmental data sets for an invasive herb, Flaveria bidentis. This species is native to South America, but has established populations in China that pose a threat to agriculture and animal husbandry. Relevant environmental data sets were selected via five statistical approaches: permutation importance (PI) and jackknife test (JK) in Maxent, variable importance identified by boosting regression trees (BRT), ecological niche factor analysis (ENFA) and a newly released algorithm based on a fluctuation index (FI). Climate spaces occupied by native South American and introduced Chinese populations were compared based on these environmental data sets. Native niche model predictions in China were compared across environmental data sets and model settings (i.e. default versus fine‐tuned Maxent settings). Results suggest that native and introduced populations occupy two distinct climate spaces, but that this divergence likely results from background effects. Niche models based on fine‐tuned Maxent settings generally showed better discrimination ability than those based on default settings. The best model discrimination in China was attained in the FI model using fine‐tuned settings, followed by the BRT model on default settings. The best models suggest that highly suitable areas at risk of invasion in China are to the west and north‐east of present distributional areas. Results presented here provide predictions for F. bidentis in particular, but also shed light on procedures for selecting ecologically relevant predictors for invasive species distributional predictions more generally.

Mapping pedomemory of spodic morphology using a species distribution model

Red spruce (Picea rubens) ecosystems in the high elevations of the central Appalachians of the eastern United States are the focus of ongoing restoration efforts due to the valuable ecosystem services these forests provide. Recent research has shown that spodic materials still present in the soil represent the pedomemory of the historic extent of red spruce forests in the region. A dataset containing 221 points with varying spodic intensities and 29 environmental variables collected from the Monongahela National Forest in West Virginia, USA, was used to evaluate the utility of a species distribution model, Maximum Entropy (MaxEnt), for predicting the presence of spodic properties. MaxEnt was selected for evaluation because, as a presence-only model, it inherently omits absence locations and thereby reduces the risk of including false absences (i.e., herein, locations that have undergone some level of depodzolization) unlike other models previously used to predict pedomemory. Model outputs that employed three different spodic intensity class inputs—very weak to strong expression, weak to strong expression, and strong expression—resulted in similar spodic probability predictions, though there was less area mapped as transitional probabilities in the strong expression model than the two models that included weaker spodic intensity input data. Permutation importance indicated that no single or small subset of environmental variables controlled the three model outputs, perhaps because the environmental covariates may have been too coarse or not strongly enough associated with podzolization processes to be very important. When the output from the MaxEnt model using the full range of spodic intensities (very weak to strong) was compared to an output produced using a presence-absence model (random forests), there was approximately 62% agreement (where both models predicted presence or both predicted absence) for the cells in the top 40% of the predicted probabilities.

Modeling Raccoon (Procyon lotor) Habitat Connectivity to Identify Potential Corridors for Rabies Spread.

The United States Department of Agriculture (USDA), Animal and Plant Health Inspection Service (APHIS), Wildlife Services National Rabies Management Program has conducted cooperative oral rabies vaccination (ORV) programs since 1997. Understanding the eco-epidemiology of raccoon (Procyon lotor) variant rabies (raccoon rabies) is critical to successful management. Pine (Pinus spp.)-dominated landscapes generally support low relative raccoon densities that may inhibit rabies spread. However, confounding landscape features, such as wetlands and human development, represent potentially elevated risk corridors for rabies spread, possibly imperiling enhanced rabies surveillance and ORV planning. Raccoon habitat suitability in pine-dominated landscapes in Massachusetts, Florida, and Alabama was modeled by the maximum entropy (Maxent) procedure using raccoon presence, and landscape and environmental data. Replicated (n = 100/state) bootstrapped Maxent models based on raccoon sampling locations from 2012–2014 indicated that soil type was the most influential variable in Alabama (permutation importance PI = 38.3), which, based on its relation to landcover type and resource distribution and abundance, was unsurprising. Precipitation (PI = 46.9) and temperature (PI = 52.1) were the most important variables in Massachusetts and Florida, but these possibly spurious results require further investigation. The Alabama Maxent probability surface map was ingested into Circuitscape for conductance visualizations of potential areas of habitat connectivity. Incorporating these and future results into raccoon rabies containment and elimination strategies could result in significant cost-savings for rabies management here and elsewhere.

Reliability Optimization and Importance Analysis of Circular-Consecutive k-out-of-n System

The circular-consecutive k-out-of-n:F(G) system (Cir/Con/k/n:F(G) system) usually consists of n components arranged in a circle where the system fails (works) if consecutive k components fail (work). The optimization of the Cir/Con/k/n system is a typical case in the component assignment problem. In this paper, the Birnbaum importance-based genetic algorithm (BIGA), which takes the advantages of genetic algorithm and Birnbaum importance, is introduced to deal with the reliability optimization for Cir/Con/k/n system. The detailed process and property of BIGA are put forward at first. Then, some numerical experiments are implemented, whose results are compared with two classic Birnbaum importance-based search algorithms, to evaluate the effectiveness and efficiency of BIGA in Cir/Con/k/n system. Finally, three typical cases of Cir/Con/k/n systems are introduced to demonstrate the relationships among the component reliability, optimal permutation, and component importance.

Short-Term Traffic Flow Forecasting for Urban Roads Using Data-Driven Feature Selection Strategy and Bias-Corrected Random Forests

Urban traffic flow forecasting is essential to proactive traffic control and management. Most existing forecasting methods depend on proper and reliable input features, for example, weather conditions and spatiotemporal lagged variables of traffic flow. However, the feature selection process is often done manually without comprehensive evaluation and leads to inaccurate results. For that challenge, this paper presents an approach combining the bias-corrected random forests algorithm with a data-driven feature selection strategy for short-term urban traffic flow forecasting. First, several input features were extracted from traffic flow time series data. Then the importance of these features was quantified with the permutation importance measure. Next, a data-driven feature selection strategy was introduced to identify the most important features. Finally, the forecasting model was built on the bias-corrected random forests algorithm and the selected features. The proposed approach was validated with data collected from three types of urban roads (expressway, major arterial, and minor arterial) in Kunshan City, China. The proposed approach was also compared with 10 existing approaches to verify its effectiveness. The results of the validation and comparison show that even without further model tuning, the proposed approach achieves the lowest average mean absolute error and root mean square error on six stations while it achieves the second-best average performance in mean absolute percentage error. Meanwhile, the training efficiency is improved compared with the original random forests method owing to the use of the feature selection strategy.

A Permutation Importance-Based Feature Selection Method for Short-Term Electricity Load Forecasting Using Random Forest

The prediction accuracy of short-term load forecast (STLF) depends on prediction model choice and feature selection result. In this paper, a novel random forest (RF)-based feature selection method for STLF is proposed. First, 243 related features were extracted from historical load data and the time information of prediction points to form the original feature set. Subsequently, the original feature set was used to train an RF as the original model. After the training process, the prediction error of the original model on the test set was recorded and the permutation importance (PI) value of each feature was obtained. Then, an improved sequential backward search method was used to select the optimal forecasting feature subset based on the PI value of each feature. Finally, the optimal forecasting feature subset was used to train a new RF model as the final prediction model. Experiments showed that the prediction accuracy of RF trained by the optimal forecasting feature subset was higher than that of the original model and comparative models based on support vector regression and artificial neural network.

Africanized honeybee habitat suitability: a comparison between models for southern Utah and southern California

The Africanized honeybee (AHB; Apis mellifera scutellata) is an invasive species which poses a threat to the United States' agricultural industry because of potential decline in pollination services. Previous research has confirmed that the AHB may still expand its range farther north and that limiting environmental factors for AHB distribution vary across the country. This study examines similarities and differences in AHB distribution and the relative importance of environmental factors between two regions of the southwestern United States: southern Utah and southern California. The Maximum Entropy (MaxEnt) modeling approach was used to create two species distribution models of the AHB. First a model was created based on AHB presence data in Utah. This model was used to estimate the Utah distribution and also to project the California distribution based on Utah environmental preferences. The second model was created to predict distribution in California and to project distribution in Utah based on AHB environmental preferences in California. The level of influence of each variable was measured through percent contribution and permutation importance. Model performance was assessed through the area under the receiver operating characteristic curve (AUC). Models estimated AHB presence with high accuracy (AUC > 0.95) in original environments, but were less accurate (AUC < 0.8) in novel environments. Minimum temperature was the primary controlling factor of AHB distribution in each model, and other variables followed similar ranking of importance between the two models. Species response curves varied substantially between the two models. Models did not transfer well from one region to the other because of local differences in response curves and the relative importance of environmental variables, suggesting that AHB in these regions may not have realized their potential geographic range.

Do little interactions get lost in dark random forests?

BackgroundRandom forests have often been claimed to uncover interaction effects. However, if and how interaction effects can be differentiated from marginal effects remains unclear. In extensive simulation studies, we investigate whether random forest variable importance measures capture or detect gene-gene interactions. With capturing interactions, we define the ability to identify a variable that acts through an interaction with another one, while detection is the ability to identify an interaction effect as such.ResultsOf the single importance measures, the Gini importance captured interaction effects in most of the simulated scenarios, however, they were masked by marginal effects in other variables. With the permutation importance, the proportion of captured interactions was lower in all cases. Pairwise importance measures performed about equal, with a slight advantage for the joint variable importance method. However, the overall fraction of detected interactions was low. In almost all scenarios the detection fraction in a model with only marginal effects was larger than in a model with an interaction effect only.ConclusionsRandom forests are generally capable of capturing gene-gene interactions, but current variable importance measures are unable to detect them as interactions. In most of the cases, interactions are masked by marginal effects and interactions cannot be differentiated from marginal effects. Consequently, caution is warranted when claiming that random forests uncover interactions.Electronic supplementary materialThe online version of this article (doi:10.1186/s12859-016-0995-8) contains supplementary material, which is available to authorized users.

Correlation and variable importance in random forests

This paper is about variable selection with the random forests algorithm in presence of correlated predictors. In high-dimensional regression or classification frameworks, variable selection is a difficult task, that becomes even more challenging in the presence of highly correlated predictors. Firstly we provide a theoretical study of the permutation importance measure for an additive regression model. This allows us to describe how the correlation between predictors impacts the permutation importance. Our results motivate the use of the Recursive Feature Elimination (RFE) algorithm for variable selection in this context. This algorithm recursively eliminates the variables using permutation importance measure as a ranking criterion. Next various simulation experiments illustrate the efficiency of the RFE algorithm for selecting a small number of variables together with a good prediction error. Finally, this selection algorithm is tested on the Landsat Satellite data from the UCI Machine Learning Repository.