Nonlinear Machine Learning Techniques Research Articles

Should Humans Lie to Machines? The Incentive Compatibility of Lasso and GLM Structured Sparsity Estimators

We consider situations where a user feeds her attributes to a machine learning method that tries to predict her best option based on a random sample of other users. The predictor is incentive-compatible if the user has no incentive to misreport her covariates. Focusing on the popular Lasso estimation technique, we borrow tools from high-dimensional statistics to characterize sufficient conditions that ensure that Lasso is incentive compatible in the asymptotic case. We extend our results to a new nonlinear machine learning technique, Generalized Linear Model Structured Sparsity estimators. Our results show that incentive compatibility is achieved if the tuning parameter is kept above some threshold in the case of asymptotics.

Early Detection of Late Onset Sepsis in Extremely Preterm Infants Using Machine Learning: Towards an Early Warning System

A significant proportion of babies that are admitted to the neonatal intensive care unit (NICU) suffer from late onset sepsis (LOS). In order to prevent mortality and morbidity, the early detection of LOS is of the utmost importance. Recent works have found that the use of machine learning techniques might help detect LOS at an early stage. Some works have shown that linear methods (i.e., logistic regression) display a superior performance when predicting LOS. Nevertheless, as research on this topic is still in an early phase, it has not been ruled out that non-linear machine learning (ML) techniques can improve the predictive performance. Moreover, few studies have assessed the effect of parameters other than heart rate variability (HRV). Therefore, the current study investigates the effect of non-linear methods and assesses whether other vital parameters such as respiratory rate, perfusion index, and oxygen saturation could be of added value when predicting LOS. In contrast with the findings in the literature, it was found that non-linear methods showed a superior performance compared with linear models. In particular, it was found that random forest performed best (AUROC: 0.973), 24% better than logistic regression (AUROC: 0.782). Nevertheless, logistic regression was found to perform similarly to some non-linear models when trained with a short training window. Furthermore, when also taking training time into account, K-Nearest Neighbors was found to be the most beneficial (AUROC: 0.950). In line with the literature, we found that training the models on HRV features yielded the best results. Lastly, the results revealed that non-linear methods demonstrated a superior performance compared with linear methods when adding respiratory features to the HRV feature set, which ensured the greatest improvement in terms of AUROC score.

Multistage quality control in manufacturing process using blockchain with machine learning technique

Information security has more demand for digital technology. Every industry transfers its data through computer networks for legal communication. The Internet of Things (IoT), sensor-based, is one of the most current advanced tools that can handle appropriate measures to control data operations across manufacturing industries. The demand for predictive machine reliability and quality drives the development of intelligent manufacturing technologies. To this goal, a variety of machine learning algorithms are being studied. Data protection and monitoring is also another concern that is a critical component of the organization. To overcome these issues, the proposed method uses Blockchain Technology (BCT) and Machine Learning to secure the information operations and manage a dataset. Significant data approaches were employed to organize and evaluate the obtained dataset. BCT allows collecting sensor user access data, whereas ML classifiers distinguish between normal and malicious behavior to detect attacks. DoS, DDoS, intrusion, a man in the middle (MitM), brute force, cross-site scripting (XSS), and searching are the attacks detected by BCT. Furthermore, the hybrid prediction technique assessed the fault detection prediction component. The program's quality control was set using non-linear machine learning techniques that represented the complicated world and determined the actual positive rate of the standard control methodology used by the platform. The experimental result shows that the proposed method outperforms empirical metrics such as accuracy, precision, recall, and response time. The proposed method efficiently provides security between innovative manufacturing transactions.

Stable isotope and trace element analyses with non-linear machine-learning data analysis improved coffee origin classification and marker selection.

This study investigated the geographical origin classification of green coffee beans from continental to country and regional levels. An innovative approach combined stable isotope and trace element analyses with non-linear machine learning data analysis to improve coffee origin classification and marker selection. Specialty green coffee beans sourced from three continents, eight countries, and 22 regions were analyzed by measuring five isotope ratios (δ13 C, δ15 N, δ18 O, δ2 H, and δ34 S) and 41 trace elements. Partial least squares discriminant analysis (PLS-DA) was applied to the integrated dataset for origin classification. Origins were predicted well at the country level and showed promise at the regional level, with discriminating marker selection at all levels. However, PLS-DA predicted origin poorly at the continental and Central American regional levels. Non-linear machine learning techniques improved predictions and enabled the identification of a higher number of origin markers, and those that were identified were more relevant. The best predictive accuracy was found using ensemble decision trees, random forest and extreme gradient boost, with accuracies of up to 0.94 and 0.89 for continental and Central American regional models, respectively. The potential for advanced machine learning models to improve origin classification and the identification of relevant origin markers was demonstrated. The decision-tree-based models were superior with their embedded variable identification features and visual interpretation. © 2023 The Authors. Journal of The Science of Food and Agriculture published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.

Viscosity prediction of ionic liquids using NLR and SVM approaches

Using the molecular functional group, two alternative models for predicting the dynamic viscosity of ionic liquids (ILs) as function of temperature are presented. The group contributions were regressed by traditional nonlinear regression (NLR) and machine learning technique including support vector machine (SVM) regression. A new group contribution scheme was proposed that allows to describe a wide variety of ILs and to distinguish among isomers. The obtained models were developed using the largest experimental viscosity database assembled to date (15 251 accepted experimental data points covering 1654 distinct ILs). The hyperparameters of SVM regression model were obtained through Bayesian optimization method with k-folds cross-validation. The two proposed models were tested using n-times repeated k-folds cross-validation. The proposed models are discussed in detail according to various statistics, both in global values and in detailed values for each chemical family, and compared to other models reported in the literature. The results indicate that SVM regression model is much better than the traditional nonlinear model. For the sake of transparency, the proposed models were implemented using common spreadsheets and given in the supplementary material as Excel file.

Exploring the Impacts of Data Source, Model Types and Spatial Scales on the Soil Organic Carbon Prediction: A Case Study in the Red Soil Hilly Region of Southern China

Rapid and accurate mapping of soil organic carbon (SOC) is of great significance to understanding the spatial patterns of soil fertility and conducting soil carbon cycle research. Previous studies have dedicated considerable efforts to the spatial prediction of SOC content, but few have systematically quantified the effects of environmental covariates selection, the spatial scales and the model types on SOC prediction accuracy. Here, we spatially predicted SOC content through digital soil mapping (DSM) based on 186 topsoil (0–20 cm) samples in a typical hilly red soil region of southern China. Specifically, we first determined an optimal covariate set from different combinations of multiple environmental variables, including multi-sensor remote sensing images (Sentinel-1 and Sentinel-2), climate variables and DEM derivatives. Furthermore, we evaluated the impacts of spatial resolution (10 m, 30 m, 90 m, 250 m and 1000 m) of covariates and the model types (three linear and three non-linear machine learning techniques) on the SOC prediction. The results of the performance analysis showed that a combination of Sentinel-1/2-derived variables, climate and topographic predictors generated the best predictive performance. Among all variables, remote sensing covariates, especially Sentinel-2-derived predictors, were identified as the most important explanatory variables controlling the variability of SOC content. Moreover, the prediction accuracy declined significantly with the increased spatial scales and achieved the highest using the XGBoost model at 10 m resolution. Notably, non-linear machine learners yielded superior predictive capability in contrast with linear models in predicting SOC. Overall, our findings revealed that the optimal combination of predictor variables, spatial resolution and modeling techniques could considerably improve the prediction accuracy of the SOC content. Particularly, freely accessible Sentinel series satellites showed great potential in high-resolution digital mapping of soil properties.

Inverse groundwater salinization modeling in a sandstone’s aquifer using stand-alone models with an improved non-linear ensemble machine learning technique

Groundwater (GW) salinization is among several hydro-environmental processes that essentially influence water resources management, freshwater supply, water scarcity and sustainability. Machine Learning (ML) is an effective and robust tool for GW monitoring, primarily when chaotic and interrelated parameters influence the system. The current study presents the first inverse modelling of GW salinization in sandstone aquifer using stand-alone (Elman NN, GRNN, BPNN, and ILR) models coupled with nonlinear ensemble ML (GPR-E, SVR-E, and ANFIS-E) techniques. For this purpose, 25 input parameters were imposed on data mining; subsequently, inversely correlated variables including bicarbonate (HCO3−), nitrate (NO3−), Temperature (Temp), Lithium (Li), and Manganese (Mn) were used for modelling the Electrical conductivity (EC). Neuro-sensitivity feature selection (NSFS) ranking was used to determine the input combinations. The Modeling schema was based on the actual field and experimental data. The results indicated the potential of GRNN-C1 with a value of NSE = 0.998, and 0.928 in both calibration and validation phases, respectively. The outcomes revealed weaknesses in all other stand-alone models with low efficiency ranged from 0 to 60%. The attained ensemble results proved promising for all, with ANFIS-E-C1 (RMSE = 2.993, MAE = 0.879, and NSE = 0.999) emerging best in the validation phase. The overall proposed ensemble learning proved the excellent capability of ML to handle complex GW processes and served as a reliable predictive tool for water resources management. The contribution-based methodologies explored in this study involved inverse modelling, nonlinear feature selection, and nonlinear ensemble techniques, which was the first study in GW salinization.

Multi-task learning framework for predicting water quality using non-linear machine learning technique

Predicting the quality of water is a very important issue in an ecosystem and it can be used to control the increase of water contamination. Also, water quality prediction is a prominent complex non-linear multi-target learning problem and extracting a relevant subset of features from a large number of features with multiple targets is a challenging task. Existing water quality prediction model not focused on multi-target learning process simultaneously and not identifying the non-linear relationship between the features and target variables. Therefore, this study proposes a multi-task learning method dealing with multi-target regression using non-linear machine learning technique. Finally, experiments are conducted to build a prediction model based on the proposed methods to evaluate accuracy on water quality dataset. The experimental results indicate that our method increases the overall accuracy of the experimental dataset compared with the existing methods with the reduced number of significant features.

Application of machine and deep learning techniques to estimate NMR-derived permeability from conventional well logs and artificial 2D feature maps

Nuclear magnetic resonance (NMR) logs can provide information on some critical reservoir characteristics, such as permeability, which are rarely obtainable from conventional well logs. Nevertheless, high cost and operational constraints limit the wide application of NMR logging tools. In this study, a machine learning (ML)-based procedure is developed for fast and accurate estimation of NMR-derived permeability from conventional logs. Following a comprehensive preprocessing on the collected data, the procedure is trained and tested on a well log dataset, with selected conventional logs as inputs, and NMR-derived permeability as target, shallow and deep learning (DL) methods are applied to estimate permeability from selected conventional logs through artificial production of NMR-derived information from the input data. Three supervised ML algorithms are utilized and evaluated, including random forest (RF), group method of data handling (GMDH), and one-dimensional convolutional neural network (1D-CNN). Additionally, a modified two-dimensional CNN (named as Residual 2D-CNN) is developed which is fed by artificial 2D feature maps, generated from available conventional logs. The hyper-parameters of the ML and DL models are optimized using genetic algorithm (GA) to improve their performances. By comparing the output of each model with the permeability derived from NMR log, it is illustrated that nonlinear machine and deep learning techniques are helpful in estimation of NMR permeability. The obtained accuracy of RF, GMDH, 1D-CNN and Res 2D-CNN models, respectively, is 0.90, 0.90, 0.91 and 0.97 which indicate that Res 2D-CNN model is the most efficient method among the other applied techniques. This research also highlights the importance of using generated feature maps for training Res 2D-CNN model, and the essential effect of the applied modifications (i.e., implementing residual and deeper bottleneck architectures) on improving the accuracy of the predicted output and reducing the training time.

OSACN-Net: Automated Classification of Sleep Apnea Using Deep Learning Model and Smoothed Gabor Spectrograms of ECG Signal

Obstructive sleep apnea (OSA) is a severe sleep-associated respiratory disorder, caused due to periodic disruption of breath during sleep. It may cause a number of serious cardiovascular complications, including stroke. Generally, OSA is detected by polysomnography (PSG), a costly procedure, and may cause discomfort to the patient. Nowadays, electrocardiogram (ECG) signal-based detection techniques have been explored as an alternative to PSG for OSA detection. Usual linear and nonlinear machine learning techniques are mainly focused on handcrafted feature extraction and classification that are time-consuming and may not be suitable for huge data. Therefore, in this work, a deep learning model (DLM) using smoothed Gabor spectrogram (SGS) of ECG signals is proposed for automated OSA detection to obtain high performance. The proposed framework fed Gabor spectrogram and SGS of ECG signals as input to the pretrained Squeeze-Net, Res-Net50, and developed DLM called obstructive sleep apnea convolutional neural network (OSACN-Net). The proposed OSACN-Net achieved an average classification accuracy of 94.81% with SGS using a tenfold cross-validation strategy. Compared to Squeeze-Net and Res-Net50, developed OSACN-Net is more accurate and lightweight as it requires few learnable parameters, which makes it computationally fast and efficient. The comparison results showed that the proposed framework outperformed all existing state-of-the-art methodologies.

Multiclass moisture classification in woodchips using IIoT Wi-Fi and machine learning techniques

For the pulping process in a pulp & paper plant that uses woodchips as raw material, the moisture content (MC) of the woodchips is a major process disturbance that affects product quality and consumption of energy, water and chemicals. Existing woodchip MC sensing technologies have not been widely adopted by the industry due to unreliable performance and/or high maintenance requirements that can hardly be met in a manufacturing environment. To address these limitations, we propose a non-destructive, economic, and robust woodchip MC sensing approach utilizing channel state information (CSI) from industrial Internet-of-Things (IIoT) based Wi-Fi. While these IIoT devices are small, low-cost, and rugged to stand for harsh environment, they do have their limitations such as the raw CSI data are often very noisy and sensitive to woodchip packing. To address this, statistics pattern analysis (SPA) is utilized to extract physically and/or statistically meaningful features from the raw CSI data, which are sensitive to woodchip MC but not to packing. The SPA features are then used for developing multiclass classification models using various linear and nonlinear machine learning techniques to provide potential solutions to woodchip MC estimation for the pulp and paper industry. This work also demonstrates that classification accuracy alone is not a good performance metric for industrial applications, and the practical implications of misclassification must also be considered.

Integration of Blockchain, IoT and Machine Learning for Multistage Quality Control and Enhancing Security in Smart Manufacturing.

Smart manufacturing systems are growing based on the various requests for predicting the reliability and quality of equipment. Many machine learning techniques are being examined to that end. Another issue which considers an important part of industry is data security and management. To overcome the problems mentioned above, we applied the integrated methods of blockchain and machine learning to secure system transactions and handle a dataset to overcome the fake dataset. To manage and analyze the collected dataset, big data techniques were used. The blockchain system was implemented in the private Hyperledger Fabric platform. Similarly, the fault diagnosis prediction aspect was evaluated based on the hybrid prediction technique. The system’s quality control was evaluated based on non-linear machine learning techniques, which modeled that complex environment and found the true positive rate of the system’s quality control approach.

A Comparison of Linear and Non-Linear Machine Learning Techniques (PCA and SOM) for Characterizing Urban Nutrient Runoff

Urban stormwater runoff represents a significant challenge for the practical assessment of diffuse pollution sources on receiving water bodies. Given the high dimensionality of the problem, the main goal of this study was the comparison of linear and non-linear machine learning (ML) methods to characterize urban nutrient runoff from impervious surfaces. In particular, the principal component analysis (PCA) for the linear technique and the self-organizing map (SOM) for the non-linear technique were chosen and compared considering the high number of successful applications in the water quality field. To strengthen this comparison, these techniques were supported by well-known linear and non-linear methods. Those techniques were applied to a complete dataset with precipitation, flow rate, and water quality (sediments and nutrients) records of 577 events gathered for a watershed located in Southern Italy. According to the results, both linear and non-linear techniques can represent build-up and wash-off, the two main processes that characterize urban nutrient runoff. In particular, non-linear methods are able to capture and represent better the rainfall-runoff process and the transport of dissolved nutrients in urban runoff (dilution process). However, their computational time is higher than the linear technique (0.0054 s vs. 15.24 s, for linear and non-linear, respectively, in our study). The outcomes of this study provide significant insights into the application of ML methods for the water quality field.

Kernel Matrix Approximation on Class-Imbalanced Data With an Application to Scientific Simulation

Generating low-rank approximations of kernel matrices that arise in nonlinear machine learning techniques holds the potential to significantly alleviate the memory and computational burdens. A compelling approach centers on finding a concise set of exemplars or landmarks to reduce the number of similarity measure evaluations from quadratic to linear concerning the data size. However, a key challenge is to regulate tradeoffs between the quality of landmarks and resource consumption. Despite the volume of research in this area, current understanding is limited regarding the performance of landmark selection techniques in the presence of class-imbalanced data sets that are becoming increasingly prevalent in many applications. Hence, this paper provides a comprehensive empirical investigation using several real-world imbalanced data sets, including scientific data, by evaluating the quality of approximate low-rank decompositions and examining their influence on the accuracy of downstream tasks. Furthermore, we present a new landmark selection technique called Distance-based Importance Sampling and Clustering (DISC), in which the relative importance scores are computed for improving accuracy-efficiency tradeoffs compared to existing works that range from probabilistic sampling to clustering methods. The proposed landmark selection method follows a coarse-to-fine strategy to capture the intrinsic structure of complex data sets, allowing us to substantially reduce the computational complexity and memory footprint with minimal loss in accuracy.

Understanding mass spectrometry images: complexity to clarity with machine learning

The application of artificial intelligence and machine learning to hyperspectral mass spectrometry imaging (MSI) data has received considerable attention over recent years. Various methodologies have shown great promise in their ability to handle the complexity and size of MSI data sets. Advances in this area have been particularly appealing for MSI of biological samples, which typically produce highly complicated data with often subtle relationships between features. There are many different machine learning approaches that have been applied to MSI data over the past two decades. In this review, we focus on a subset of non-linear machine learning techniques that have mostly only been applied in the past 5 years. Specifically, we review the use of the self-organizing map (SOM), SOM with relational perspective mapping (SOM-RPM), t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP). While not their only functionality, we have grouped these techniques based on their ability to produce what we refer to as similarity maps. Similarity maps are color representations of hyperspectral data, in which spectral similarity between pixels-that is, their distance in high-dimensional space-is represented by relative color similarity. In discussing these techniques, we describe, briefly, their associated algorithms and functionalities, and also outline applications in MSI research with a strong focus on biological sample types. The aim of this review is therefore to introduce this relatively recent paradigm for visualizing and exploring hyperspectral MSI, while also providing a comparison between each technique discussed.

Integrating hyperspectral imaging with machine learning techniques for the high-resolution mapping of soil nitrogen fractions in soil profiles

Soil nitrogen (N) plays a central role in soil quality and biogeochemical cycles. However, little is known about the distribution and spatial variability of the different fractions of soil N within entire soil profiles. This study aimed to investigate the potential of laboratory-based hyperspectral imaging (HSI) spectroscopy to retrieve and map total N (TN), available N (AvailN), ammonium N (NH4-N), nitrate N (NO3-N), and microbial biomass N (MBN) in soil profiles at a high resolution. HSI images of eleven intact soil profiles of 100 ± 5 cm depth from three typical soil types were recorded. A variety of nonlinear machine learning techniques, such as artificial neural networks (ANN), cubist regression tree (Cubist), k-nearest neighbour (KNN), support vector machine regression (SVMR) and extreme gradient boosting (XGBoost), were compared with a partial least squares regression (PLSR) to determine the most suitable model for the prediction of the various soil N fractions. Overall, the results showed that nonlinear techniques performed better than PLSR in most cases, with a high coefficient of determination (R2) and low root mean square error (RMSE). Among the models, SVMR was found to be superior to the other tested models for TN (R2P = 0.94, RMSEP = 0.17 g kg−1), AvailN (R2P = 0.94, RMSEP = 13.35 mg kg−1), NO3-N (R2P = 0.82, RMSEP = 7.31 mg kg−1), and NH4-N (R2P = 0.70, RMSEP = 1.51 mg kg−1) based on independent validation, whereas MBN (R2P = 0.63, RMSEP = 6.62 mg kg−1) was predicted best by KNN. In addition, SVMR required less computational time and was less sensitive to spectral noise. It can therefore be concluded that HSI spectroscopy combined with SVMR is suitable for the high-resolution mapping of various soil N fractions in soil profiles.

On the Differential Analysis of Enterprise Valuation Methods as a Guideline for Unlisted Companies Assessment (I): Empowering Discounted Cash Flow Valuation

The Discounted Cash Flow (DCF) method is probably the most extended approach used in company valuation, its main drawbacks being probably the known extreme sensitivity to key variables such as Weighted Average Cost of Capital (WACC) and Free Cash Flow (FCF) estimations not unquestionably obtained. In this paper we propose an unbiased and systematic DCF method which allows us to value private equity by leveraging on stock markets evidences, based on a twofold approach: First, the use of the inverse method assesses the existence of a coherent WACC that positively compares with market observations; second, different FCF forecasting methods are benchmarked and shown to correspond with actual valuations. We use financial historical data including 42 companies in five sectors, extracted from Eikon-Reuters. Our results show that WACC and FCF forecasting are not coherent with market expectations along time, with sectors, or with market regions, when only historical and endogenous variables are taken into account. The best estimates are found when exogenous variables, operational normalization of input space, and data-driven linear techniques are considered (Root Mean Square Error of 6.51). Our method suggests that FCFs and their positive alignment with Market Capitalization and the subordinate enterprise value are the most influencing variables. The fine-tuning of the methods presented here, along with an exhaustive analysis using nonlinear machine-learning techniques, are developed and discussed in the companion paper.

Assessment of coastal geomorphological changes using multi-temporal Satellite-Derived Bathymetry

The present study demonstrates the usability of Satellite-Derived Bathymetry (SDB) to understand the geomorphological changes that have occurred in a coastal region located along Puducherry, India, where a beach restoration project was taken up in 2017 to arrest the shoreline erosion that is prevalent due to installation of hard structures. In the study, multi-temporal bathymetry data is generated by applying a non-linear machine learning technique of Support Vector Regression (SVR) on Landsat 8 OLI satellite datasets of 30 m resolution. The empirically driven SVR is calibrated and validated using eco-sounder data collected during field measurement campaigns and fairly accurate SDBs with Root Mean Square Errors and Mean Absolute Errors ranging between 0.40-1.07 m and 0.31–0.85 m, respectively are obtained. Subsequently, the derived temporal depth maps are studied to understand the morphological changes that have occurred in this coastal stretch and the results clearly show the development of a beach, north of the pier, followed by the stabilization of the coastline. The outcomes are further validated through an independent ArcGIS- DSAS based shoreline change analysis which suggests similar trends of accretion and erosion as observed through the bathymetry change analysis. The study thus substantiates that the beach restoration step has yielded positive results between 2017 and 2018 and throws light on the significance of SDBs in coastal monitoring, modelling and assessment.

Nonlinear Machine Learning Models for Insulin Bolus Estimation in Type 1 Diabetes Therapy.

Type 1 diabetes (T1D) therapy requires multiple daily insulin injections to compensate the lack of endogenous insulin production due to β-cells destruction. An empirical standard formula (SF) is commonly used for such a task. Unfortunately, SF does not include information on glucose dynamics, e.g. the glucose rate-of-change (ROC) provided by continuous glucose monitoring (CGM) sensor. Hence, SF can sometimes lead to under/overestimations that can cause critical hypo/hyperglycemic episodes during/after the meal. Recently, to overcome this limitation, we proposed new linear regression models, integrating ROC information and personalized features. Despite the first encouraging results, the nonlinear nature of the problem calls for the application of nonlinear models. In this work, random forest (RF) and gradient boosting tree (GBT), nonlinear machine learning methodologies, were investigated. A dataset of 100 virtual subjects, opportunely divided into training and testing sets, was used. For each individual, a single-meal scenario with different meal conditions (preprandial ROC, BG and meal amounts) was simulated. The assessment was performed both in terms of accuracy in estimating the optimal bolus and glycemic control. Results were compared to the best performing linear model previously developed. The two tree-based models proposed lead to a statistically significant improvement of glycemic control compared to the linear approach, reducing the time spent in hypoglycemia (from 32.49% to 27.57-25.20% for RF and GBT, respectively). These results represent a preliminary step to prove that nonlinear machine learning techniques can improve the estimation of insulin bolus in T1D therapy. Particularly, RF and GBT were shown to outperform the previously linear models proposed.Clinical Relevance- Insulin bolus estimation with nonlinear machine learning techniques reduces the risk of adverse events in T1D therapy.

Analyzing driving factors of land values in urban scale based on big data and non-linear machine learning techniques

Land value plays a vital role in the real estate market. It is a critical reference for urban planners to reallocate land resources and introduce valid policies. Studying the influential factors on land value can help better understand the spatial-temporal variation of land values and design effective control policies. This attracted a number of scholars to study the spatial and temporal relationships between land value and its possible influential factors from the perspective of macro and micro. However, the majority of the existing studies have the problems of linear assumption and multicollinearity in research models. Limited features and the lack of feature selection procedure are another two commonly seen limitations. To overcome the gaps, this paper adopts non-linear machine learning (ML) methods to investigate the influential factors on land values per square foot based on “big data” in New York City. More than one thousand potential factors are considered, covering from the land attribute, point of interest, demographics, housing, to economic, education, and social. They are further selected using a feature extraction model named Recursive Feature Elimination (RFE). Six ML algorithms, including Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Multi Linear Regression (MLR), Linear Support Vector Regression (SVR), Multilayer Perceptron (MLP) Regression, and K-Nearest Neighbor (KNN) Regression are evaluated and compared. The optimal one with an R-square value of 0.933 is used to calculate the feature importance further. Several important impact features are disclosed, including the number of newsstands, and the vacant housing percentage.