Multivariate Adaptive Regression Splines Model Research Articles

Investigating Landfill Leachate and Groundwater Quality Prediction Using a Robust Integrated Artificial Intelligence Model: Grey Wolf Metaheuristic Optimization Algorithm and Extreme Learning Machine

The likelihood of surface water and groundwater contamination is higher in regions close to landfills due to the possibility of leachate percolation, which is a potential source of pollution. Therefore, proposing a reliable framework for monitoring leachate and groundwater parameters is an essential task for the managers and authorities of water quality control. For this purpose, an efficient hybrid artificial intelligence model based on grey wolf metaheuristic optimization algorithm and extreme learning machine (ELM-GWO) is used for predicting landfill leachate quality (COD and BOD5) and groundwater quality (turbidity and EC) at the Saravan landfill, Rasht, Iran. In this study, leachate and groundwater samples were collected from the Saravan landfill and monitoring wells. Moreover, the concentration of different physico-chemical parameters and heavy metal concentration in leachate (Cd, Cr, Cu, Fe, Ni, Pb, Mn, Zn, turbidity, Ca, Na, NO3, Cl, K, COD, and BOD5) and in groundwater (Cd, Cr, Cu, Fe, Ni, Pb, Mn, Zn, turbidity, EC, TDS, pH, Cl, Na, NO3, and K). The results obtained from ELM-GWO were compared with four different artificial intelligence models: multivariate adaptive regression splines (MARS), extreme learning machine (ELM), multilayer perceptron artificial neural network (MLPANN), and multilayer perceptron artificial neural network integrated with grey wolf metaheuristic optimization algorithm (MLPANN-GWO). The results of this study confirm that ELM-GWO considerably enhanced the predictive performance of the MLPANN-GWO, ELM, MLPANN, and MARS models in terms of the root-mean-square error, respectively, by 43.07%, 73.88%, 74.5%, and 88.55% for COD; 23.91%, 59.31%, 62.85%, and 77.71% for BOD5; 14.08%, 47.86%, 53.43%, and 57.04% for turbidity; and 38.57%, 59.64%, 67.94%, and 74.76% for EC. Therefore, ELM-GWO can be applied as a robust approach for investigating leachate and groundwater quality parameters in different landfill sites.

Correlates and Predictors of NT-proBNP in Life Insurance Applicants.

-To document the various laboratory and demographic/historical correlates of NT-proBNP levels in applicants for life insurance, and to explore the accuracy of a prediction model based on those variables. -NT-proBNP blood test results were obtained from 1.34 million insurance applicants between the age of 50 and 85 years, beginning in 2003. Exploratory data analysis was carried out to document correlations with other laboratory variables, sex, age, and the presence of relevant diseases. Further, predictive models were used to quantify the proportion of the variance of NT-proBNP, which can be explained by a combination of these other, easier to determine variables. -NT-proBNP shows the expected, negative correlation with estimated glomerular filtration rate (eGFR) is markedly higher in those with a history of heart disease and is somewhat higher in those with a history of hypertension. A strong, unexpected, negative correlation between NT-proBNP and albumin was discovered. Of the variables evaluated, a multivariate adaptive regression spline (MARS) model automated selection procedure selected 7 variables (age, sex, albumin, eGFR, BMI, systolic blood pressure, cholesterol, and history of heart disease). Variable importance evaluation determined that age, albumin and eGFR were the 3 most important continuous variables in the prediction of NT-proBNP levels. An ordinary least squares (OLS) model using these same variables achieved a R-squared of 24.7%. -Expected ranges of NT-proBNP may vary substantially depending on the value of other variables in the prediction equation. Albumin is significantly negatively correlated with NT-proBNP levels. The reasons for this are unclear.

ANOVA Decomposition and Importance Variable Process in Multivariate Adaptive Regression Spline Model

This article reviews one of the non-parametric functions, namely the MARS (Multivariate Adaptive Regression Spline) method, a complex combination of recursive partitioning and spline regression. The many advantages of the MARS function over other non-parametric regression functions are of interest to researchers. One of them is it can accommodate the additive and interaction model to improve the prediction and interpretation of the data. There are some important things in the MARS method, namely, ANOVA decomposition and Importance Variable. Decomposition ANOVA is a technique in MARS that is useful for grouping basis function based on variables engagement, whether they enter by one variable or interactions with other variables, making it easier to interpret in graphical form. In comparison, the important variables are a technique that can be used to determine which predictor variables most influence the MARS modeling. This study assesses ANOVA decomposition, and the important variables process in MARS modeling based on GCV and MSE criteria. We use the poverty rate modeling data on Java Island to implement the study results. The results show that the MARS model's interpretation of the poverty rate can be better done through ANOVA decomposition. Besides that, based on GCV and MSE criteria, the result also shows that the biggest variable importance in poverty rate modeling on Java Island is the percentage of per capita expenditure for food, while the smallest is the economic growth variable.

Estimation of Multivariate Adaptive Regression Splines (MARS) Model Parameters by Using Generalized Least Square (GLS) Method

Estimation of Multivariate Adaptive Regression Splines (MARS) Model Parameters by Using Generalized Least Square (GLS) Method

Modeling Human Development Index in Papua and West Sumatera with Multivariate Adaptive Regression Spline

The Human Development Index (HDI), which is an indicator as a comprehensive description of the measure of achievment development in the quality of human life in a region. In Indonesia, there are still many areas with low HDI, especially in Papua Province. This study aims to model and find out what factors affect HDI in Papua Province and West Sumatera Province, using Multivariate Adaptive Regression Spline (MARS). MARS is a modeling methods that can handle high-dimensional data, namely data that has independent variables and a sampel size of data with unknown data patterns in advance, and can be applied to see interaction between the variables used. The result of this study obtained that the best MARS model for Papua Province is a combination of BF=24, MI=2, and MO=0 with GCV=0.55953. while the best MARS model for West Sumatera Province ia a combination of BF=24, Mi=2, and MO=0 with GCV=0.02697. the factors that significantly affect HDI in Papua Province and West Sumatera Province are average lenght of schooling (X2), adjusted per-capita income (X6), life expectancy (X1), percentage of poor population (X4), anf gross regional domestic product (X3). The level of importance of each variable for Papua Province is 100%, 45.26%, 29.24%, 6.55%, and 6.27%. while for West Sumatera Province it is 100%, 96.73%, 57.54%, 34.13%, and 29.6%,respectively. So that in this case based on research results the average lenght schooling (X2) is the variable that most influences HDI in the two regions.

Development of hybrid computational data-intelligence model for flowing bottom-hole pressure of oil wells: New strategy for oil reservoir management and monitoring

Among several metric parameters concerning the assessment of oil and gas well production, the flowing bottom-hole pressure (FBHP) is considered essential. Accurate prediction of FBHP is crucial for petroleum engineering and management. Several related parameters are associated with the FBHP magnitude influence; thus, proper inspection of those parameters is another vital concern. This research proposes a hybrid modeling framework based on the hybridization of machine learning (ML) models (i.e., Extreme Learning Machine (ELM), Support Vector Machine Regressor (SVR), Extreme Gradient Boosting (XGB), and Multivariate Adaptive Regression Spline (MARS)) and nature-inspired Differential Evolutional (DE) optimization for FBHP prediction. The adjustment of the internal parameters of the ML-based models and the input feature selection is formulated as an incremental learning problem that is solved by the evolutionary algorithm. Problem-specific samples were collected from the open-source literature for this investigation. Modeling results are adaptable, automatically determining the most relevant variables for the context of the ML model. The adaptive polynomial structure of hybridized MARS model attained the best average performance for the FBHP modeling with correlation (R = 0.94) and minimum root mean square (RMSE = 97.88). The proposed modeling framework produces an alternative efficient computer aid model for FBHP prediction, resulting in reliable automated technology to assist oil and gas well management.

Bankruptcy Prediction for Micro and Small Enterprises Using Financial, Non-Financial, Business Sector and Macroeconomic Variables: The Case of the Lithuanian Construction Sector

Credit-risk models that are designed for general application across sectors may not be suitable for the construction industry, which has unique characteristics and financial risks that require specialised modelling approaches. Moreover, advanced bankruptcy-prediction models are often used to achieve the highest accuracy in large modern datasets. Therefore, the aim of this research is the creation of enterprise-bankruptcy prediction (EBP) models for Lithuanian micro and small enterprises (MiSEs) in the construction sector. This issue is analysed based on classification models and the specific types of variable used. Firstly, four types of variable are proposed. In EBP models, financial variables substantially explain an enterprise’s financial statements and performance from different perspectives. Including enterprises’ non-financial, construction-sector and macroeconomic variables improves the characteristics of EBP models. The inclusion of macroeconomic variables in the model has a particularly significant impact. These findings can be of great significance to investors, creditors, policymakers and practitioners in assessing financial risks and making informed decisions. The second question is related to the classification models used. To develop the EBP models, logistic regression (LR), artificial neural networks (ANNs) and multivariate adaptive regression splines (MARS) were used. In addition, this study developed two-stage hybrid models, i.e., the LR is combined with ANNs. The findings show that two-stage hybrid models do not improve bankruptcy prediction. It cannot be argued that ANN models are more accurate in predicting bankruptcy. The MARS model demonstrates the best bankruptcy prediction, i.e., this model could be a valuable tool for stakeholders to evaluate enterprises’ financial risk.

Hydrogen solubility in ionic liquids: Application of a structure-based deep learning approach and equations of state

Following the increasing threat to the environment, the application of hydrogen as an energy carrier is one of the solutions that has received much attention. Considering the importance of gas solubility in separation and gasification processes, the solubility of hydrogen in various solvents has particular importance. In this study, an extensive database containing 580 experimental data points (13 ionic liquids (ILs)) in a vast range of temperature (278.2–453.15 K) and pressure (0.433–552 bar) was employed to determine the solubility of hydrogen in ILs using intelligent models. In this regard, four intelligent models, including Random Forest (RF), Adaptive Boosting-Support Vector for Regression (AdaBoost-SVR), Deep Belief Network (DBN), and Multivariate Adaptive Regression Splines (MARS) models, were developed with two different approaches. In the first method, the chemical substructures were considered as inputs; in the second manner, the thermodynamic attributes of ILs were considered as inputs. Temperature and pressure were also two inputs of both methods. The results show that in both methods, the DBN paradigm has the best proficiency. The best root mean square error (RMSE) and coefficient of determination (R2) values in the former way were 0.00106 and 0.9991, respectively, and in the latter way, 0.00066 and 0.9996, respectively. The findings of sensitivity analysis show that among the chemical substructures and thermodynamic properties, –CH3 and pressure, with absolute relevancy factor values of 0.425 and 0.517, respectively, have the most impression on the hydrogen solubility in ILs. Also, the investigation of the influence of different parameters on hydrogen solubility shows that raising the pressure and alkyl chain length increases the hydrogen solubility in ILs. The acquired findings were also compared with equations of state (EOSs), and it was found that the intelligent models have better performance and high efficiency than EOSs. Finally, to validate the model data, the leverage method was used, which shows that over 96.7% of the data is in the authentic domain and only about 1% of the data is in the suspected data region.

Evaluation of River Water Quality Index Using Remote Sensing and Artificial Intelligence Models

To restrict the entry of polluting components into water bodies, particularly rivers, it is critical to undertake timely monitoring and make rapid choices. Traditional techniques of assessing water quality are typically costly and time-consuming. With the advent of remote sensing technologies and the availability of high-resolution satellite images in recent years, a significant opportunity for water quality monitoring has arisen. In this study, the water quality index (WQI) for the Hudson River has been estimated using Landsat 8 OLI-TIRS images and four Artificial Intelligence (AI) models, such as M5 Model Tree (MT), Multivariate Adaptive Regression Spline (MARS), Gene Expression Programming (GEP), and Evolutionary Polynomial Regression (EPR). In this way, 13 water quality parameters (WQPs) (i.e., Turbidity, Sulfate, Sodium, Potassium, Hardness, Fluoride, Dissolved Oxygen, Chloride, Arsenic, Alkalinity, pH, Nitrate, and Magnesium) were measured between 14 March 2021 and 16 June 2021 at a site near Poughkeepsie, New York. First, Multiple Linear Regression (MLR) models were created between these WQPs parameters and the spectral indices of Landsat 8 OLI-TIRS images, and then, the most correlated spectral indices were selected as input variables of AI models. With reference to the measured values of WQPs, the WQI was determined according to the Canadian Council of Ministers of the Environment (CCME) guidelines. After that, AI models were developed through the training and testing stages, and then estimated values of WQI were compared to the actual values. The results of the AI models’ performance showed that the MARS model had the best performance among the other AI models for monitoring WQI. The results demonstrated the high effectiveness and power of estimating WQI utilizing a combination of satellite images and artificial intelligence models.

Predicting reference evapotranspiration based on hydro-climatic variables: comparison of different machine learning models

ABSTRACT This paper aimed to estimate the reference evapotranspiration (ET0) due to some limitations of the Food and Agriculture Organization-56 Penman-Monteith (FAO 56-PM) approach by using five alternative machine learning models. The study makes an important contribution to the ET0 estimation success for of the ET0 of 12 stations with variable climate characteristics in the Central Anatolian Region (CAR). The performances of the models were compared with the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) metrics that are frequently cited in the literature, and also with the performance index (PI). Long short-term memory (LSTM), artificial neural networks (ANN), and multivariate adaptive regression splines (MARS) models provided the best performance in eight, three, and one stations, respectively. The R2, MAE, RMSE, and PI values of the selected models from each station vary in the range of 0.987-0.999, 1.948-4.567, 2.671-6.659, and 1.544-4.018, respectively.

Predicting Earthquake-Induced Landslides by Using a Stochastic Modeling Approach: A Case Study of the 2001 El Salvador Coseismic Landslides

In January and February 2001, El Salvador was hit by two strong earthquakes that triggered thousands of landslides, causing 1259 fatalities and extensive damage. The analysis of aerial and SPOT-4 satellite images allowed us to map 6491 coseismic landslides, mainly debris slides and flows that occurred in volcanic epiclastites and pyroclastites. Four different multivariate adaptive regression splines (MARS) models were produced using different predictors and landslide inventories which contain slope failures triggered by an extreme rainfall event in 2009 and those induced by the earthquakes of 2001. In a predictive analysis, three validation scenarios were employed: the first and the second included 25% and 95% of the landslides, respectively, while the third was based on a k-fold spatial cross-validation. The results of our analysis revealed that: (i) the MARS algorithm provides reliable predictions of coseismic landslides; (ii) a better ability to predict coseismic slope failures was observed when including susceptibility to rainfall-triggered landslides as an independent variable; (iii) the best accuracy is achieved by models trained with both preparatory and trigger variables; (iv) an incomplete inventory of coseismic slope failures built just after the earthquake event can be used to identify potential locations of yet unreported landslides.

Streamflow prediction using machine learning models in selected rivers of Southern India

ABSTRACT The need for adequate data on the spatial and temporal variability of freshwater resources is a significant challenge to the water managers of the world in water resource planning and management. The problems will be acute in the coming years because of the increase in frequency and intensity of hydrologic extremes due to climate change. Therefore, streamflow prediction has become an important area of research because of its importance in flood mitigation, reservoir operation, and water resource management. In this paper, we have tested four Machine Learning models (ML models): Support Vector Machines (SVM), Random Forest (RF), Long Short-Term Memory (LSTM), and Multivariate Adaptive Regression Splines (MARS) for streamflow prediction at daily and monthly time scales in three rivers draining in the different climatic and geological settings. The SVM, RF, LSTM, and MARS models have been trained and tested in the Suvarna, Aghanashini, and Kunderu River Basins in peninsular India. Model intercomparison was made to identify the best suitable model for streamflow prediction. The RF outperforms other models for daily streamflow, and MARS outperforms other models for monthly streamflow prediction in the Suvarna river with Nash-Sutcliffe efficiency (NSE) values of 0.676 and 0.924, respectively. SVM (NSE = 0.741) and RF (NSE = 0.826) are found to be the best models for daily and monthly streamflow prediction in the Aghanashini river. MARS outperformed other models in the case of high, severe, and extreme flow simulation with NSE values of 0.481, 0.374, and 0.455, respectively, in the Aghanashini river. Other hydrological variables (groundwater level data, antecedent soil moisture, potential evapotranspiration data) and a better spatial resolution of rainfall data can be used to develop more accurate machine-learning models for streamflow predictions.

Prediction of Sediment Yields Using a Data-Driven Radial M5 Tree Model

Reliable estimations of sediment yields are very important for investigations of river morphology and water resources management. Nowadays, soft computing methods are very helpful and famous regarding the accurate estimation of sediment loads. The present study checked the applicability of the radial M5 tree (RM5Tree) model to accurately estimate sediment yields using daily inputs of the snow cover fraction, air temperature, evapotranspiration and effective rainfall, in addition to the flow, in the Gilgit River, Upper Indus Basin (UIB) tributary, Pakistan. The results of the RM5Tree model were compared with support vector regression (SVR), artificial neural network (ANN), multivariate adaptive regression spline (MARS), M5Tree, sediment rating curve (SRC) and response surface method (RSM) models. The resulting accuracy of the models was assessed using Pearson’s correlation coefficient (R2), the root-mean-square error (RMSE) and the mean absolute percentage error (MAPE). The prediction accuracy of the RM5Tree model during the testing period was superior to the ANN, MARS, SVR, M5Tree, RSM and SRC models with the R2, RMSE and MAPE being 0.72, 0.51 tons/day and 11.99%, respectively. The RM5Tree model predicted suspended sediment peaks better, with 84.10% relative accuracy, in comparison to the MARS, ANN, SVR, M5Tree, RSM and SRC models, with 80.62, 77.86, 81.90, 80.20, 74.58 and 62.49% relative accuracies, respectively.

Calculating soil titratable acidity from routine soil analyses of cultivated South African soils using various model structures

Soil titratable acidity (TA) is determined through extraction with an acetate-buffered salt solution (potassium sulphate, K2SO4) and is routinely used to predict the soil lime requirement using the Eksteen method. Soil TA is not always reported by commercial soil testing laboratories, but rather exchangeable acidity (1M KCl) is determined. Therefore, a need exists to use other routinely determined soil properties in a model to derive TA. In this study, routine soil-analysis results from a commercial laboratory (n = 5 915 measurements) were used to generate various model structures to calculate a theoretical value of TA and to evaluate the accuracy thereof. Measured TA was found to be significantly correlated to the calculated TA using the Eksteen R-value, soil pH, organic carbon percentage, exchangeable Ca2+ and Mg2+, and a function of the interaction between pH and organic carbon. Soil TA could be calculated most accurately using these factors in multivariate adaptive regression splines model (r 2 = 0.69, mean absolute error = 0.16) at TA values of < 2 cmolc kg−1. However, the exponential model calculations remained stable over a larger range of actual TA values up to 4 cmolc kg−1. Given the ease of use and interpretability, it is recommended that an exponential function model is used to calculate TA.

MARSplines-Based Soil Moisture Sensor Calibration

Soil moisture sensor calibration based on the Multivariate Adaptive Regression Splines (MARSplines) model is studied in this paper. Different from the generic polynomial fitting methods, the MARSplines model is a non-parametric model, and it is able to model the complex relationship between the actual and measured soil moisture. Rao-1 algorithm is employed to tune the hyper-parameters of the calibration model and thus the performance of the proposed method is further improved. Data collected from four commercial soil moisture sensors is utilized to verify the effectiveness of the proposed method. To assess the calibration performance, the proposed model is compared with the model without using the temperature information. The numeric studies prove that it is promising to apply the proposed model for real applications.

MARS와 PAM을 활용한 순항 단계 고도 이탈 이상치 탐지

Recently, methods for analyzing and detecting precursor signs of aviation accidents through navigation data have been studied. However, most of the studies are limited to the takeoff and landing stages, and there are no rules specifying the cruise stage, so the cruise stage search studies are rare. At this stage, break away according to altitude and path rather than speed occur, and in this paper, we try to detect break away in altitude of cruise stage by clustering method. In the ADS-B flight data received from January to December 2019, 10% of flights from Jeju Airport to Gimpo Airport, the most frequently operated domestic flights, were sampled and studied about 4,400 flights. Through the Multivariate Adaptive Regression Splines (MARS) model and the simple linear regression model, the slope between the trajectories was calculated to define the cruise slope, an interval that can be defined as the cruise stage. In addition, the non-similarity between flights was calculated using the distance measurements, EDR and DTW. Two cluster methods were used: PAM, Hierarchical clustering, and flights were clustered through PAM with the largest average silhouette width. An outlier was determined through a non-similar value between the center of the cluster classified through PAM and the flights within the cluster, and about 12% of flights were detected. It is expected to be able to prevent aviation accidents by detecting altitude break away, one of the premonitory signs, earlier. In subsequent studies, outlier detection can be carried out using a Semi-supervised learning method to detect outliers based on normal data collected with the help of experts in the aviation field.

Crop Yield Prediction Using Hybrid Machine Learning Approach: A Case Study of Lentil (Lens culinaris Medik.)

This paper introduces a novel hybrid approach, combining machine learning algorithms with feature selection, for efficient modelling and forecasting of complex phenomenon governed by multifactorial and nonlinear behaviours, such as crop yield. We have attempted to harness the benefits of the soft computing algorithm multivariate adaptive regression spline (MARS) for feature selection coupled with support vector regression (SVR) and artificial neural network (ANN) for efficiently mapping the relationship between the predictors and predictand variables using the MARS-ANN and MARS-SVR hybrid frameworks. The performances of the algorithms are com-pared on different fit statistics such as RMSE, MAD, MAPE, etc., using numeric agronomic traits of 518 lentil genotypes to predict grain yield. The proposed MARS-based hybrid models outperformed individual models such as MARS, SVR and ANN. This is largely due to the enhanced feature ex-traction capability of the MARS model coupled with the nonlinear adaptive learning ability of ANN and SVR. The superiority of the proposed hybrid models MARS-ANN and MARS-SVM in terms of model building and generalisation ability was demonstrated.

Biomarkers, Socioeconomic Factors, and Right Ventricular Function After Surgical Repair for Tetralogy of Fallot.

Right ventricular (RV) dysfunction early after tetralogy of Fallot (TOF) increases post-operative morbidity. We investigated associations of circulating biomarkers and socioeconomic factors with early post-operative RV systolic function. Single-center prospective cohort study of infants undergoing TOF repair. Six serologic biomarkers of myocardial fibrosis and wall stress collected at the time of surgery were measured with immunoassay. Geocoding was performed for socioeconomic factors. Multivariate adaptive regression splines (MARS) models identified factors associated with RV function parameters: fractional area change (FAC), global longitudinal strain and strain rate, andfree wall strain and strain rate. Seventy-one patients aged 3.5months (IQR 2.4, 5.2) were included. Galectin-3 was the highest ranked predictor for FAC, global longitudinal strain, and free wall strain, and procollagen type-I carboxy-terminal propeptide (PICP) was the highest ranked predictor for global longitudinal strain rate and free wall strain rate. Several neighborhood characteristics were also highly ranked. Models adjusted R2 ranged from 0.71 to 0.85 (FAC, global longitudinal strain/strain rate), and 0.55-0.57 (RV free wall strain/strain rate). A combination of serologic biomarkers, socioeconomic, and clinical variables explain a significant proportion of the variability in RV function after TOF repair. These factors may inform pre-operative risk-stratification for these patients.

Safety Performance Evaluation of Flashing Yellow Arrow: Time-of-Day Versus 24-Hour Operation

The flashing yellow arrow (FYA), as the permissive left-turn indication, has been widely implemented to mitigate driving confusion caused by the permissive steady green indication. The selection of an appropriate FYA operation type, however, remains an open-ended question as few studies have evaluated the safety effectiveness of FYA under different operation types, that is, 24-hour, time-of-day (TOD), and switching from 24-hour to TOD. Moreover, since FYA has been typically used at intersections with a single left-turn lane, its safety effectiveness at intersections with dual left-turn lanes has not been well investigated. To bridge these gaps, this paper focuses on evaluating the safety effectiveness of FYA at intersections with varying numbers of left-turn lanes under different operation types and examining whether the FYA operation type switch affects safety performance. Empirical Bayes before/after studies are conducted for 24-hour, TOD, and switching from 24-hour to TOD FYA operation types using multivariate adaptive regression splines models in comparison with traditional negative binomial models. Safety performance functions are developed for different combinations of crash types and a different number of left-turn lanes. Results show that for intersections with either a single left-turn lane or dual left-turn lanes, 24-hour and TOD FYA operation types reduce crashes by 8.76% to 50%. However, intersections with dual left-turn lanes experience a 31.2% increase in total crashes when switching from 24-hour to TOD FYA operation type, while intersections with a single left-turn lane see a 60% decrease in rear-end crashes.

A methodology for prioritizing safety indicators using individual vehicle trajectory data

A methodology for assessing crash risk using vehicle driving trajectories based on data mining techniques was developed in this study. A variety of safety indicators reflecting the characteristics of traffic and road geometric conditions were evaluated in terms of their capability of capturing hazardous traffic flow. Comprehensive data preparation was conducted by matching driving trajectory data obtained from in-vehicle digital tachograph devices and crash data to classify and analyze hazardous and normal traffic flows. The random forest approach was adopted to quantify the importance of safety indicators. The crash risks were evaluated using the logistic regression model and multivariate adaptive regression splines model based on the set of safety indicators with high importance. The results show that the dangerous driving events rate and driving volatility indicators were found to be particularly significant in identifying hazardous conditions. The multivariate adaptive regression splines model showed better performance and a classification accuracy of 86% was achieved. The proposed methodology will be useful for deriving effective countermeasures to prevent crashes, which is the backbone of proactive traffic safety management.