Developing an Optimal Ensemble Model to Estimate Building Demolition Waste Generation Rate

Abstract

Smart management of construction and demolition (C&D) waste is imperative, and researchers have implemented machine learning for estimating waste generation. In Korea, the management of demolition waste (DW) is important due to old buildings, and it is necessary to predict the amount of DW to manage it. Thus, this study employed decision tree (DT)-based ensemble models (i.e., random forest—RF, extremely randomized trees—ET, gradient boosting machine—GBM), and extreme gradient boost—XGboost) based on data characteristics (i.e., small datasets with categorical inputs) to predict the demolition waste generation rate (DWGR) of buildings in urban redevelopment areas. As a result of the study, the RF and GBM algorithms showed better prediction performance than the ET and XGboost algorithms. Especially, RF (6 features, 450 estimators; mean, 1169.94 kg·m⁻²) and GBM (4 features, 300 estimators; mean, 1166.25 kg·m⁻²) yielded the top predictive performances. In addition, feature importance affecting DWGR was found to have a significant impact on the order of gross floor area (GFA) > location > roof material > wall material. The straightforward collection of features used here can facilitate benchmarking as a decision-making tool in demolition waste management plans for industry stakeholders and policy makers. Therefore, in the future, it is required to improve the predictive performance of the model by updating additional data and building a reliable dataset.

1. Introduction

Waste management-related issues are rising due to rapid urban growth [1]. Increasing populations are translating into growing housing demands, rapid town growth, and ultimately greater waste generation [2]. The construction industry was estimated to account for 35% of the total waste generation [3] and the considerable amounts of solid waste and greenhouse gases emitted from construction and demolition (C&D), as well as refurbishment, are posing a serious challenge to global environments [4]. C&D waste generation has tended to increase steadily [5,6,7], with 70–90% of this total attributed to demolition waste [8,9]. Accordingly, appropriate C&D waste management is essential for urban sustainability, with the specific need for achieving maximum economic and environmental values during building demolition [10]. To this end, accurate data regarding waste generation amounts are required to estimate the scale of waste generation, economic values, costs, and environmental impacts [11]. Moreover, C&D waste generation data can be used to provide essential information on waste management to the relevant industry stakeholders (e.g., clients, architects, engineers, contractors, planners, etc.; [12]).

Smart management of C&D waste is an essential component of modern information and communication technology. Accordingly, numerous researchers have implemented artificial intelligence (AI) technology for predicting C&D waste generation. In particular, early research on waste generation using machine learning (ML) has primarily focused on the development of a single algorithm predictive model. For example, Jalali and Nouri [13], Milojkovic and Litovski [14], Noori et al. [15], and Patel and Meka [16] developed a municipal solid waste (MSW) generation predictive model by applying artificial neural networks (ANNs). Noori et al. [17] and Dai et al. [18] developed a C&D waste generation predictive model using support vector machine (SVM) algorithms; whereas other analyses have employed linear regression (LR) to predict MSW [19,20,21,22].

The selection of proper ML algorithms suitable for the data characteristics, the application of data preprocessing and verification methods, as well as the selection of proper hyper-parameters, are essential for developing an optimal ML predictive model [23,24,25,26]. Recently, studies have employed various other algorithms and adjusted hyper-parameters to derive an optimal ML model (Table 1). Song et al. [10] developed a hybrid predictive model by combining the gray model (GM) and support vector machine (SVR) to predict C&D waste generation across 31 cities in China. The annual total area of construction (ATAC) from 2015 to 2018 was used as input variables, and the average value of the relative percent error of the GM-SVR model was <0.1 (i.e., good performance). Johnson et al. [27] developed a predictive model for MSW (refuse, paper, and metal, glass, plastic—MGP) generation in New York City using weekly MSW data from 2003 to 2015 and 28 features. Here, a gradient boosting machine (GBM) algorithm was used, the best predictive performance was achieved when all features were utilized. The coefficients of determination (R²) for the model of refuse, paper, and MGP according to the GBM algorithm were 0.889–0.906, 0.744–0.791, and 0.685–0.694, respectively. Kontokosta et al. [23] constructed 41,412 timeseries datasets from 609 New York City Department of Sanitation subsections from 2013 to 2016 and developed a predictive model for total waste, refuse, and MGP by applying 31 features in a GBM algorithm. In addition, hyper-parameters (number of trees, 200; tree depth, 6; learning rate, 0.1) were adjusted before applying the algorithm, and the R² of the GBM models were 0.87, 0.87, 0.73, and 0.78, respectively. Kumar et al. [28] predicted plastic generation rates by collecting data from 120 households, where ANN, random forest (RF), and SVM algorithms were employed, along with four features (education, occupation, income, and house type). As for the predictive performance of the model, the ANN (R² = 0.75) and SVM (R² = 0.74) outperformed RF (R² = 0.66). Kannangara et al. [29] developed a waste generation predictive model for MSW (1553 samples) and paper datasets (1867 samples) using eight socio-economic features, and ANN–DT algorithms were used, producing R² values of 0.54, 0.72, 0.31, and 0.35 for MSW-DT, MSW-ANN, paper-DT, and paper-ANN, respectively, concluding that the low predictive performance of DT was due to the characteristics of a single model. Alternatively, Lu et al. [30] applied multiple linear regression (MLR), GM, ANN, and DT analyses to predict construction waste generation based on five features—population, GDP per capita, total construction output, floor space of newly started buildings, and floor space of completed buildings completed—across 43 datasets. R² of the test models were 0.977, 0.918, 0.777, and 0.764 for GM, ANN, MLR, and DT, respectively. Akanbi et al. [24] developed models for building-level recycle, reuse, and landfill waste generation using deep neural network (DNN) algorithms. A dataset constructed from demolition records of 2280 buildings and five features—gross floor area (GFA), building volume, number of floors, building archetype, and usage type—were used for developing the model, producing R² values of 0.9475, 0.9789, and 0.9944 for recycle, reuse, and landfill DNN, respectively, corresponding to high prediction performance. Ghanbari et al. [31] developed a municipal solid waste generation (MSWG) predictive model based on timeseries data of MSWG, and four features (income, population, GDP, and month) of Tehran, Iran from 1991 to 2013. Here, ANN, RF, multivariate adaptive regression splines (MARS), and MARS-crow search algorithm (CSA) were applied, and the predictive performance of the MARS-CSA model (R² = 0.90) was superior to that of MARS (R² = 0.88), ANN (R² = 0.74), and RF (R² = 0.77). Nguyen et al. [25] developed a predictive model for MSW generation in residential areas of Vietnam, where the dataset (189 MSW samples) collected from 2015 to 2017 across nine features—urban population, total retail sales of consumer goods, average per capita monthly income, average per capita size of the home, population density, average per capita monthly consumption expenditure, total hospital beds, total residential land per province, and total solid waste collected per day) were used. The K-nearest neighbor (KNN), RF, and DNN algorithms were applied through hyper-parameters adjustment, and the resulting R² were 0.96, 0.97, and 0.91, respectively. Jayaraman et al. [26] developed an MSW predictive model using SARIMA (season ARIMA) and XGboost (extreme gradient boosting) algorithms, in conjunction with a timeseries dataset of MSW (1129 rows and 40 columns) from 2006 to 2018, revealing that XGboost (R² = 0.4145) was superior to SARIMA (R² = −0.8885). Moreover, the prediction performance of XGboost was improved by adjusting the tree number and max-depth. Namoun et al. [32] developed a predictive model for daily household waste generation using SVR, XGBoost, LightGBM, RF, extremely randomized trees (ETs), and ANN based on weekly waste generation data from 2011 to 2021, producing R² values of 0.692, 0.67, 0.745, 0.714, 0.7368, and 0.685, respectively.

Table 1. ML models applied to predict waste generation in previous studies.

Studies	Waste Type	Applied Algorithm	Performance
Song et al. [10]	C&D waste	GM-SVR	the average value of relative percent error < 0.1
Johnson et al. [27]	MSW	GBM	R2 value: 0.685–0.906
Kontokosta et al. [23]	MSW	GBM	R2 value: 0.73–0.87
Kumar et al. [28]	MSW	ANN	R2 value: 0.75
		RF	R2 value: 0.66
		SVM	R2 value: 0.74
Kannangara et al. [29]	MSW	ANN	R2 value: 0.72
		DT	R2 value: 0.54
	Paper	ANN	R2 value: 0.31
		DT	R2 value: 0.35
Lu et al. [30]	C&D waste	GM	R2 value: 0.977
		ANN	R2 value: 0.918
		MLR	R2 value: 0.777
		DT	R2 value: 0.764
Akanbi et al. [24]	C&D waste	DNN	R2 value: 0.948–0.994
Ghanbari et al. [31]	MSW	Multivariate adaptive regression splines (MARS)	R2 value: 0.90
		ANN	R2 value: 0.74
		RF	R2 value: 0.88
Nguyen et al. [25]	MSW	KNN	R2 value: 0.96
		RF	R2 value: 0.97
		DNN	R2 value: 0.91
Jayaraman et al. [26]	MSW	ARIMA	R2 value: −0.89
		XGboost	R2 value: 0.41
Namoun et al. [32]	Household waste	SVR	R2 value: 0.692
		XGboost	R2 value: 0.67
		LightGBM	R2 value: 0.745
		RF	R2 value: 0.714
		ET	R2 value: 0.737
		ANN	R2 value: 0.685

Data characteristics (e.g., types of independent and dependent variables, data size, etc.) used in the above studies for C&D and MSW generation prediction were diverse; whereas a single ML algorithm was mainly used in early studies, while these algorithms appear to be expanding recently, likely due to the improvement of ML algorithms, data processing (e.g., outlier and noise removal, data preprocessing, etc.), and verification methods (e.g., k-fold or leave one out cross-validation) must be varied according to data characteristics (e.g., types of independent and dependent variables, such as categorical or numerical variables) and environment (e.g., the size of data). In addition, recent studies have developed an optimized ML predictive model through the adjustment of proper hyper-parameters.

The purpose of the present study was to select proper ML algorithms for a relatively small dataset primarily composed of categorical variables and develop an ML model for predicting demolition waste generation (DWG) in the end-of-life stage of building to serve as decision-making support for proper waste management and plan establishment. Specifically, the detailed purposes were to: (i) apply proper ML algorithms for model design, (ii) evaluate the performance of various submodels, and (iii) derive an optimal demolition waste generation rate (DWGR) predictive model in consideration of the data characteristics.

Subsequent to this introductory Section 1, the remainder of the paper is organized as follows: Section 2 presents the methods and materials, together with a description of the data used; Section 3 describes the results of the study; Section 4 discusses several key points related to the findings; and, in Section 5, the main findings, applications, limitations, and future research of this study were discussed.

2. Methods and Materials

2.1. Data Source Description

The demolition waste (DW) generation (kg·m⁻²) records surveyed from 782 buildings in three redevelopment areas (Project A and B in Daegu, and Project C in Busan, Republic of Korea), within two cities were used in the present study. Table 2 presents the building status and statistical analysis of the collected data according to location and building characteristics. The dataset included information on six building features—location, structure, usage, gross floor area (GFA), as well as wall and roof materials—in addition to the corresponding building DWG rates. These building features correspond to the main factors affecting DWGR, and in this study, the six building features were used to estimate DWGR. Accordingly, the correlation between DWGR and six building features was expressed by Equation (1); whereas DWGR was defined by Equation (2):

DWGR = f (some or all of six features),

(1)

$$\mathrm{D}\mathrm{W}\mathrm{G}\mathrm{R}=\frac{\sum \mathrm{A} \mathrm{o}\mathrm{f} {\mathrm{b}\mathrm{u}\mathrm{i}\mathrm{l}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}}_{\mathrm{i}}}{\mathrm{G}\mathrm{F}\mathrm{A} \mathrm{o}\mathrm{f} {\mathrm{b}\mathrm{u}\mathrm{i}\mathrm{l}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}}_{\mathrm{i}}},$$

(2)

where DWGR is in kg·m⁻², A is the amount of a building (quantity in kg) and GFA is in m⁻².

Table 2. Building status and statistical analysis of raw data used in this study.

Category		Numbers	GFA (m2)				DWGR (kg·m−2)
			Total	Min	Mean	Max	Total	Min	Mean	Max
Location	1	343	31,542	21	92	275	450,310	298	1313	6034
	2	356	40,653	19	114	1127	485,037	83	1362	8574
	3	83	13,851	26	167	414	101,531	736	1223	1808
Usage	1	595	54,929	19	92	514	767,578	83	1290	8574
	2	172	28,706	22	167	1127	251,381	418	1462	5718
	3	15	2410	28	161	790	19,510	607	1301	2474
Structure	1	87	20,783	47	239	1127	169,538	418	1949	6034
	2	604	56,975	19	94	688	788,042	83	1305	8574
	3	91	8288	24	91	206	80,889	298	889	2237
Wall material	1	9	3693	48	410	1127	10,357	871	1151	4696
	2	236	32,584	23	138	790	391,259	252	1658	6034
	3	500	47,089	19	94	688	596,799	83	1194	8574
	4	37	2679	24	72	137	40,056	517	1083	2591
Roof material	1	289	43,565	21	151	1127	479,356	252	1659	6034
	2	33	4414	76	134	282	38,877	252	1178	1808
	3	178	12,439	23	70	206	227,923	306	1280	8574
	4	282	25,627	19	91	688	292,314	83	1037	2527

Location: 1—Project A, 2—Project B, 3—Project C. Structure: 1—Reinforced concrete (RC), 2—Masonry, 3—Wood. Usage: 1—Residential, 2—Residential and commercial, 3—Commercial. Wall material: 1—Concrete, 2—Brick, 3—Block, 4—Mud plastered and mortar wall. Roof material: 1—Slab, 2—Slab and roofing tile, 3—Slate, 4—Roofing tile.

2.2. Data Preprocessing and Dataset Size

To improve the predictive performance of the ML model, it is necessary to create a stable dataset. In this study, data preprocessing including outlier removal and standardization was performed to reduce data distortion and outliers’ impacts. Outliers were removed from the raw data, according to Equation (3), and the number of samples in the dataset after outlier removal was 690. The size of the dataset before and after data preprocessing and the change in DWGR descriptive statistics are shown in Table 3. Accordingly, ensemble predictive models were developed, and the data were standardized according to Equation (4) to create a dataset with the same scale:

Q1 − 1.5 × IQR < selecting data < Q3 + 1.5 × IQR,

(3)

where IQR is interquartile range, equal to Q3 minus Q1; and Q1 and Q3 are the 25th and 75th percentile, respectively.

$$x_{standardization}=\frac{x-\overline{x}}{\sigma },$$

(4)

where $x$ is the element of data, $\overline{x}$ is the average data value, and $\sigma$ is the standard deviation of the data.

Table 3. Changes in DWGR (kg·m⁻²) data statistics before and after data preprocessing.

Data Preprocessing	Number of Samples	Minimum	Maximum	Average	Median	Standard Deviation	Variance
Before	782	83.34	8573.79	1327.97	1162.25	809.2	654,032.4
After	690	298.30	3024.04	1165.04	1138.30	407.7	166,016.7

2.3. Applied Machine Learning Algorithms

The input variables used in this study consisted of categorical variables across five features—location, usage, structure, as well as wall and roof materials—and one numerical variable, GFA. Accordingly, the non-parametric DT algorithm selected can handle both categorical and numerical variables [29,32,33]; however, Kannangara et al. [29] found a single model of DT resulted in poor predictive performance, possibly due to overfitting of big and complex models [34]. To address this limitation, the present study considered DT-based ensemble algorithms and adopted two different ensemble techniques: bagging and boosting. Ensemble learning has been shown to outperform individual base models in various studies [35,36,37], due to its capacity to reduce the risk of selecting a poor classifier through individual classifier votes [38]. In a bagging approach, multiple bootstraps are created from a given training dataset, and an independent weak learner can be generated for each bootstrap. Accordingly, bagging can improve the stability and accuracy of ML algorithms [39]. Alternatively, boosting is an iterative and dependent-based learner that creates a strong classifier from weak classifiers by weighting. In the following subsections, the applied ensemble methods (i.e., RF, Extreme tree, GBM, and XGBoost) are described in detail.

2.3.1. Random Forest

RF proposed by Breiman [39] is a representative bagging-based ensemble method that generates bootstrap sampling. RF builds numerous subsets (bootstrap sampling) from training data and trains the same algorithm multiple times. The final predictive result is considered the average of all submodel predictions. With increasing tree numbers, RF can avoid overfitting and is less affected by outliers. In addition, it has superior predictive power compared to other ML algorithms, even when classes are imbalanced [39].

2.3.2. Extremely Randomized Trees

ETs is a recent bagging-based algorithm, where, unlike RF, ET uses whole origin data as is to create weak classifiers without bootstrapping, allowing it to maintain lower bias compared to RF models [40,41]. Moreover, instead of choosing the most discriminative split in each node, ET picks the best among K randomly generated splits, as random selections are advantageous for reducing variance and simultaneously, shortening the computational time [40,41].

2.3.3. Gradient Boosting Machine

GBM is a boosting method, where its iterative approach generates weak learners sequentially [42]. The GBM model has similar characteristics to the bagging approach, in that it is composed of weak learners; whereas the primary difference between GBM and RF is that the former’s model bias can be reduced by iteratively correcting errors made in the former tree, and building new trees [42,43]. Alternatively, the RF model reduces variance by averaging weak learners. The GBM model has been shown superior to RF in terms of overfitting and computational costs [43]; however, as a sequential learning method, boosting has the disadvantage of slow processing speeds due to difficult parallel processing.

2.3.4. Extreme Gradient Boosting

XGboost is a boosting-based ensemble tree algorithm generating boosted trees, and operating in parallel so that it can more efficiently account for regression and classification compared to GBM [44,45,46,47,48]. XGboost is well known for ‘regularized boosting’ technology; whereas the implementation of standard gradient boosting has no such regularization step [47]. Accordingly, such characteristics of XGboost can improve GBM model accuracy [47], and prevent overfitting [43].

2.4. Feature Selection and Hyper-Parameter Tuning

Several hyper-parameters must be carefully considered in the DT-based ensemble model [49,50], including the number of estimators (n_estimators) required to obtain the optimal performance (which is dependent upon the dataset’s properties; [51]), and the number of features (n_features) when finding the best split [50]. Therefore, in the present study, n_estimators and n_features were adjusted before applying each ensemble (i.e., RF, ET, GBM, and XGboost) algorithm. Further, to select the optimal number of estimators for the submodels, bagging and boosting ensemble models (50 each) with 100, 150, 200, … 500 component submodels were established. Further, the submodels for n_features included some or all of the six dataset variables. Each submodel contained 3 (3 features), 4, 5, and 6 variables. Recursive feature selection (RFE) was used for selecting submodel variables with 3, 4, and 5 features; thus, in this study, 36 predictive submodels were created by one ensemble algorithm with different numbers of estimators and features. For performance evaluation, the optimal n_estimators and n_features were selected based on R². Furthermore, this study tested various hyper-parameters to develop an optimal ensemble model (Table 4).

Table 4. Hyper-parameters applied in ensemble model development for DWGR prediction.

Algorithm	Hyper-Parameter	Definition	Tested Value
RF	n_estimators	The number of trees in the forest	100, 150, 200, 250, 300, 350, 400, 450, 500
	min_samples_split	The minimum number of samples required to split an internal node	1, 2, 3, 4, 5
	min_samples_leaf	The minimum number of samples required to be at a leaf node	1, 2, 3, 4, 5
	max_depth	The maximum depth of the tree	Maximum possible
ET	n_estimators	The minimum number of samples required to split an internal node	100, 150, 200, 250, 300, 350, 400, 450, 500
	min_samples_split	The minimum number of samples required to be at a leaf node	1, 2, 3, 4, 5
	min_samples_leaf		1, 2, 3, 4, 5
	max_depth	The maximum depth of the tree	None
	max_leaf_nodes		None
GBM	n_estimators	The number of boosting stages	100, 150, 200, 250, 300, 350, 400, 450, 500
	min_samples_split	The minimum number of samples required to split an internal node	1, 2, 3, 4, 5
	loss	Least squares	Least squares
	learning rate	Amount of learning	0.01, 0.1, 1
	subsample	Rate of sampling data to control overfitting	1.0
XGboost	n_estimators	The minimum number of samples required to split an internal node	100, 150, 200, 250, 300, 350, 400, 450, 500
	eta	0.3 step size shrinkage used in update to prevent overfitting	0.3
	max_depth	The maximum depth of the tree	10
	min_child_weight		1
	max_delta_step		0
	subsample		1

2.5. Model Validation and Evaluation

Leave one out cross-validation (LOOCV) was adopted as the model validation method. LOOCV is a special case of k-fold cross-validation, with the number of observations k = n. LOOCV is feasible when the sample size is small [52,53]; thus, it has been adopted in numerous studies to evaluate algorithm performance when the number of instances is small [7,54]. LOOCV uses all samples as testing and training data to ensure sufficient subset sizes and has the advantage of obtaining more stable results than that of the k-fold CV method for small datasets when compared to the validation set approach (e.g., 10-fold or k-fold; [55,56,57,58]). Accordingly, LOOCV was employed as a model validation method in consideration to the size of the dataset here (N = 690 samples).

Root mean square error (RMSE, Equation (5)), R² (Equation (6)), and the coefficient relationship (R, Equation (7)) were used to evaluate the performance accuracy of the ensemble predictive models, with high R² and R values, and lower RMSE values indicating improved performance:

$$\mathit{RMSE}=\sqrt{{\sum }_{i=1}^n\frac{{\left(y_i-x_i\right)}^2}{n}},$$

(5)

$$R^2 = 1-\frac{{\sum }_{=1}^n{\left(y_i-x_i\right)}^2}{{\sum }_{=1}^n{\left(y_i-{\underset{¯}{x}}_i\right)}^2},$$

(6)

$$R=\frac{{\sum }_{i=1}^n{\left(x_i-{\overline{x}}_i\right)}^2{\left(y_i-{\overline{y}}_i\right)}^2}{{\sum }_{i=1}^n\sqrt{{\left(x_i-{\overline{x}}_i\right)}^2}{\sum }_{i=1}^n\sqrt{{\left(y_i-{\overline{y}}_i\right)}^2}},$$

(7)

where $x_i$ is the observed value of the generated DW, $y_i$ is the predicted quantity of the generated DW, ${\overline{x}}_i$ is the average observed quantity of generated DW, ${\overline{y}}_i$ is the average predicted quantity of generated DW, and n is the number of samples.

To build a good model, we need to find a good balance between bias and variance such that it minimizes the total error and the best bias and variance balance was evaluated through prediction model error [59]. Accordingly, the prediction error of the model along with RMSE, R², and R as performance evaluation indicators were considered here to find the final ensemble model with the best predictive performance, as defined in Equation (8); [59,60]:

$$Error value\left[{\left(f\left(x\right)-\widehat{f}\left(x\right)\right)}^2\right]={\left(Bias\left[\widehat{f}\left(x\right)\right]\right)}^2+Var\left[\widehat{f}\left(x\right)\right]+Var\left(\epsilon \right)$$

(8)

where var($\epsilon$) is the irreducible error, which is the variance of the noise term in the true underlying function ($f\left(x\right)$), which cannot be reduced by any model [60].

3. Results

3.1. Optimal Number of Estimators and Features

The R² values corresponding to the number of features and estimators of RT, ET, GBM, and XGboost algorithms are shown in Figure 1a–d. The 6-feature model exhibited better predictive performance than the 5-, 4-, or 3-feature models, with the best predictive performance achieved when the number of estimators was 450. Further, the 5-, 4-, and 3-feature models exhibited the best predictive performance at 350, 350, and 300, respectively. For the ET model, the 6-feature model performed best when the number of estimators was 400, while the 5-, 4-, and 3-feature models had the best predictive performance at the number 350, 250, and 250 predictors, respectively. The GBM model achieved the best predictive performance for the 4-feature, 400 estimator model, while the 6-, 5- and 3-feature models showed the best results when the number of estimators was 400, 400, and 300, respectively. Lastly, the XGboost model achieved the best results with the 6-feature 400-estimator models, while the 5-, 4-, and 3-features models performed best when the number of estimators was 300, 300, and 150, respectively. According to the above results, the RT, ET, and XGboost models produced the best predictive performance using the 6-feature (i.e., all feature) model; however, the optimal number of estimators generated a different result for each model. Further, the GBM model achieved the best predictive performance in the 4-feature model, and unlike the RT, ET, and XGboost models, it exhibited similar predictive performances as the RF model, even with a small number of features.

Figure 1. (a) R² values according to number of estimators and features of RF model. (b) R² values according to number of estimators and features of ET model. (c) R² values according to number of estimators and features GBM model. (d) R² values according to number of estimators and features of XGboost model.

3.2. Feature Importance

Six features—location, structure, usage, GFA, wall material, and roof material—were utilized to estimate DWGR, and a difference in the contribution of features affecting the performance of RF, ET, GBM, and XGboost predictive models was recorded. Figure 2a–d shows the feature importance analysis results of the models exhibiting the best predictive performance using the optimal estimator number for each RF, ET, GBM, and XGboost algorithm (6, 450; 6, 400; 4, 300; and 6, 400, respectively). For RF, the most influential feature on DWGR was GFA (0.432), while the feature importance of location, roof material, wall material, structure, and usage was 0.205, 0.186, 0.070, 0.074, and 0.033, respectively. For ET, GFA importance (0.436) was highest, while that of location, roof material, wall material, structure, and usage tended to be similar to that of RF. For the GBM model, the 4 feature (GFA, location, roof, and wall material) had the greatest impact on DWGR, and the feature importance of GFA was the highest at 0.443. Alternatively, the XGboost predictive model exhibited a different trend of feature importance from RF, ET, and GBM (Figure 2d). In the XGBoost predictive model, the most important feature on DWGR was roof material (0.382), followed by wall material, structure, location, usage, and GFA (0.190, 0.173, 0.147, 0.069, and 0.040, respectively). This finding notably contrasts others, with regards to GFA’s impact on DWGR in the three predictive models (i.e., RF, ET, and GBM). Accordingly, it was revealed that if even a specific feature plays an important role in numerous ML algorithms, its effect is in other models. Therefore, the application of suitable ML algorithms for data characteristics, and the development of proper features for various ML algorithms, are essential for developing ML models with optimal predictive performance.

Figure 2. (a) Feature importance of RFmodel with optimal numbers of features and estimators. (b) Feature importance of ET model with optimal numbers of features and estimators. (c) Feature importance of GBM model with optimal numbers of features and estimators. (d) Feature importance of XGboost model with optimal numbers of features and estimators.

3.3. Performance Evaluation and Optimal Ensemble Model

Figure 3 displays the correlation results according to the feature combinations of the four ensemble models for DWGR, and Table 5 shows the performance indicator results of the model having the optimal number of estimators. The RF model achieved the best predictive performance at the 6-feature level (RMSE, 253.727; R², 0.6171; R, 0.7855), while the 3-feature predictive model produced the lowest prediction performance (RMSE, 261.836; R², 0.6006; R, 0.7750). Similarly, in the ET and XGboost models, the best and worst predictive performances were achieved in the 6- and 3-feature predictive models, respectively. In contrast, the 4-feature predictive model (RMSE, 253.085; R², 0.6142; R 0.7837) produced the best GBM model; whereas the 5-, 4-, and 3-feature predictive models produced similar results. Interestingly, the predictive performance of the 6-feature model was the lowest (RMSE, 265.834; R², 0.5806; R, 0.7620). Further, the accuracy of the RF 6-feature model was best among the 16 ensemble models in terms of R² and R values; however, the RMSE results showed that the GBM 4-feature predictive model with an RMSE value of 253.085 was slightly better than the RF 6-feature model (RMSE 253.727). In addition, the GBM 4-feature predictive model (R², 0.6142; R, 0.7837), presented prediction performances close to the RF 6-feature model. Similarly to the above, the RF and GBM models were determined to exhibit better predictive performances than either ET and XGboost algorithms based on the RMSE, R², and R performance evaluation results.

Figure 3. Correlation between observed and predicted values according to the number of features of ensemble models. (The yellow line means the ideal line where the predicted value and the observed value are the same. The blue dotted line represents the R² value).

Table 5. Comparison of model performance by RMSE, R², and R.

Model	n_Feature	n_Estimator	RMSE	R Square	R
RF	6	450	253.727	0.6171	0.7855
	5	350	261.772	0.6002	0.7747
	4	350	260.989	0.6026	0.7763
	3	300	261.836	0.6006	0.7750
ET	6	400	269.396	0.5936	0.7704
	5	350	277.334	0.5726	0.7567
	4	250	278.221	0.5709	0.7556
	3	250	277.690	0.5720	0.7563
GBM	6	400	265.834	0.5806	0.7620
	5	400	254.362	0.6103	0.7812
	4	300	253.085	0.6142	0.7837
	3	300	254.020	0.6114	0.7819
XGboost	6	400	281.262	0.5565	0.7460
	5	300	287.480	0.5539	0.7442
	4	300	287.790	0.5537	0.7441
	3	150	288.590	0.5516	0.7427

Model prediction error was also investigated to obtain the best variance–bias balance, together with the accuracy performance evaluation ensemble models (Figure 4). Since the prediction error of the GBM (4, 300) predictive model was the lowest (64,052), it was considered the best in terms of variance–bias tradeoff. The prediction errors of RF (6, 450), GBM (3, 300), and GBM (5, 400) models were 64,377, 64,526, and 64,700, respectively, notably similar to GBM (4, 300). Conversely, the prediction error was >72,574 and >79,122 for the ET and XGboost models. Accordingly, the GBM (4, 300), RF (6, 450), GBM (3, 300), and GBM (5, 400) predictive models were deemed the most appropriate model for achieving optimal variance–bias balance based on the prediction error.

Figure 4. Comparison of model prediction error values by variance–bias tradeoff.

Combining the performance evaluation results of the ensemble models above, the RF (6, 450) predictive model was deemed the best model in terms of R² and R values; whereas the GBM (4, 300) predictive model was best in terms of RMSE and prediction error. Accordingly, the RF (6, 450) and GBM (4, 300) predictive models were considered the most optimal for predicting DWGR (kg·m⁻²) based on the variable data of the five categorical and one numerical features. The observed and predicted values by the RF (6, 450) and GBM (4, 300) models are shown in Figure 5, where the mean observed value was 1171.2 kg·m⁻², and the means of the RF (6, 450) and GBM (4, 300) models were 1169.94 and 1166.25 kg·m⁻², respectively. The observed and predicted values by the GBM (3, 300), GBM (5, 400), ET (6, 400), and XGboost (6, 400) models are shown in Figure A1, Figure A2, Figure A3 and Figure A4.

Figure 5. Comparison of the observed and predicted DWGRs of RF (6 feature, 450 estimator) and GBM (4 feature, 300 estimator) models proposed as best.

4. Discussion

Building characteristics (e.g., GFA, usage, structure, location, etc.) were determined as the major key factors affecting DWGR [61]. Poon et al. [62] and Lu et al. [9] presented the correlation between GFA and DWGR, while Banias [63] that between usage and DWGR. Andersen et al. [64], Bergsdal et al. [65], and Bohne et al. [66] studied the effects of regional factors on DWGR; whereas the results of the present study indicated the highest feature importance of GFA and location for the developed RF, ET, and GBM models, in notable agreement with the existing research results (Figure 2a–d). Notably, the feature importance of wall and roof materials was higher than that of structure and usage in the present study; however, the wall and roof materials were not considered as major factors affecting DWGR in the previous studies. Further, the wall and roof materials exhibited higher feature importance than the GFA, usage, location, and structure in the XGboost model; accordingly, the feature importance of input variables in this model was quite different from that of the RT, ET, and GBM models, and there was a significant difference in the factors affecting DWGR from those presented in previous studies.

Section 2.3 presents a review of the existing research literature, and indicates that ET is an improved bagging method compared to RF in terms of bias and variance; whereas XGboost is a superior boosting method than GBM for accuracy improvement and overfitting prevention. The results of the predictive models employing the RF, ET, GBM, and XGboost ensemble algorithms in this study (Figure 3 and Figure 4; Table 5), however, were varied. The predictive performance of RF and GBM models in the present study was better than that of ET and XGboost models in terms of accuracy and prediction error. The existing research has employed DT-based ensemble algorithms [32], together with LightGBM, ET, RF, and Xgboost for time series datasets to estimate daily household waste generation. Specifically, Namoun et al. [32] found that LightGBM and ET algorithms exhibited better predictive performances than RF and XGboost. According to the results of a study applying a DT-based ensemble, except for C&D waste generation, Byeon [67] developed a predictive model to identify hypokinetic dysarthria by employing a DT classification and regression tree (CART), RF, GBM, and XGboost algorithms. The authors found that GBM (accuracy 83.1%) and RF (accuracy 83.8%) models achieved better predictive performances than the XGboost (accuracy 81.1%) and DT (accuracy 70.3%) models, where 16 input features (14 numerical, 2 categorical) were used. Ahmad et al. [68] used DT, ET, SVR, and RF to estimate useful solar thermal energy with 9 input features (8 numerical, 1 categorical), producing predictive performances (R²) of 0.957, 0.954, 0.930, and 0.903 for RF, ET, DT, and SVR, respectively, indicating the strength of the RF and ET models. Papadopoulos et al. [41] applied the GBM, RT, and ET algorithms using 8 numerical input features to estimate energy (cooling and heating) load, and revealed that the GBM model exhibited the best predictive performance in terms of mean square error (MSE) scores. In the case of heating load prediction, ET and RF produced similar results; whereas for the cooling load, ET was superior to RF.

Accordingly, previous studies using DT-based ensembles demonstrated various results regardless of the type of algorithms due to the difference in the characteristics of data used in each study (e.g., input feature types or sample size). Further, this may also be related to the considerable variation of the models’ predictive performances depending on the selection and adjustment of hyper-parameters before the algorithm applications in each study. Thus, the proper selection of algorithms and hyper-parameters in consideration of data characteristics are important issues during the development of ML models, and the process of deriving the optimal ML model in consideration of these issues becomes a key factor associated with predictive performance results. Considering these facts, this study developed an optimal ML model for predicting DWGR. To this end, a new set of input variables was developed, including the input variables used for DWGR prediction in previous studies [64,65,66]. In addition, this study applied DT-based algorithms considering the characteristics of the dataset and developed submodels applying various hyper-parameters. Based on this, a prediction model was developed with hyper-parameters for the final optimal performance model for DWGR prediction.

5. Conclusions

In this study, DT-based ensemble algorithms (i.e., RT, ET, GBM, and XGboost) were applied in consideration of data characteristics (relatively small-sized dataset consisting of categorical variables) to estimate DWGR (kg·m⁻²) in the end-of-life stage of a building. To develop a model, submodels were created according to the input features (GFA, location, usage, structure, wall materials, and roof materials), and the number of estimators. Subsequently, the optimal DT-based ensemble models were derived using performance indicators, such as R², R, RMSE, and prediction error. The findings of this study are summarized as follows.

(1)

RF and GBM exhibited superior predictive performances compared to ET and XGboost for the relatively small, categorical data environment.

(2)

The most suitable models were RF (6 features, 450 trees) and GBM (4 features, 300 trees), where the predictive performance of the former was: R², 0.6171; R, 0.7855; RMSE, 253.727; prediction error, 64,377; and latter was R², 0.6142; R, 0.7837; RMSE, 253.085; and prediction error, 64,052. The mean observed value was 1171.2 kg·m⁻², while the means of the RF (6, 450) and GBM (4, 300) models were 1169.94 and 1166.25 kg·m⁻², respectively. The GBM model presented excellent performance, even with 3- and 5-features, or 300 and 400 estimators. The mean of the GBM (3, 300) and GBM (5, 400) predictive models was 1167.14 and 1165.22 kg·m⁻², respectively.

(3)

A different result in feature importance was observed in the RT, ET, GBM, and XGboost models. In particular, the feature importance of GFA (0.432) had the greatest impact on RT, ET, and GBM models, followed by location, roof materials, and wall materials. For XGboost, the highest feature importance of 0.382 was determined to be the roof material, followed by wall material, structure, location, and usage; whereas the lowest feature importance was GFA.

The features used in this study correspond to the building exterior and characteristics that can be easily obtained from the building register provided by administrative agencies. This can facilitate decommissioning engineers or company officials for the rapid collection of applicable modeling features and is beneficial for benchmarking the DWGR predictive models here. In particular, the proposed GBM model could be employed to estimate DWGR, even with 4-features. Therefore, it was concluded that the proposed methods here can be easily used as a decision-making tool in demolition waste management plans for decommissioning engineers and companies.

As the model developed in this study was derived using a relatively small dataset, the limitation in ML modeling results due to dataset size was unavoidable. As the data used in this study referred to the field-collected values acquired through a direct survey before building demolition, future limitations in terms of time and manpower will be encountered. Thus, additional research should be conducted to derive a model that can improve accuracy through data collection methods such as surveys in future development. In addition, the performance of the model developed in this study can be seen as another limitation of this study that it is difficult to say that the R² value has an excellent predictive performance of about 0.62. It is difficult to see this reason as a problem of the type of ML algorithm selected or data preprocessing technology, and it is judged that there is a high possibility that distortion of the collected data has occurred by the investigators who participated in the data collection. In this respect, it seems necessary to properly control the uncertainty problem caused by the data collection method, and it seems that various DWGR data collection methods should be applied to secure reliable data.