Handling missing data for construction waste management: machine learning based on aggregated waste
generation behaviors
Zhongze Yang 1, Fan Xue 2* and Weisheng Lu 3
This is the peer-reviewed post-print version of the paper:
Yang, Z., Xue, F., & Lu, W. (2021). Handling missing data for construction waste
management: machine learning based on aggregated waste generation behaviors.
Resources, Conservation & Recycling, 175, 105809. Doi:
10.1016/j.resconrec.2021.105809
The final version of this paper is available at https://doi.org/10.1016/j.resconrec.2021.105809.
The use of this file must follow the Creative Commons Attribution Non-Commercial No
Derivatives License, as required by Elsevier’s policy
Highlights
• A set of 821 waste generation behavioral features defined and analyzed for
construction projects
• ‘Missing not at random’ project data handled by machine learning of aggregated
behavioral data
• Adaboost selected based on experiments on 2,451 construction projects
• Satisfactory results in 10-fold cross-validation (F1 = 0.87) and real-world tests (F1 =
0.80)
• Effective and inexpensive approach to portraying project behaviors in waste big data

5

Abstract
In the era of big data, data is increasingly driving the construction waste management (CWM)
for minimizing the impacts on the environment and recycling construction materials. However,
missing data, led by various information barriers, often undermines the decision-making and
hinders effective CWM. This paper applies aggregated behavior-based machine learning (ML)
1

Zhongze YANG, Research Assistant, M.Sc.
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong
Kong SAR, China
Email: zhongze@hku.hk
2

Fan XUE, Assistant Professor, Ph.D. Homepage: https://frankxue.com
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong
Kong SAR, China
Email: xuef@hku.hk, ORCiD : https://orcid.org/0000-0003-2217-3693
*: Corresponding author, Tel: +852 3917 4174, Fax: +852 2559 9457, Email: xuef@hku.hk
3

Weisheng LU, Professor, Ph.D.
Department of Real Estate and Construction, The University of Hong Kong, Pokfulam, Hong
Kong SAR, China
Email: wilsonlu@hku.hk, ORCiD : https://orcid.org/0000-0003-4674-0357

10

15

methods to handling the project-level ‘missing not at random’ (MNAR) data by using
aggregated waste generation behaviors as a case study. First, we define a set of 821 waste
generation behavioral features based on waste big data, then screen the indicative and decisive
behaviors using automatic feature selection. Then, the most predictive ML method, trained via
data of 2,451 construction projects in 2011-2016 in Hong Kong, is selected for handling the
MNAR data. The experiments showed that the prediction of project missing data was
satisfactory (validation F1 = 0.87, test F1 = 0.80). The contribution of this paper is to pinpoint
the potential of waste big data in portraying project behaviors for more value-added
applications, at the same time, to present a handling method for MNAR data that is automatic,
fast, and low-cost from the CWM practitioner’s perspective.
Keywords
Construction waste management, waste generation behavior, missing data handling, big data
analytics, machine learning

20

25

30

35

40

1 Introduction
Construction wastes have great impacts on the environment through different aspects like
ecosystem, environmental sustainability, and natural resources, while recycling of construction
materials from wastes can loosen the pressure on constrained natural materials, such as
running-out river sands (UNEP 2019; Osmani et al. 2008; Tang et al. 2020; Ma et al. 2020).
To make construction waste management (CWM) more effective, the industry has adopted
various informatization technologies including Geographic Information System (GIS), sensor
technology, Internet of Things (IoT), computer vision (You & Wu 2019; You et al. 2020; Chen
et al. 2021). Thanks to the rapid development of hardware and software, such as Building
Information Modelling (BIM), embedded devices, and various sensors, the construction
industry has entered the big data era for digitalizing construction waste handling processes
(Kerzner 2017; Eastman et al. 2011; Bilal et al. 2016; Xue et al. 2021). The unprecedented
amount of big data from the explosive growth of business corporates, government, and
scientific databases also enabled the construction project management to fully embrace the big
data analytics (Soibelman & Kim 2002).

The quality and completeness of data are crucial for data analytics (Sattari et al. 2017); however,
the data of construction activities is often incomplete. Missing data, such as ‘Null’, ‘N.A.’, and
sometimes misleading values led by information barriers, is an omnipresent challenge to the
monitoring and evaluation in the CWM (Bilal et al. 2016; Callistus & Clinton 2016). The
related information barriers in the construction industry arose from the distinctive and
accelerating complexities in modern construction projects (Luo et al. 2017). First, a modern
construction project often has multiple stakeholders, such as developer, contractor, designer,
and suppliers, and subsequently complex intra-organizational structures and relationships,

45

50

55

60

65

70

75

80

inconsistent interest, and poor communications among the stakeholders (Olander 2007; Atkin
& Skitmore 2008). Furthermore, construction work has applied diversified technologies,
numerous trades, and complex processes, which are often overlapping or parallel,
uncoordinated and highly variable (Kagioglou et al. 2000). In addition to complexity and
flexibility, many construction processes are with uncertainties, such as the change of materials,
mechanism, environment, and policies (Paslawski 2017). Other factors such as frequent
turnovers of site personnel and temporary multi-organization also lead to missing data of
construction projects (Love et al. 2002). From statistics and information perspectives, missing
data is the missingness of information and hinders people from understanding or finding a
phenomenon hiding in massive data (McKnight et al. 2007).

Mechanisms of missing data, which describe relationships between measured variables and the
probability of missing data, were studied to describe, understand, and handle missing data
based on specific analysis and assumptions (Baraldi & Enders 2010). Rubin (1976) classified
such mechanisms into three types according to the reason why data is missing: missing
completely at random (MCAR), missing at random (MAR), and missing not at random
(MNAR). MCAR means the missingness is completely independent of all observed and
unobserved data; for example, a worker forgets to return the questionnaire by chance. If the
missingness is independent of all unobserved data but may be dependent on the observed data,
it is said to be MAR; for example, high-paid workers may be less willing to share their income
information than those with lower salaries. As for MNAR, the missingness comes from nonrandom reasons, such as non-preferred values or ambiguous data collection instructions.
Therefore, handling MNAR data requires additional conjectures and explorations for hidden
reasons (Jia 2019); and MNAR is known more challenging than MCAR and MAR. However
in the construction industry, numerous missing data falls in the challenging MNAR type, due
to the information barriers in construction projects.

Researchers have applied several schools of statistical methods to handle general MNAR data
(Baraldi & Enders 2010), such as deletion, regression imputation, multiple imputation, and
maximum likelihood estimation (Peugh & Enders 2004; N.Baraldi & K.Enders 2010; Peeters
et al. 2015). Statistically based approaches try to make reasonable and justifiable assumptions
for model establishing, however, for MNAR data such assumptions are untestable and
unverifiable, which is the main limitation for estimation (Rabe et al. 2018). After entering the
big data era, advanced big data analytics and machine learning (ML) methods have thrown
light upon many research domains (Ma et al. 2020), including MNAR data handling. Unlike
the conventional statistics, ML methods’ predictions are often more accurate due to the
assumed nonlinear relationships of feature data and the ability of capturing high-order
interactions between features (Jerez et al. 2010; Sattari et al. 2017; Nugroho et al. 2020).

However, most studies only focused on the numerical complement and prediction, but did not
take the information barriers and the characteristics of target industries into consideration. Few
studies have paid attention to the problem of missing data handling, especially MNAR data, in
the construction industry.
85

90

95

100

105

110

115

There exist an increasing amount of well-structured feature data in CWM, which is essential to
statistical analysis and ML. A structured feature referring to the knowledge of a partial
subsystem could allow us easily guess the state of other parts of the same system, or a more
detailed state of parts in the same classification category (Li et al. 2002). For example, the
waste big data, and the aggregated waste generation behaviors of projects, waste facilities,
companies, and individual trucks can offer such features. Some governments have extensively
collected waste big data for environmental protection and cost control (Poon et al. 2004). For
example, the Government of Hong Kong implemented the Construction Waste Disposal
Charging Scheme (CWDCS) in 2006 to strengthen CWM. Based on millions of waste disposal
records in the CWDCS dataset, Lu (2019) identified illegal dumping behaviors and found
previously unknown characteristics of illegal dumpers. Xu et al. (2020a) also found that
different types of construction projects had different waste generation patterns using passive
bigger data in the same dataset. Through deep exploration of well-structured waste big dataset,
MNAR data related to the information barriers could be predicted based on the characteristic
waste generation behaviors with specific ML patterns.

This paper applies a ML approach based on big data-driven waste generation behavioral
features to handle MNAR data for construction projects. Unlike the handling methods in
quantity studies, this paper focuses on extracting and distilling the hidden behavioral features
through excessive definitions and selection to facilitate existing ML methods. Besides, trained
ML models provide new patterns and insights to understand the projects in the construction
industry. The proposed approach was tested on the real data from CWDCS and preliminarily
validated with a few developers and clients.

The remainder of the paper is organized as follows. After this introductory section, Section 2
is the literature review focusing on the missing data handling and waste generation behaviors.
Section 3 presents the detailed research methods, including data collection and processing,
feature definition and selection, training and evaluation of the ML methods, as well as the
predictive handling of MNAR data. Section 4 reports the experimental results, sensitivity
analyses, and findings from tests, followed by an in-depth discussion in Section 5. Conclusions
are drawn in Section 6.

120

125

130

135

140

145

150

155

2 Literature review
2.1 Waste generation behaviors
Behavior is “anything an organism does in response to a particular situation or stimulus;” and
classical theories state that “all behavior is due to a complex interaction between genetic
influence and environmental experience” (Pierce & Cheney 2017). In the domain of waste
management analysis, the generation laws or characteristics of construction waste can also be
seen as “behaviors”, since the waste generation and disposal schedules or flows are the
response to the particular waste management regulation and construction process. Such
behaviors, involving social, economic, environmental and human behavioral factors, can be
heterogeneous in different regions and situations regarding to waste treatment (Bakshan et al.
2017; Luangcharoenrat et al. 2019; Alcay et al. 2020). For example, governments are
developing and implementing new waste management systems, regulations, and techniques,
which profound influence the waste generation and recycling (Tam & Tam 2006; Guerrero et
al. 2013). Waste charging schemes, economic incentives of waste recycling, and personal
awareness of environmental protection significantly affect waste generation and disposal
(Corsini et al. 2018; Tamayo-Orbegozo et al. 2017). Many existing studies has aimed to
identify the waste generation behavior to predict and explore the influence factors using
dynamics and data-driven modeling techniques (Kontokosta et al. 2018). In general, current
research trends on waste generation behaviors include: (i) how individuals’ attitudes and
decisions affect the waste generation behaviors (Lingard et al. 2000; Begum et al. 2009; Mattar
et al. 2018); (ii) significant factors that influence the waste generation behaviors (Keser et al.
2012; Zhang et al. 2015); (iii) what models and techniques identify and predict the patterns of
waste generation behaviors (Karadimas & Loumos 2008; Rimaitytė et al. 2012; Johnson et al.
2017). Among the studies in (iii), supervised ML methods, such as SVM, K-means clustering,
and decision tree, have widely been adopted, sometimes with feature selection (Márquez et al.
2008; Kontokosta et al. 2018; Meza et al. 2019; Abbasi & El Hanandeh 2016).

Waste generation behaviors about solid construction wastes are profound and consistent
project-level features forming different behavior patterns (Tonglet et al. 2004). Watanabe
(1985) defined general patterns as an entity represented by a set of properties and feature
variables. And pattern recognition is usually posed as a classification problem using supervised
or unsupervised approaches (Jain et al. 2000). In the literature, abundant behavioral features
can be defined on big datasets, to represent many waste generation patterns. Examples are
trucks’ behavioral features in Lu (Lu 2019), accumulative waste generation flows in Xu et al.
(2020b), and combinations of prefabricated components in Lu et al. (2021). However, not every
feature is essential, informative, or helpful for missing data handling. Feature selection aims to
find the minimum subset automatically with some criteria to improve ML performance (Koller
& Saharni 1996; Lu 2019). Thus, feature selection can improve the efficiency (lower time cost)
and effectiveness (higher correctness) of an ML method at the same time by removing some
less informative features (Vafaie & Jong 1992). Compared with principal component analysis

160

165

170

175

180

(PCA), which is widely used in studies and linearly transforms all the features into integrated
principal components, feature selection pertains to the original meaning of features with no
changes (Vafaie & Jong 1992). To sum up, waste generation behaviors of construction projects,
together with dimensionality reduction, feature extraction, and feature selection, can improve
ML methods for handling missing data (Tang et al. 2014).

2.2 Missing data handling
Conventional missing data techniques, such as deletion and single imputation approaches,
usually perform poorly for MNAR. Although maximum likelihood estimation and multiple
imputation tend to perform better than most traditional techniques, these approaches still have
obvious drawbacks in handling MNAR data (Meeyai 2016). The approaches to handling
missing data in the literature can be divided into four different groups: complete case analysis,
imputation, maximum likelihood, and embedded ML (García-Laencina et al. 2009). Complete
case analysis, as the name suggests, is a procedure based on the complete data, which ignores
the missing values and requires the support of a big volume of data. Both imputation and the
maximum likelihood approaches aim at estimating; while imputation methods estimate the
missing values directly and maximum likelihood methods estimate the model parameters to
handle missing data (Little & Rubin 2019; Schafer 1997). In the MNAR case, the observed
data are no longer representative of the population, which leads to selection bias in the sample,
and therefore to bias in the parameters estimation. For imputation and maximum likelihood
methods, which are to model missing data distribution, will have a computational burden and
are often restricted to a limited number of MNAR variables (Sportisse et al. 2020). The
evaluation of the methods is in Table 1 (Pigott 2010; Shylaja & Kumar 2018; N.Baraldi &
K.Enders 2010; Nugroho et al. 2020).
Table 1. Assessment of statistical missing data approaches
Methods
Deletion

Complete case analysis
(List wise deletion)
Available case analysis
(Pair wise deletion)
Single imputation

Imputation
Multiple imputation

Maximum likelihood estimation

Advantages
1) Simple to implement
2) Acceptable with small
number of missing data
3) Suitable for MCAR data

Disadvantages
1) Need big volume of data
2) Not efficient, may cause
problems in prediction

1) Full sample size is
preserved
2) Administrative data is not
needed
3) Estimated variations that
can be justified by a statistic.
4) More suitable for MAR and
MCAR data

1) Single imputation may
lead bias to the estimation
2) Limited by the MAR
assumption
3) Not practical with large
databases
4) Sensitive to missing rate

1) Could produce accurate
estimates
2) Use all the available
information

1) Limited by the MAR
assumption
2) Model dependent

185

190

195

The embedded ML methods are ML approaches, such as decision trees and fuzzy neural
networks (Ishibuchi et al. 1993), which have been proven to be much more effective in handling
MNAR (Twala 2009; Song et al. 2008). Mitchell (2006) defined ML as ‘machine learns with
respect to a particular task T, performance metric P, and type of experience E, if the system
reliably improves its performance P at task T, following experience E’. In the four different
types of ML, i.e., unsupervised, supervised, semi-supervised, and reinforcement, supervised
ML algorithms learn or approximate a behavior of a function which maps a vector into one of
several classes by looking at several input outputs examples of the function (Osisanwo et al.
2017). Therefore, supervised ML algorithms can produce a general hypothesis of missing data
from externally supplied features, which applies to missing data handling (Singh et al. 2016;
Ge et al. 2017). Ding & Simonoff (2010) studied three types of missing data mechanisms and
proved that supervised ML classification methods could provide accurate predictions and are
more robust than traditional approaches, especially for the case when data in the test group is
also uncompleted. Currently, there exist many supervised ML algorithms, such as Support
Vector Machines (SVM), quadratic classifiers, k-means clustering, k-nearest neighbors (k-NN),
boosting, decision tree, random forest (RF), Artificial Neural Networks, Bayesian networks,
and ensemble methods combining them thereof.

200

205

210

215

Supervised ML methods have been integrated with non-ML methods to complete missing
values and collaborate to improve prediction accuracy. For example, Twala & Cartwright
(2010) ensembled decision trees with two imputation methods for predicting missing data, of
which the results showed that the ensemble method improved the accuracy. Nanni et al. (2012)
proposed a novel ensemble approach for completing missing data and tested its performance
on different databases. Rahman & Islam (2013) presented two novel techniques based on
Decision Tree impute missing values and used nine publicly available datasets to indicate the
superiority of these techniques. Tran et al. (2017) combined multiple imputations and ensemble
ML methods to handle missing data, and found it significantly better than decision tree and kNN in terms of classification accuracy.

It is generally accepted that different characteristics of missing data can produce different
effects on the completion, and some researchers have suggested that these characteristics
should be taken into consideration when selecting handling methods (Garciarena & Santana
2017). Thus, the performances of different ML methods have also been compared for handling
missing data based on different datasets in different domains. For example, Gavankar &
Sawarkar (2015) reviewed different algorithms of decision trees for handling missing data, and
employed a database to benchmark the performance of the algorithms. Abidin et al. (2018)
compared three methods, i.e., k-NN, decision tree, and Bayesian networks, for handling

220

225

230

missing data on a medical dataset. Perkowski (2020) focused on the bagging and boosting
ensemble ML methods on handling missing data and proved the significances. Therefore, for
analyzing well-structured construction waste data, the characteristics of missing data should
also be considered together with ML methods. In addition, features should be demanded based
on well-structured data for handling missing data in the field of construction project
management.

3 Research methods
Figure 1 shows the presented ML method for handling missing waste data. In general, there
are four steps. After the data preparation, the second step is the aggregation of waste
generation behaviors, which is the core characteristic that distinguishes this paper in the
literature. The third step is iterated ML training with feature selection and cross-validation,
while the last step is the tests with ground truth data obtained from interviews.

Figure 1. Flowchart of the presented ML method for handling missing waste data
235

240

3.1 Data source and preparation
The data source in this paper is the CWDCS dataset, which is a construction waste big data
containing over 12 million transaction records of Hong Kong since 2011 (EPD 2020). Besides,
the information of projects, facilities, vehicles, and their indexed relationships form a
completed data structure regarding the construction waste transactions. The whole dataset is
composed of structured data, such as numerical data and descriptive text of waste generation
and disposal. In light of excluding the uncompleted projects, the study period is set to between
January 2011 and December 2016 (both inclusive) including 19,626 projects. Figure 2 shows
the structure of the data tables in the CWDCS dataset, where the four data tables are:

•

Waste disposal data records every daily waste disposal information, including facility,
account, date of transaction, vehicle, net weight, waste depth, etc.

•

Facility data contains all government waste disposal facilities for construction waste,
including facility type, name, and code. The abbreviation of the facility corresponds to
that in the waste disposal database.

•

Vehicle data contains information on vehicle plate numbers and the maximum weight
capacities of 11,374 waste transportation trucks. The vehicle number corresponds to
that in the waste disposal database.

•

Project data contains detailed billing account information, such as contract number,
contract name, site address, department, type of construction work, etc. The account
number corresponds to that in the waste disposal database.

245

250

255

Figure 2. Excerpts and cross-references of waste disposal data, facility data, vehicle data, and
project data in the CWDCS dataset

260

265

The missing data column to handle is the ‘type of construction work’ in the project data.
According to the local regulations, the main contractor, who undertakes construction work
under contracts with values of HK$1 million or above, must register a billing account in
CWDCS solely for the contract. During the construction activities, construction waste will be
transported and dumped into the facilities by truck. The ‘type of construction work’ is required
to report the contract details, such as ‘new building’ , ‘renovation’ (general renovation, i.e.,
renovation, maintenance, and improvement (RMI)), ‘demolition’, ‘site formation’, and
‘foundation work’, along with project sum, site address, and contract number, as shown in
Figure 2. The ‘type of construction work’ is essential for the government’s waste charging
policy reviews, strategic management of waste facilities, and regular statistical reports. For
example, in Hong Kong, Monthly Digest from Buildings Department, Quarterly Report of

270

275

Employment and Vacancies at Construction Sites, and Report on the Quarterly Survey of
Construction Output from the Census and Statistics Department all have specialized statistics
of the performances of all construction works.

Many projects have missing project data in the CWDCS dataset. Due to the information barriers
in the construction industry, the detailed information of the projects is usually vague and
uncertain for information fillers, such as truck drivers, to fill confidently. Thus, fuzzy options,
such as ‘others’, or even giving up filling, are a much preferable choice, see in Figure 3. It can
be seen in Figure 3 that the majority of project type values are missing, which hinders many
value-added analyses and applications for CWM.

280

Figure 3. Project types in the CWDCS dataset (2011-2016), where ‘NA’ and ‘others’ are
regarded as missing values

285

290

This paper focuses on the top two types, i.e., ‘new building’ and ‘renovation,’ of all registered
projects in 2011-2016. As a result, 2,451 projects were selected to produce the training data,
including 625 ‘new building’ projects and 1,826 ‘renovation’ projects. A data cleansing is
conducted via the quartiles method, which removes data outliers and is neutral to non-normal
distributions. After the data cleansing, there are 895,063 records of waste disposal transactions,
to be aggregated for the 2,451 projects. The 10-fold cross-validation is used for the training
ML model, instead of the conventional 80-20 training-test splitting, because of the less bias.
Moreover, an extra test data set consists of our interviewee’s projects originally reported as
‘others’ and ‘NA’ to further verify the selected ML method and explore the practical value,
which is the final target of our research. The results of the test data set can at the same time,
verify whether the model is overfitting or not.

295

300

3.2 Aggregation of waste generation behaviors
We define 821 behavioral features of waste generation for each construction project to
represent the corresponding waste generation pattern thoroughly. The definitions are partially
based on the pattern definition in Watanabe (1985) and partially by extending the 54 yearly
indicators in Lu (2019). As listed in Table 2, there are 21 types of behavioral features clustered
into three groups: (i) truck usage behavioral features, (ii) waste disposal behavioral features,
and (iii) facility usage behavioral features. The definitions of the behavioral features well
exploit the waste transactions, anonymous truck information, project information, and facility
information interlinked in the CWDCS dataset.

305

Table 2. Definition of 821 features in 21 types of project waste generation behaviors
Group
(Total no.
of features)

Type
id
1
2

Truck
usage
behaviors
(226)

3
4
5
6

Waste
disposal
behaviors
(55)

7
8

Feature type definition
The truck usage for the pattern
which the maximum load <=16t
The truck usage for the pattern
which the maximum load =24t
The truck usage for the pattern
which the maximum load =30t
The truck usage for the pattern
which the maximum load >=38t
The usage of all trucks for every
waste disposal
The usage count of distinct
trucks for every waste disposal
Activeness of waste disposal
record
Amount of waste generation

9

Facility
usage
behaviors
(540)

Statistics of waste generation
10~18 The usage of the nine facilities
19
The usage of all public facilities
20

The usage of all sorting facilities

21

The usage of all landfill facilities

Unit

Daily
(d)

Features*
Weekly Monthly Yearly
（y）
(w)
(m)

Total
(t)

1
1
1
1
1

𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑗𝑗𝑖𝑖 , 𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 , 𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 ,
𝑝𝑝𝑝𝑝𝑝𝑝5𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝25𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝50𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝75𝑖𝑖𝑗𝑗 ,
𝑝𝑝𝑝𝑝𝑝𝑝95𝑖𝑖𝑗𝑗 , 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑗𝑗𝑖𝑖 , 𝐼𝐼𝐼𝐼𝐼𝐼𝑗𝑗𝑖𝑖 ,
𝑖𝑖 = 1, … ,5, 𝑗𝑗 = 𝑑𝑑, 𝑤𝑤, 𝑚𝑚, 𝑦𝑦

/#

1
1
t
t
1
1
1

/#

𝑎𝑎𝑎𝑎𝑎𝑎𝑗𝑗7 ,
𝑗𝑗 = 𝑑𝑑, 𝑤𝑤, 𝑚𝑚, 𝑦𝑦

/#

/#

𝑠𝑠𝑡𝑡𝑖𝑖 =

𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑖𝑖
𝑠𝑠𝑡𝑡5

𝑖𝑖 = 1, … ,4
𝑠𝑠𝑡𝑡5
𝑠𝑠𝑡𝑡6

/#

𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑡𝑡𝑛𝑛 ,
𝑛𝑛 = 11,
… ,16

𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑗𝑗𝑖𝑖 , 𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 , 𝑚𝑚𝑚𝑚𝑚𝑚𝑗𝑗𝑖𝑖 ,
𝑝𝑝𝑝𝑝𝑝𝑝5𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝25𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝50𝑖𝑖𝑗𝑗 , 𝑝𝑝𝑝𝑝𝑝𝑝75𝑖𝑖𝑗𝑗 ,
𝑝𝑝𝑝𝑝𝑝𝑝95𝑖𝑖𝑗𝑗 , 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑗𝑗𝑖𝑖 , 𝐼𝐼𝐼𝐼𝐼𝐼𝑗𝑗𝑖𝑖 ,

𝑠𝑠𝑡𝑡8

/#

𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑖𝑖

𝑠𝑠𝑡𝑡𝑖𝑖 = ∑

𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑖𝑖

𝑖𝑖 = 10, … ,21

𝑖𝑖 = 9, … ,21, 𝑗𝑗 = 𝑑𝑑, 𝑤𝑤, 𝑚𝑚, 𝑦𝑦
1
*: Stddev: standard deviation; pct: percentile; extremum: max − min; IQR: interquartile range.
#: Feature omitted. Reasons include insufficient variations from 𝑠𝑠𝑡𝑡6 (Type 6), no definitions (Types 7 and 9), and
represented by yearly features (Type 8).

310

For the first group of truck usage behaviors in Table 2, the 226 features are divided into six
types by the maximum load: <=16t, =24t, =30t, >=38t, all trucks, and distinct trucks. For every
behavioral feature type i, there exists a 2D array of features 𝑠𝑠𝑗𝑗𝑖𝑖 , where s = mean, standard

315

deviation, maximum, minimum, extremum, quartiles—i.e., 5th, 25th, 50th (median), 75th,
95th—or interquartile range is a statistic, and j = daily, weekly, monthly, and yearly is the
period of behaviors. Besides, a total or percentage number is defined for each feature type.

320

The second group of 55 waste disposal behaviors represents a project’s transaction activeness,
yearly summation amount of waste, and the statistics with granularities j = daily, weekly,
monthly, and yearly. In the third group of facility usage behaviors, the 540 features are grouped
into 12 types. The first nine types represent the nine waste facilities operated in Hong Kong,
while the remaining three types are the groups of the nine facilities, as listed in Table 3. Not
that the ‘outlying island transfer’ facilities are omitted due to the low usage, which reflects
construction solid wastes were rarely generated on the outlying islands.
Table 3. List of the waste handling facilities and associated feature type Ids
Facility group
Feature
Facility name with site
(feature type Id) type Id
10
Chai Wan Public Fill Barging Point
11
Mui Wo Temporary Public Fill Reception Facility
Public fill
(19)
12
Fill Bank at Tseung Kwan O Area 137
13
Fill Bank at Tuen Mun Area 38
14
Sorting Facilities at Tseung Kwan O Area 137
Sorting
(20)
15
Sorting Facilities at Tuen Mun Area 38
16
North East New Territories Landfill
Landfill
17
South East New Territories Landfill
(21)
18
West New Territories Landfill

Required proportion
of inert materials
100%

≥ 50%
≤ 50%

325

In total, 821 features are defined for the 21 types to represent the behaviors of one of the 2,451
projects thoroughly, as listed in Table 2. Thus, after aggregation, the data set is a table of
aggregated features in 2,451 rows and 821 columns of waste generation behaviors.

330

335

3.3 ML training with cross-validation and feature selection
The ML method in this paper is then applied to process the behavioral features and behavioral
patterns represented by the features. Given a set X = [f1, f2, …, f821]T of waste generation
behavioral features, the task of missing project data handling H in this paper is thus to find a
prediction function:
H (X) ∈ C, for any X

(1)

where C is the set of possible values of the missing data, i.e., ‘new building’ and ‘renovation’
in this study. The whole dataset for the supervised ML methods includes training input data in
a 2,451 × 821 matrix and 2,451 × 1 class labels.

340

Four supervised ML methods are selected to represent the four popular groups of ML according
to our pilot tests of the handling process H in Eq. (1). The ‘decision tree’ is selected from the
group of trees and rules, ‘k-NN’ from the group of instance-based ML, ‘SVM’ from the group
of function ML, and the ‘ensemble ML’ from the group of meta-models. Other ML methods,

345

such as Artificial Neural Networks and Bayesian Networks, are also tested but dropped, as
shown in Tables A4 and A5 in the Appendix.
i.

350

ii.

iii.
355

iv.
360

365

370

Decision tree learning is a data mining technique to classify data based on complete
information. The structure of a decision tree contains three parts: a) root node, contains
the complete set of samples; b) internal node, contains corresponding characteristic
attribute tests; c) leaf node, represents the decision result (Myles et al. 2004).
K-Nearest Neighbors (k-NN) is a basic ML algorithm that can compare features in a
test set with those in a training set and find the top k most similar record, so that the
classification of those inputted data is that with the most occurrences in k data
(Laaksonen & Oja 1996).
Support vector machines (SVM) is a binary classification model. Its basic model is a
linear classifier with the most considerable interval defined in the feature space, and the
main idea is to solve the separating hyperplane that can correctly divide the training
dataset and have the largest geometric interval (Hearst et al. 1998).
Ensemble ML combines multiple weakly-supervised models to obtain a better and
more comprehensive strong-supervised model. The underlying idea of ensemble ML is
that even if an individual weak classifier gets a wrong prediction, other weak classifiers
can also correct it. Example ensemble ML algorithms are AdaBoost, RUSBoost,
LogitBoost, GentleBoost, and Bag. (Dietterich 2000).

The F1 scores from 10-fold cross-validation to compare the ML methods. The evaluation of a
trained ML model’s performance is primarily based on the F1 score, with reference to precision
and recall rates, where
Precision = True Positive / (True Positive + False Positive)

(2)

Recall = True Positive / (True Positive + False Negative)

(3)

F1= 2 × Precision × Recall / (Precision + Recall).

(4)

In Eqs. (2) and (3), ‘true positive’ counts the missing data that is estimated as ‘positive’ while
the prediction is ‘true’ according to the actual data (same as the predicted); similar rules apply
to ‘false positive’ and ‘false negative.’ The F1 score in Eq. (4) integrates the precision and
recall rates into one harmonic mean, evaluates the two classes at the same time, and can solve
the problem of the unbalanced sample sizes of the training classes.

375

Feature selection is applied to trim the noisy behavioral features prior to the ML training, as
suggested in the literature (Koller & Saharni 1996; Lu, Big data analytics to identify illegal
construction waste dumping: A Hong Kong study 2019). The feature selection algorithm is the
permutation importance in the open-source library scikit-learn (ver. 0.24) in Python (ver. 3.9.1)

380

385

390

395

400

(Pedregosa et al. 2011). The parameters in the algorithm usually include estimator, X, y, scoring,
and n_repeats. The estimator is the corresponding ML classifier; the X is the training data; the
y was the prediction (handed MNAR data in this paper); the scoring function of permutation
importance is the F1 metric, and n_repeats indicates how many iterations to loop for. The
permutation importance indicates the contribution of features to achieve a higher F1 score and
to choose useful features. The permutation importance is calculated as follows: i) first, a
baseline metric, defined by scoring, is evaluated on X dataset; ii) next, a feature column from
the validation set is permuted and the metric is re-evaluated. After calculating the difference
between the baseline metric and metric from the feature column permutation, this function will
return the importance score of the original ranking. Features with a positive importance score
will be selected for ML.

Classification Learner in MATLAB (ver. R2019b) is also used to train the ML methods
comparing with the results in scikit-learn. In order to obtain the best ML model with the best
parameter settings, the Hyperparameter Optimization in MATLAB is applied to fine-tuning the
ML parameters automatically. Meanwhile, the hyperparameter optimization could efficiently
avoid the overfitting of model training, since it will cease automatically based on the
performance of the learning curve. The four ML methods are, therefore, fine-tuned and
compared in iterations using the F1 of 10-fold cross-validation on the training dataset. The ML
method with the top F1 metric is finally determined.
3.4 Test
This step applies the trained and optimized ML method to predict the missing data in other
construction projects. A set of construction project cases are selected based on three screening
conditions:
i)

405

ii)
iii)

410

Public construction projects with complete contract number and contract name for
obtaining trustworthy and ethical interview data; and
Sufficient (> 1,000) construction waste transaction records from 2011 to 2016 or
having data records for at least three years from 2011 to 2016
The ‘type of construction work’ in the dataset now is ‘Others’ or ‘NA’.

After the screening, 36 projects labeled as “Others” are filtered. All the data is organized by
the same approach in Section 3.2 but used the optimized ML method with the same parameters
directly. Many details and background information about the 36 projects are available freely
on the internet. Furthermore, we can select typical cases for interviewing the developer and
client, the Hong Kong Housing Authority (HKHA), on the ground truth and reasons of
reporting as missing.

415

420

425

4 Experiments
4.1 ML training and validation results
The experiments were conducted on a laptop computer, with an Intel i5-5200U 2.20 GHz CPU
and 8GB memory, where the hyperparameter optimizer was ‘Bayesian’ and the maximum
number of iteration was 50. Table 4 compares the evaluation results of 10-fold cross-validation
of the four ML methods. In general, all the four ML methods can produce > 0.8 F1 scores with
selected feature sets sized from 74 to 378. For the decision tree method, 124 features were
selected and F1 = 0.82. For the k-NN method, 222 features were selected, and F1 = 0.81. There
were 378 features selected for SVM, and F1 = 0.82. The AdaBoost method received 74
features—the least in the methods, while F1 = 0.87 was considerably higher than the rest three
methods. Therefore, the AdaBoost method was applied for handling missing data in later steps.
Table 4. Average performance metrics of the ML models in 10-fold cross-validation, where
the best value in each row is in bold.
No. of selected features
New building
Renovation
Precision
Overall

430

435

440

Decision tree
124

SVM

k-NN

AdaBoost

222

378

0.76
0.90
0.83

0.76
0.89
0.83

0.76
0.90
0.83

Recall

New building
Renovation
Overall

0.69
0.93
0.81

0.67
0.93
0.80

0.68
0.93
0.81

Validation
F1 score

New building
Renovation
Overall

0.72
0.91
0.82

0.71
0.91
0.81

0.72
0.92
0.82

74
0.89
0.91
0.90
0.72
0.97
0.85
0.80
0.94
0.87

The decision tree concluded from the training data is shown in Figure 4, and the pruning level
is 13 out of 21. Although the decision tree failed to return the best predictions of missing data,
its interpretable presentation can reveal general patterns and behavioral characteristic of the
two types of construction work. The root node in Figure 4 has two sub-trees. The upper subtree with the condition ‘the mean of yearly waste generation ≥ 185.635t’ represents the projects
with high-level yearly waste generation. There were 453 projects labeled as ‘New building’,
with higher weekly waste disposal records, less usage of trucks whose maximum load is below
16t, and smaller extremum. In other words, construction projects with continuous high-level
waste generation amount and lower-level fluctuations are likely ‘New buildings.’ In the lower
sub-tree in Figure 4, 1741 projects are labeled as ‘Renovation’, where the ‘yearly 95% quantile
of the usage of all trucks for every waste disposal <201.5’ branch has 1723 ‘Renovation’
projects. It means that with lower-level yearly waste generation, construction projects having
less truck usage are likely ‘Renovation.’

Figure 4. The decision tree learned by Classification Learner, where the number indicates the
sample size of each leaf
445

450

AdaBoost method returned the best results on top of 74 features, as listed in Table 5 (other ML
methods’ features are in the Appendix). Among three groups of features, 6.2% (14 out of 226)
truck usage features were selected, together with 38.2% (21 out of 55) waste disposal and 7.2%
(39 out of 540) facility usage features. The importance scores of these 74 features were shown
in Figure 5, and the statistical features of waste generation (Type 9) in the group of waste
disposal behaviors was considerably the most critical than the rest in terms of predicting project
type. Two following-up feature types fell in the rest two groups, respectively: one was the
usage of Sorting Facilities at Tseung Kwan O Area 137 (Type 14) and the other was yearly
statistics of waste generation amount (Type 8).

455

Table 5. List of 74 decisive features selected for AdaBoost from the pool of 821 features
Group
(subtotal)

Truck usage
behaviors
(14)

Feature
type id
1
2
3
4
5
6
7

Daily (d)
stddev, IQR

pct95
stddev

Weekly (w)
pct95
mean

Feature
Monthly (m)
pct75
pct50, pct5

Yearly (y)

Total (t)

mean, pct50
pct75

stddev
act

𝑠𝑠𝑡𝑡6

Waste
disposal
behaviors
(21)

9
10
11
12
13
14

Facility
usage
behaviors
(39)

weightn,
n = 11 to 14, 16

8
mean, max,
pct25, pct75,
IQR
stddev
IQR
mean
mean, stddev

15
16

stddev

17

mean

min,
extremum,
pct5, pct95

460

465

mean, pct95
pct50

stddev
extremum,
pct5

pct95
pct5
pct50

pct25
IQR

𝑠𝑠𝑡𝑡10
𝑠𝑠𝑡𝑡13
𝑠𝑠𝑡𝑡14
𝑠𝑠𝑡𝑡15

max

𝑠𝑠𝑡𝑡16

pct25, pct75

18
19

mean, max,
min, pct25

pct95
stddev, pct75

20
21

stddev

Subtotal

19

pct75

pct5

pct95

stddev

mean
mean, pct5

pct75
stddev

14

17

16

𝑠𝑠𝑡𝑡19

𝑠𝑠𝑡𝑡20

8

Figure 5. The importance scores of 74 decisive features, where the blocks from bottom to top
represent the order from Daily (d) to Total (t).

As shown in Figure 5, the sub-feature with the highest importance score is the usage of trucks
for daily waste generation, followed by the sub-features in the usage of the Sorting Facilities
at Tseung Kwan O Area 137 (Type 14: yearly IQR), the amount of the waste disposal (Type 8:
yearly weight in 2012), the statistics of the waste generation (Type 9: monthly pct25), and the
usage of the Public Fill facilities (Type 19: weekly pct75). In other words, there were no distinct
time periods of waste generation behaviors. Yet, there were unbalanced pairs in the truck types

470

475

and facilities in Table 5 and Figure 5. For the group of truck usage, the selection of the first
two truck types stands for a disparity in using trucks with capacities ≤ 24t among four different
truck types. As for the facility uses, Mui Wo Temporary Public Fill Reception Facility (Type
11) and West New Territories Landfill (Type 18) had less informative features. In our later
investigation, it was found that the first facility received very low wastes from both new
building and renovation, while the latter one had little differences between the two construction
types. Sorting Facilities at Tseung Kwan O Area 137 (Type 14) and Public fill facilities (Type
19) have much more informative features. It was found that there were two disposal facilities
at Tseung Kwan O Area 137, one is a public fill facility while the other is a sorting facility. It
seems that the usage of these two facilities at Tseung Kwan O Area 137 contributes a lot to the
distinction of the two construction types. In summary, the selected features are informative to
correlate the targeted construction types.

480

485

4.2 Parameters sensitivity
Table 6 shows the F1 scores from 10-fold cross-validations of the four ML methods on two ML
libraries, i.e., scikit-learn and MATLAB. The ML libraries were found notably impactful to
the results, even for the same ML method. In general, MATLAB’s Classification Learner
produced consistently higher F1, which were preferred for handling the missing data in this
study. The reason was that Classification Learner exploits an iterated ‘Hyperparameter
Optimization’ improvement for the ML methods.
Table 6. Comparison of F1 scores of same ML methods implemented in different scientific
packages, the best value in each row in bold
ML library

Language

Decision tree

k-NN

SVM

AdaBoost

scikit-learn

Python

0.76

0.72

0.79

0.84

MATLAB (iter=50)

Object C

0.82

0.81

0.82

0.87

490

495

Table 7 lists the F1 scores of the AdaBoost method, with the number of iterations increased
from 1 to 70. The key parameter of the Hyperparameter Optimization in MATLAB is the
number of iterations. The scores hint at an inverse U-shape, i.e., highest in the middle. A series
of 30 to 50 iterations in total should train AdaBoost for the best predictions. Similarly, the
minor parameters of AdaBoost were determined as ‘maximum number of splits’ = 16, ‘number
of learners’ = 492, and ‘learning rate’ = 0.3145.
Table 7. Validation F1 scores of AdaBoost against different number of iterations, the best value
in each row in bold
No. of iterations
Validation
F1 score

New building:
Renovation:
Overall:

1

10

0.76
0.93
0.85

0.78
0.93
0.86

30

0.80
0.94
0.87

50

0.80
0.94
0.87

70
0.79

0.94
0.87

500

505

510

4.3 Results of tests
Application values are tested by further tests through new construction projects. First,
Adaboot’s selected 74 features were calculated for the test data of 36 public projects. Then, the
missing data can be predicted by the trained AdaBoost model. As a result, there were 26
projects classified as ‘new buildings’, and other 10 projects are predicted as ‘renovations’. The
public project documents and reports available on the Internet claimed that there were 24 ‘new
buildings’, 12 general renovations (or RMI), as listed in Table 8. Thus, the test results was
satisfactory, with precision = 0.82, recall = 0.79, F1 = 0.80. Therefore, it is plausible to
complete the missing data of construction project types using the presented ML method.
Table 8. Confusion matrix of the test results on the 36 public projects, where correct prediction
are in bold
Actual
Predicted

515

520

New building

General renovation (RMI)
Renovation
Maintenance and
improvement

New building

22

0

4

Renovation

2

1

7

The HKHA, as the developer and client of the test cases, was contacted for double-checking
the predictions. Two construction projects, one predicted as a new building and the other
predicted as renovation, were selected. We found one reason for the missing data was the
inconsistent definition of the ‘renovation’ type. In EPD’s CWDCS system, ‘renovation’ stands
for the general renovation, or RMI; however, in some other systems, like HKHA’s, the general
renovation has subcategories, such as ‘maintenance and improvement’, a narrowly-defined
‘renovation’, ‘commercial building’, and ‘civil engineering’. The inconsistent definitions are
also one prevailing information barrier in the industry.
Table 9. Two project cases in the tests
Proj.
ID
1

2

Project Description
“The works comprise the building
work on total construction area
132,047 m2 with retail area, estate
management office, carpark,
refuse collection facilities, etc.”
“The works comprise the
maintenance and repairs of and
alterations and additions to any
properties, site and slopes

Project
value
(HK$
million)
2,821.8

374.0

Type of construction work
Period
2014 –
2016

Our
prediction
New
building

Ground truth
from client
New building

Correct
?
Yes

2014 –
2017

(General)
renovation

Maintenance
and
improvement

Yes

including vacant flat
refurbishment.”

525

530

535

5 Discussion
For practitioners in the construction industry, the methodology presented in this paper can
summarize quantitative waste generation behaviors, predict missing data for each project, and
help eliminate the information barriers. Although construction projects are known to have
distinctive characteristics, their aggregated waster generation behaviors are relatively stable for
analyses and applications. For example, the findings in this paper show ‘New building’ and
‘renovation’ projects had different behavioral patterns in Hong Kong (2011-2016). ‘New
building’ projects had higher yearly waste generation amount, less usage of trucks with lower
maximum load, and more usage of all the trucks; while the ‘Renovation’ showed opposite
trends. From the perspective of CWM, the waste generation behaviors utilize the existing big
data and enable new data analytics like missing data handling. Participators in CWM can
explore the essential and subtle behavioral characteristics of the waste generation according to
the feature selection and ML results. The completion of the construction work type may also
benefits further construction waste management and studies, e.g., construction work typespecific waste facility policies. In the experiments, the waste generation behaviors showed
correlations with the vehicles and facilities, and further concluded in the interpretable decision
trees and non-readable ML methods.

540

545

550

555

The proposed aggregated waster generation behavior-based ML method has several advantages
in handling missing project data. First, it is much more efficient (less time and human resource)
than interviews to complete the missing information. The predictions were satisfactory (F1 =
0.87) and provided a quantitative basis for further studies. Meanwhile, the decisive attributes
can explain the core differences in the aggregated behaviors regarding the missing data. From
the perspective of construction management, the quantitative behavioral features can help
understand the projects and characteristics of different construction work types. The excessive
feature generation is inclusive for the features that are easily overlooked. Based on the
investigation on the predicted results, possible data management policies can be recommended,
e.g., renaming the ‘renovation’ to ‘general renovation’ or ‘RMI’ in EPD’s system. Last but not
least, the proposed methods are general, so that they can be applied to other types of project
data elsewhere, not limited to ‘new building’ and ‘renovation’ in Hong Kong.

However, there exist a few limitations in this study. First, only two significant values of missing
data were selected to illustrate the presented methods. More extensive ranges of missing data
can be tested in future studies. Secondly, one ML method, decision tree, can be interpreted in
a human-readable fashion. More interpretable ML methods should be investigated in the future.

560

565

570

575

580

585

590

There was a slight unbalance between the missing data, e.g., ‘Renovation’ got higher precision,
recall, and F1 scores in Table 3. Thus, the balance, apart from correctness, should be noted in
the predictions as well. Lastly, the waste big data in this paper was from Hong Kong during
2011 and 2016, the socio-industrial environments and project behaviors may change in another
time in another place.
6 Conclusion
Information barriers in the construction industry lead to missing data that hinders CWM from
digitalization and data-driven decision-making. This paper presents a machine learning (ML)
method to handle quantitative missing not at random (MNAR) data based on aggregated
projects’ waste generation behaviors. Experiments on waste big data, including 895,063 rows
of disposal transactions data in 2,451 construction projects, confirmed the presented method.
The ensemble ML was selected as the best prediction method with F1 = 0.87. The critical
features were recognized and analyzed, besides the characteristics of the waste generation of
two types of construction projects were summarized. As shown in Table 5, the No. 9 statistical
feature “Statistics of waste generation” (by time) in the group “waste disposal behavior” was
the most critical characteristic; the second-tier characteristics included one “truck usage
behavior” (No. 2), one “waste disposal behavior” (No. 8) and two “facility usage behaviors”
(Nos. 14 and 19). The major parameters in the ML method were analyzed, and the results of
tests were preliminarily validated by the true values from clients.

The contribution of this paper is two-fold. From the theoretical perspective, it pinpoints that
waste big data has the potential to articulate projects’ waste generation behaviors for more
value-added applications. From the CWM practitioner’s perspective, the presented handling
method is an automatic, fast, and low-cost pipeline to complete MNAR data and understand
the projects’ behaviors. Meanwhile, the more suitable and consistent classification category is
beneficial to decrease the confusion due to the information barriers and improve the completion
and efficiency of the big data system in CWM. The recognized and summarized waste
generation behaviors of the two construction types, like the truck usage and the usage
preference of some specific facilities, could also help CWM practitioners and policymakers to
have a deeper understanding of the waste generation characteristics, which further benefits the
rulemaking.

Future work to extend this study lies in several aspects. First, missing data having three or more
values should be studied to validate the methodological scalability. Furthermore, interpretable
ML should be studied besides correctness, such as the F1 scores. The balance of predictions
across different target values is another issue to investigate. Meanwhile, researchers should be
reminded to pay attention to ethical and privacy issues while processing more information.

595

600

Acknowledgments
The work presented in this paper was supported by the Policy Innovation and Co-ordination
Office (PICO) of the Government of Hong Kong SAR under the Strategic Public Policy
Research (SPPR) (No.: S2018.A8.010.18S). The funding agency had no role in research design,
data collection and analysis, decision to publish, or preparation, of the manuscript.

605

Data Availability
The training datasets, including the full set consisting of 821 features and those selected for the
four ML methods, are provided as supplemental materials of this paper. Other data, models, or
codes that support the findings of this study are available from the corresponding author upon
reasonable request.

610

615

620

625

630

References
Abbasi, M. & El Hanandeh, A. (2016). Forecasting municipal solid waste generation using
artificial intelligence modelling approaches. Waste management, 56, 13-22.
doi:10.1016/j.wasman.2016.05.018
Abidin, N. Z., Ismail, A. R. & Emran, N. A. (2018). Performance analysis of machine
learning algorithms for missing value imputation. International Journal of Advanced
Computer Science and Applications (IJACSA), 9(6), 442-447.
doi:10.14569/IJACSA.2018.090660
Alcay, A., Montañés, A. & Simón-Fernández, M. B. (2020). Waste generation in Spain. Do
Spanish regions exhibit a similar behavior? Waste Management, 112, 66-73.
doi:10.1016/j.wasman.2020.05.029
Atkin, B. & Skitmore, M. (2008). Stakeholder management in construction. Construction
Management and Economics, 26(6), 549-552. doi:10.1080/01446190802142405
Bakshan, A., Srour, I., Chehab, G., El-Fadel, M. & Karaziwan, J. (2017). Behavioral
determinants towards enhancing construction waste management: A Bayesian
Network analysis. Resources, Conservation and Recycling, 117, 274-284.
doi:10.1016/j.resconrec.2016.10.006
Baraldi, A. N. & Enders, C. K. (2010). An introduction to modern missing data analyses.
Journal of school psychology, 48(1), 5-37. doi:10.1016/j.jsp.2009.10.001
Begum, R. A., Siwar, C., Pereira, J. J. & Jaafar, A. H. (2009). Attitude and behavioral factors
in waste management in the construction industry of Malaysia. Resources,
Conservation and Recycling, 53(6), 321-328. doi:10.1016/j.resconrec.2009.01.005
Bilal, M., Oyedele, L. O., Qadir, J., Munir, K., Ajayi, S. O., Akinade, O. O., ... & Pasha, M.
(2016). Big Data in the construction industry: A review of present status,
opportunities, and future trends. Advanced engineering informatics, 30(3), 500-521.
doi:10.1016/j.aei.2016.07.001

635

640

645

650

655

660

665

Callistus, T. & Clinton, A. (2016). Evaluating Barriers to Effective Implementation of Project
Monitoring and Evaluation in the Ghanaian Construction Industry. Procedia
engineering, 164, 389-394. doi:10.1016/j.proeng.2016.11.635
Chen, J., Lu, W. & Xue, F. (2021). “Looking beneath the surface”: A visual-physical feature
hybrid approach for unattended gauging of construction waste composition. Journal
of Environmental Management, 286, 112233. doi:10.1016/j.jenvman.2021.112233
Corsini, F., Gusmerotti, N. M., Testa, F. & Iraldo, F. (2018). Exploring waste prevention
behaviour through empirical research. Waste Management, 79, 132-141.
doi:10.1016/j.wasman.2018.07.037
Dietterich, T. (2000). Ensemble Methods in Machine Learning. International workshop on
multiple classifier systems (pp. 1-15). Berlin: Springer. doi:10.1007/3-540-45014-9_1
Ding, Y. & Simonoff, J. S. (2010). An investigation of missing data methods for
classification trees applied to binary response data. Journal of Machine Learning
Research, 11(1), 131-170. doi:abs/10.5555/1756006.1756012
Eastman, C. M., Eastman, C., Teicholz, P., Sacks, R. & Liston, K. (2011). BIM handbook: A
guide to building information modeling for owners, managers, designers, engineers
and contractors (2nd ed.). John Wiley & Sons.
EPD. (2020). Monitoring of Solid Waste in Hong Kong: Waste Statistics for 2019. Hong
Kong: Environmental Protection Department, Government of Hong Kong SAR.
Retrieved from
https://www.wastereduction.gov.hk/en/assistancewizard/waste_red_sat.htm
García-Laencina, P. J., Sancho-Gómez, J. L., Figueiras-Vidal, A. R. & Verleysen, M. (2009).
K nearest neighbours with mutual information for simultaneous classification and
missing data imputation. Neurocomputing, 72(7-9), 1483-1493.
doi:10.1016/j.neucom.2008.11.026
Garciarena, U. & Santana, R. (2017). An extensive analysis of the interaction between
missing data types, imputation methods, and supervised classifiers. Expert Systems
with Applications, 89, 52-65. doi:10.1016/j.eswa.2017.07.026
Gavankar, S. & Sawarkar, S. (2015). Decision Tree: Review of Techniques for Missing
Values at Training, Testing and Compatibility. 2015 3rd International Conference on
Artificial Intelligence, Modelling and Simulation (AIMS) (pp. 122-126). Kota
Kinabalu: IEEE. doi:10.1109/AIMS.2015.29
Ge, Z., Song, Z., Ding, S. X. & Huang, B. (2017). Data mining and analytics in the process
industry: The role of machine learning. Ieee Access, 5, 20590-20616.
doi:10.1109/ACCESS.2017.2756872.
Guerrero, L. A., Maas, G. & Hogland, W. (2013). Solid waste management challenges for
cities in developing countries. Waste management, 33(1), 220-232.
doi:10.1016/j.wasman.2012.09.008

670

675

680

685

690

695

700

705

Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J. & Scholkopf, B. (1998). Support vector
machines. IEEE Intelligent Systems and their applications (pp. 18-28). IEEE.
doi:10.1109/5254.708428
Ishibuchi, H., Miyazaki, A., Kwon, K. & Tanaka, H. (1993). Learning from incomplete
training data with missing values and medical application. Proceedings of 1993
International Conference on Neural Networks(IJCNN-93-Nagoya, Japan), (pp. 18711874). Nagoya, Japan. doi:10.1109/IJCNN.1993.717020
Jain, A. K., Duin, R. P. & Mao, J. (2000). Statistical pattern recognition: A review. IEEE
Transactions on pattern analysis and machine intelligence, 22(1), 4-37.
doi:10.1109/34.824819
Jerez, J. M., Molina, I., García-Laencina, P. J., Alba, E., Ribelles, N., Martín, M. & Franco,
L. (2010). Missing data imputation using statistical and machine learning methods in
a real breast cancer problem. Artificial intelligence in medicine, 50(2), 105-115.
doi:10.1016/j.artmed.2010.05.002
Jia, Z. (2019). A Nonparametric Multiple Imputation Approach For MNAR Mechanism Using
the Sample Selection Model Framework. Arizona: The University of Arizona.
Retrieved from http:arizona.edu/handle/10150/632554
Johnson, N. E., Ianiuk, O., Cazap, D., Liu, L., Starobin, D., Dobler, G. & Ghandehari, M.
(2017). Patterns of waste generation: A gradient boosting model for short-term waste
prediction in New York City. Waste management, 62, 3-11.
doi:10.1016/j.wasman.2017.01.037
Kagioglou, M., Cooper, R., Aouad, G. & Sexton, M. (2000). Rethinking construction: the
generic design and construction process protocol. Engineering, construction and
architectural management, 7(2), 141-153. doi:10.1108/eb021139
Karadimas, N. V. & Loumos, V. G. (2008). GIS-based modelling for the estimation of
municipal solid waste generation and collection. Waste Management & Research,
26(4), 337-346. doi:10.1177/0734242X07081484
Kerzner, H. (2017). Project management: a systems approach to planning, scheduling, and
controlling (12th ed.). John Wiley & Sons.
Keser, S., Duzgun, S. & Aksoy, A. (2012). Application of spatial and non-spatial data
analysis in determination of the factors that impact municipal solid waste generation
rates in Turkey. Waste management, 32(3), 359-371.
doi:10.1016/j.wasman.2011.10.017
Koller, D. & Saharni, M. (1996). Toward Optimal Feature Selection. Stanford InfoLab.
Retrieved from http://ilpubs.stanford.edu:8090/208/
Kontokosta, C. E., Hong, B., Johnson, N. E. & Starobin, D. (2018). Using machine learning
and small area estimation to predict building-level municipal solid waste generation in
cities. Computers, Environment and Urban Systems, 70, 151-162.
doi:10.1016/j.compenvurbsys.2018.03.004

710

715

720

725

730

735

740

745

Laaksonen, J. & Oja, E. (1996). Classification with learning k-nearest neighbors.
Proceedings of International Conference on Neural Networks (ICNN'96) (pp. 14801483). Washington, DC: IEEE. doi:10.1109/ICNN.1996.549118
Li, R. P., Mukaidono, M. & Turksen, I. B. (2002). A fuzzy neural network for pattern
classification and feature selection. Fuzzy Sets and Systems, 130(1), 101-108.
doi:10.1016/S0165-0114(02)00050-7
Lingard, H., Graham, P. & Smithers, G. (2000). Employee perceptions of the solid waste
management system operating in a large Australian contracting organization:
implications for company policy implementation. Construction Management &
Economics, 18(4), 383-393. doi:10.1080/01446190050024806
Little, R. J. & Rubin, D. B. (2019). Statistical Analysis with Missing Data (3rd ed.). New
York: John Wiley & Sons. doi:10.1002/9781119482260
Love, P. E., Holt, G. D., Shen, L. Y., Li, H. & Irani, Z. (2002). Using systems dynamics to
better understand change and rework in construction project management systems.
International journal of project management, 20(6), 425-436. doi:10.1016/S02637863(01)00039-4
Lu, W. (2019). Big data analytics to identify illegal construction waste dumping: A Hong
Kong study. Resources, Conservation and Recycling, 141, 264-272.
doi:10.1016/j.resconrec.2018.10.039
Lu, W., Lee, W. M., Xue, F. & Xu, J. (2021). Revisiting the effects of prefabrication on
construction waste minimization: A quantitative study using bigger data. Resources,
Conservation and Recycling, 170, 105579. doi:10.1016/j.resconrec.2021.105579
Luangcharoenrat, C., Intrachooto, S., Peansupap, V. & Sutthinarakorn, W. (2019). Factors
influencing construction waste generation in building construction: Thailand’s
perspective. Sustainability, 11(13), 3638. doi:10.3390/su11133638
Luo, L., He, Q., Jaselskis, E. J. & Xie, J. (2017). Construction project complexity: research
trends and implications. Journal of construction engineering and management,
143(7), 04017019. doi:10.1061/(ASCE)CO.1943-7862.0001306
Ma, J., Cheng, J. C., Jiang, F., Chen, W., Wang, M. & Zhai, C. (2020). A bi-directional
missing data imputation scheme based on LSTM and transfer learning for building
energy data. Energy and Buildings, 216, 109941. doi:10.1016/j.enbuild.2020.109941
Ma, M., Tam, V. W., Le, K. N. & Li, W. (2020). Challenges in current construction and
demolition waste recycling: A China study. Waste Management, 118, 610-625.
doi:10.1016/j.wasman.2020.09.030
Márquez, M. Y., Ojeda, S. & Hidalgo, H. (2008). Identification of behavior patterns in
household solid waste generation in Mexicali's city: Study case. Resources,
Conservation and Recycling, 52(11), 1299-1306. doi:10.1016/j.resconrec.2008.07.011
Mattar, L., Abiad, M. G., Chalak, A., Diab, M. & Hassan, H. (2018). Attitudes and behaviors
shaping household food waste generation: Lessons from Lebanon. Journal of Cleaner
Production, 198, 1219-1223. doi:10.1016/j.jclepro.2018.07.085

750

755

760

765

770

775

780

785

McKnight, P. E., McKnight, K. M., Sidani, S. & Figueredo, A. J. (2007). Missing data: A
gentle introduction. New York: Guilford Press. Retrieved from
https://psycnet.apa/record/2007-06639-000
Meeyai, S. (2016). Logistic Regression with Missing Data: A Comparison of Handling
Methods, and Effects of Percent Missing Values. Journal of Traffic and Logistics
Engineering, 4(2). doi:10.18178/JTLE.4.2.128-134
Meza, J. K., Yepes, D. O., Rodrigo-Ilarri, J. & Cassiraga, E. (2019). Predictive analysis of
urban waste generation for the city of Bogotá, Colombia, through the implementation
of decision trees-based machine learning, support vector machines and artificial
neural networks. Heliyon, 5(11), e02810. doi:10.1016/j.heliyon.2019.e02810
Mitchell, T. M. (2006). The discipline of machine learning (Vol. 9). Pittsburgh: Carnegie
Mellon University. Retrieved from
https://web.cs.wpi.edu/~kmlee/cs539/MachineLearning
Myles, A. J., Feudale, R. N., Liu, Y., Woody, N. A. & Brown, S. D. (2004). An introduction
to decision tree modeling. Journal of Chemometrics: A Journal of the Chemometrics
Society, 18(6), 275-285. doi:10.1002/cem.873
N.Baraldi, A. & K.Enders, C. (2010). An introduction to modern missing data analyses.
Journal of School Psychology, 48(1), 5-37. doi:10.1016/j.jsp.2009.10.001
Nanni, L., Lumini, A. & Brahnam, S. (2012). A classifier ensemble approach for the missing
feature problem. Artificial intelligence in medicine, 55(1), 37-50.
doi:10.1016/j.artmed.2011.11.006
Nugroho, H., Utama, N. P. & Surendro, K. (2020). Comparison Method for Handling
Missing Data in Clinical Studies. Proceedings of the 2020 9th International
Conference on Software and Computer Applications (pp. 46–50). New York: ICSCA
2020. doi:10.1145/3384544.3384594
Olander, S. (2007). Stakeholder impact analysis in construction project management.
Construction management and economics, 25(3), 277-287.
doi:10.1080/01446190600879125
Osisanwo, F. Y., Awodele, O., Hinmikaiye, J. O., Olakanmi, O. & Akinjobi, J. (2017).
Supervised machine learning algorithms: classification and comparison. International
Journal of Computer Trends and Technology (IJCTT), 48(3), 128-138.
doi:10.14445/22312803/IJCTT-V48P126
Osmani, M., Glass, J. & Price, A. D. (2008). Architects’ perspectives on construction waste
reduction by design. Waste Management, 28(7), 1147-1158.
doi:10.1016/j.wasman.2007.05.011
Paslawski, J. (2017). Flexible approach for construction process management under risk and
uncertainty. Procedia engineering, 208, 114-124. doi:10.1016/j.proeng.2017.11.028
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,
Brucher, M., Perrot, M. & Duchesna. (2011). Scikit-learn: Machine Learning in

790

795

800

805

810

815

820

825

Python. Journal of Machine Learning Research, 12, 2825--2830.
doi:10.5555/1953048.2078195
Peeters, M., Zondervan-Zwijnenburg, M., Vink, G. & Van de Schoot, R. (2015). How to
handle missing data: A comparison of different approaches. European Journal of
Developmental Psychology, 12(4), 377-394. doi:10.1080/17405629.2015.1049526
Perkowski, E. (2020). Impact of ensemble machine learning methods on handling missing
data. Twente: University of Twente. Retrieved from
http://purl.utwente.nl/essays/82210
Peugh, J. L. & Enders, C. K. (2004). Missing Data in Educational Research: A Review of
Reporting Practices and Suggestions for Improvement. Review of Educational
Research, 74(4), 525-556. doi:10.3102/00346543074004525
Pierce, W. D. & Cheney, C. D. (2017). Behavior Analysis and Learning, Sixth Edition (6th
ed.). New York: Routledge. doi:10.4324/9781315200682
Pigott, T. D. (2010). A Review of Methods for Missing Data. Educational Research and
Evaluation , 7(4), 353-383. doi:10.1076/edre.7.4.353.8937
Poon, C. S., Yu, A. T., Wong, S. W. & Cheung, E. (2004). Management of construction
waste in public housing projects in Hong Kong. Construction Management &
Economics, 22(7), 675-689. doi:10.1080/0144619042000213292
Rabe, B. A., Day, S., Fiero, M. H. & Bell, M. L. (2018). Missing data handling in noninferiority and equivalence trials: A systematic review. Pharmaceutical statistics,
17(5), 477-488. doi:10.1002/pst.1867
Rahman, M. G. & Islam, M. Z. (2013). Missing value imputation using decision trees and
decision forests by splitting and merging records: Two novel techniques. KnowledgeBased Systems, 53, 51-65. doi:10.1016/j.knosys.2013.08.023
Rimaitytė, I., Ruzgas, T., Denafas, G., Račys, V. & Martuzevicius, D. (2012). Application
and evaluation of forecasting methods for municipal solid waste generation in an
eastern-European city. Waste Management & Research, 30(1), 89-98.
doi:10.1177/0734242X10396754
Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581-592.
doi:10.1093/biomet/63.3.581
Sattari, M. T., Rezazadeh-Joudi, A. & Kusiak, A. (2017). Assessment of different methods
for estimation of missing data in precipitation studies. Hydrology Research, 48(4),
1032-1044. doi:10.2166/nh.2016.364
Schafer, J. L. (1997). Analysis of incomplete multivariate data. New York: Chapman and
Hall/CRC. doi:10.1201/9780367803025
Shylaja, B. & Kumar, R. S. (2018). Traditional versus modern missing data handling
techniques: An overview. International Journal of Pure and Applied Mathematics,
118(14), 77-84. Retrieved from https://acadpubl.eu/jsi/2018-118-1415/articles/14/12.pdf

830

835

840

845

850

855

860

865

Singh, A., Thakur, N. & Sharma, A. (2016). A review of supervised machine learning
algorithms. 2016 3rd International Conference on Computing for Sustainable Global
Development (INDIACom) (pp. 1310-1315). New Delhi: IEEE. Retrieved from
https://ieeexplore.ieee.org/abstract/document/7724478
Soibelman, L. & Kim, H. (2002). Data preparation process for construction knowledge
generation through knowledge discovery in databases. Journal of Computing in Civil
Engineering, 16(1), 39-48. doi:10.1061/(ASCE)0887-3801(2002)16:1(39)
Song, Q., Shepperd, M., Chen, X. & Liu, J. (2008). Can k-NN imputation improve the
performance of C4. 5 with small software project data sets? A comparative
evaluation. Journal of Systems and software, 81(12), 2361-2370.
doi:10.1016/j.jss.2008.05.008
Sportisse, A., Boyer, C. & Josses, J. (2020). Estimation and imputation in probabilistic
principal component analysis with missing not at random data. 34th Conference on
Neural Information Processing Systems. Canada: NeurIPS 2020. Retrieved from
https://proceedings.neurips.cc/paper/2020/file/4ecb679fd35dcfd0f0894c399590be1aPaper.pdf
Tam, V. W. & Tam, C. M. (2006). Evaluations of existing waste recycling methods: a Hong
Kong study. Building and Environment, 41(12), 1640-1660.
doi:10.1016/j.buildenv.2005.06.017
Tamayo-Orbegozo, U., Vicente-Molina, M. A. & Villarreal-Larrinaga, O. (2017). Ecoinnovation strategic model. A multiple-case study from a highly eco-innovative
European region. Journal of Cleaner Production, 142, 1347-1367.
doi:10.1016/j.jclepro.2016.11.174
Tang, J., Alelyani, S. & Liu, H. (2014). Feature selection for classification: A review. Data
Classification: Algorithms and Applications (pp. 37-64). CRC Press.
doi:10.1201/b17320
Tang, Z., Li, W., Tam, V. W. & Xue, C. (2020). Advanced progress in recycling municipal
and construction solid wastes for manufacturing sustainable construction materials.
Resources, Conservation & Recycling: X, 6, 100036. doi:10.1016/j.rcrx.2020.100036
Tonglet, M., Phillips, P. S. & Bates, M. P. (2004). Determining the drivers for householder
pro-environmental behaviour: waste minimisation compared to recycling. Resources,
conservation and recycling, 42(1), 27-48. doi:10.1016/j.resconrec.2004.02.001
Tran, C. T. (2017). Multiple imputation and ensemble learning for classification with
incomplete data. In Intelligent and Evolutionary Systems (pp. 401-415). Springer.
doi:10.1007/978-3-319-49049-6_29
Twala, B. & Cartwright, M. (2010). Ensemble Missing Data Techniques for Software Effort
Prediction. Intelligent Data Analysis, 14(3), 299-331. doi:10.3233/IDA-2010-0423
Twala, B. (2009). An empirical comparison of techniques for handling incomplete data using
decision trees. Applied Artificial Intelligence, 23(5), 373-405.
doi:10.1080/08839510902872223

870

875

880

885

890

895

UNEP. (2019). Sand and sustainability: Finding new solutions for environmental governance
of global sand resources. Geneva, Switzerland: United Nations Environment
Programme. Retrieved from https://wedocs.unep.org/handle/20.500.11822/28163
Vafaie, H. & Jong, K. D. (1992). Genetic algorithms as a tool for feature selection in machine
learning. Proceedings Fourth International Conference on Tools with Artificial
Intelligence TAI '92. 1, pp. 200-203. Arlington: IEEE. doi:10.1109/TAI.1992.246402
Watanabe, S. (1985). Pattern recognition: human and mechanical. New York: John Wiley &
Sons, Inc.
Xu, J., Lu, W., Ye, M., Webster, C. & Xue, F. (2020a). An anatomy of waste generation
flows in construction projects using passive bigger data. Waste Management, 106,
162-172. doi:10.1016/j.wasman.2020.03.024
Xu, J., Lu, W., Ye, M., Xue, F., Zhang, X. & Lee, B. F. (2020b). Is the private sector more
efficient? Big data analytics of construction waste management sectoral efficiency.
Resources, Conservation and Recycling, 155, 104674.
doi:10.1016/j.resconrec.2019.104674
Xue, F., Wu, L. & Lu, W. (2021). Semantic enrichment of building and city information
models: A ten-year review. Advanced Engineering Informatics, 47, 101245.
doi:10.1016/j.aei.2020.101245
You, Z. & Wu, C. (2019). A framework for data-driven informatization of the construction
company. Advanced Engineering Informatics, 39, 268-277.
doi:10.1016/j.aei.2019.02.002
You, Z., Wu, C., Zheng, L. & Feng, L. (2020). An Informatization Scheme for Construction
and Demolition Waste Supervision and Management in China. Sustainability, 12(4),
1672. doi:10.3390/su12041672
Zhang, G., Lin, T., Chen, S., Xiao, L., Wang, J. & Guo, Y. (2015). Spatial characteristics of
municipal solid waste generation and its influential spatial factors on a city scale: a
case study of Xiamen, China. Journal of Material Cycles and Waste Management,
17(2), 399-409. doi:10.1007/s10163-014-0257-7

Appendix
Table A1. List of 124 decisive features selected for Decision Tree from the pool of 821 features
Group
(subtotal)

Feature
type id
1
2

Truck usage
behaviors
(28)

Daily (d)
mean
mean,
extremum

3
4
5
6

stddev, max

pct95

Feature
Monthly (m)
extremum
extremum,
pct5
mean

mean,
stddev, max

stddev, max,
min

Weekly (w)
mean
mean

Yearly (y)
pct50

Total (t)

mean,pct25,IQR

𝑠𝑠𝑡𝑡2

max, min, IQR
pct95

𝑠𝑠𝑡𝑡1

7
Waste
disposal
behaviors
(62)

act
weightn,
n = 11,14, 16

8
max,pct50,
pct5,pct25,
pct95,
extremum,
IQR, stddev

min,
pct5,pct25,p
ct75,
pct95,IQR

mean, min,
extremum,
pct25,pct75,
pct95,IQR

stddev

mean, IQR

max
pct95
pct75

IQR

extremum

14

pct25

mean,
stddev,
pct75

IQR

15

stddev

IQR

mean, max

extremum

16

stddev,pct95

17

pct95,IQR

stddev, max,
extremum
min,pct50,
pct5,pct95

mean, max,
pct75
min,pct5,
pct95,IQR
extremum

9
10
11
12
13

Facility
usage
behaviors
(34)

act

18
19

mean,pct95,
IQR

20

max

900

26

𝑠𝑠𝑡𝑡10

stddev

stddev,
pct25, pct75
pct75,IQR
max
mean,
stddev

pct50, extremum
stddev,pct95

21
Subtotal

mean, stddev,
pct50,extremum,
pct75,pct95,IQR

𝑠𝑠𝑡𝑡8

27

29

extremum

𝑠𝑠𝑡𝑡14

𝑠𝑠𝑡𝑡19

pct,pct75

𝑠𝑠𝑡𝑡20

34

8

𝑠𝑠𝑡𝑡21

Table A2. List of 222 decisive features selected for KNN from the pool of 821 features
Group
(subtotal)

Feature
type id

Daily (d)

Weekly (w)

Feature
Monthly (m)

1

2
Truck usage
behaviors
(42)

stddev

3

mean, stddev,
max, pct50,
pct75, pct95

stddev

Yearly (y)
stddev, min,
pct50, pct5,
pct25
mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95, IQR
mean, stddev,
max, min, pct50,
pct5, pct25,
pct75, pct95,
IQR

Total (t)

4
stddev, max,
extremum

5
6
7

act
weightn,
n = 11,…,16

8
Waste
disposal
behaviors
(44)

9

10

mean, stddev,
max, min,
pct50,
extremum,
pct5, pct25,
pct75, pct95,
IQR

stddev, max,
pct50, extremum,
pct95

stddev,
max, min,
pct50,
extremum,
pct5, pct25,
pct95, IQR

stddev, max,
min,
extremum,
pct5, pct25,
pct75, pct95,
IQR

mean, stddev,
max, extremum,
pct75, pct95,
IQR

max

max, pct95

mean, stddev,
max,
extremum,
pct75, pct95

mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95, IQR

stddev, max,
extremum,
pct95

stddev,
max,
extremum,
pct95

mean, stddev,
max,
extremum,
pct75, pct95,
IQR

11

Facility
usage
behaviors
(136)

12

13

stddev

14
15

16

stddev

stddev

stddev, max,
extremum,

mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95
stddev, max,
pct50, pct95
stddev
mean, stddev,
max, pct50,
extremum,
pct75, pct95,
IQR
mean, stddev,
max, pct50,
extremum,

𝑠𝑠𝑡𝑡8

pct75, pct95,
IQR
mean, max,
min, pct25,
pct95
stddev

pct75, pct95,
IQR

mean,
stddev,
max, pct50,
extremum,
pct75,
pct95, IQR

mean, stddev,
max, pct50,
extremum,
pct75, pct95,
IQR

mean, stddev,
max, min, pct50,
pct5, pct25,
pct75, pct95,
IQR

stddev

stddev

17
18

19

stddev, max,
extremum,
pct95, IQR

20

21
Subtotal

21

26

stddev
stddev

mean, max,
pct75, pct95,
IQR

mean, stddev,
max, min, pct50,
pct5, pct25,
pct75, pct95
mean, stddev,
max, pct50,
pct75, pct95

60

114

1

Table A3. List of 378 decisive features selected for SVM from the pool of 821 features
Group
(subtotal)

Feature
type id

Weekly (w)
stddev,
max, min,
extremum,
pct5, pct25,
pct95, IQR
stddev,
pct50,
pct25, pct95
mean,
stddev, min,
pct50, pct5,
pct25,
pct75,
pct95, IQR
min, pct50,
pct5, pct25

mean, stddev,
max, min,
pct50,
extremum,
pct5, pct25,
pct75, IQR
min, pct50,
pct5, pct25

stddev,
pct25

stddev, min,
pct5, pct25

act

act

stddev, min,
pct50, pct5,
pct25, pct75,
IQR

mean,
stddev, min,
pct50,
pct25,
pct75,
pct95, IQR

mean, stddev,
min, pct50,
pct5, pct25,
pct75, IQR

10

min, pct5,
pct25, IQR

mean,
stddev, min,
pct50, pct5,
pct25,
pct75, IQR

11

mean, stddev,
max, pct50,
extremum,
pct25, pct75,
pct95, IQR

mean,
stddev,
max, pct50,
extremum,
pct25,
pct75,
pct95, IQR

12

mean, stddev,
min, pct50,
extremum,
pct25, pct75,
pct95, IQR

mean,
stddev, min,
pct50, pct5,
pct25,
pct75,
pct95, IQR

13

mean, min,
pct50, pct5,
pct25

mean,
stddev, min,
pct50, pct5,

1

Truck usage
behaviors
(105)

Daily (d)
mean, stddev,
max,
extremum,
pct75, pct95,
IQR

2

stddev, min,
pct5, pct25

3

mean, stddev,
min, pct50,
pct5, pct25,
pct75, IQR

4

min, pct50,
pct5, pct25

5

stddev

6
7
Waste
disposal
behaviors
(23)

Facility
usage
behaviors
(250)

Feature
Monthly (m)
stddev, max,
pct50,
extremum,
pct75, pct95

Total (t)

mean, stddev

8

9

Yearly (y)
mean, stddev,
max, min, pct50,
extremum, pct5,
pct75, pct95,
IQR

mean, stddev,
max, min,
extremum,
pct5, pct25,
pct75, pct95,
IQR
mean, stddev,
max, min,
pct50,
extremum,
pct5, pct25,
pct75, pct95,
IQR
mean, stddev,
max, min,
pct50,
extremum,
pct5, pct25,
pct75, pct95,
IQR
mean, stddev,
min, pct50,

mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95, IQR
pct50
stddev, min,
pct50, pct5,
pct25

𝑠𝑠𝑡𝑡4

act
weightn,
n = 13,15,16
stddev, min,
pct5, pct25

mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95, IQR
mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
pct95, IQR
mean, stddev,
max, min, pct50,
extremum, pct5,
pct25, pct75,
IQR

𝑠𝑠𝑡𝑡12

14

15
16

pct5, pct25,
pct75

min, pct5,
pct25
stddev,
max, min,
extremum,
pct5

𝑠𝑠𝑡𝑡14

stddev, min,
pct5, pct25

mean, stddev,
min, pct5, pct25

pct50

min, pct5,
pct25

min, pct5

17

pct50, pct25

pct50

pct25

stddev

18

stddev, max,
min,
extremum,
pct5, IQR

stddev,
pct95

max,
extremum

stddev, min,
pct50, pct5,
pct25

19

mean, stddev,
min, pct50,
pct5, pct25,
pct75, IQR

mean, stddev,
min, pct50,
pct5, pct25,
pct75, IQR

stddev, min,
pct5, pct25,
pct75, IQR

min

stddev

21

min, pct50,
pct5, pct25
pct50, pct25

Subtotal

95

20

905

mean, min,
pct50, pct5,
pct25, IQR
stddev, max,
min,
extremum,
pct5, pct25
mean, pct50,
pct75

pct25,
pct75, IQR

mean,
stddev, min,
pct50, pct5,
pct25,
pct75,
pct95, IQR
min, pct5,
pct25

𝑠𝑠𝑡𝑡15
𝑠𝑠𝑡𝑡16
𝑠𝑠𝑡𝑡17

𝑠𝑠𝑡𝑡19

pct25, pct75
94

95

87

7

Table A4. List of 162 decisive features selected for Neural-network from the pool of 821
features
Group
(subtotal)

Feature
type id

Daily (d)

Weekly (w)

1

stddev

2
Truck usage
behaviors
(43)

3

stddev

stddev

mean,
stddev, max

extremum
stddev

max,
extremum

6
7
Waste
disposal
behaviors
(88)

stddev

4
5

Feature
Monthly (m)

stddev,
max,
extremum

stddev

act

8

9

mean, stddev,
pct5, pct25,
pct75, pct95

10

mean,
pct50,
extremum,
pct75,
pct95, IQR
stddev,
max,
extremum

mean, stddev,
max, pct50,
extremum,
pct75, pct95,
IQR
stddev, max,
extremum

11

Facility
usage
behaviors
(31)

Yearly (y)
stddev, max,
extremum, pct75,
pct95, IQR
max, min,
extremum, pct5,
pct25, pct75,
IQR
stddev, max,
pct95
stddev, max,
pct75, IQR
stddev, max,
min, extremum,
pct5, pct75,
pct95

12

max,
extremum

stddev,
max,
extremum

stddev, max,
extremum,
pct75, IQR

13

stddev

stddev

stddev,
extremum

stddev

stddev, pct95

14
15

weightn,
n = 11,12,14,16
mean, stddev,
max, pct50,
extremum
stddev, max,
extremum, pct75,
pct95
stddev
stddev, max,
min, pct50,
extremum, pct5,
pct25, pct95
stddev, max,
min, pct5, pct25,
pct95, IQR
mean, stddev,
max, extremum,
pct75, pct95,
IQR
stddev

16

stddev

stddev, min,
pct50

17

stddev

max

stddev, max,
extremum,
pct75, IQR

mean, stddev,
max, min,
extremum, pct5,

18
19

stddev, max,
extremum

stddev,
max,
extremum

Total (t)

𝑠𝑠𝑡𝑡6
𝑠𝑠𝑡𝑡8

20

stddev

21
Subtotal
910

15

25

33

pct25, pct75,
pct95, IQR
mean, stddev,
max, extremum,
pct75, pct95,
IQR
stddev
87

2

Table A5. Average performance metrics of the ML models in 10-fold cross-validation, where
the best value in each row is in bold
No. of selected features
New building
Precision
Renovation
Overall
New building
Recall
Renovation
Overall
New building
F1 score
Renovation
Overall

Decision tree
162
0.77
0.89
0.83
0.66
0.93
0.80
0.71
0.91
0.81

915

Table A6. List of 45 decisive features selected for Bayesian Networks from the pool of 821
features
Group
(subtotal)

Feature
type id

Daily (d)

Weekly (w)

Feature
Monthly (m)

stddev

stddev

1
2
Truck usage
behaviors
(17)

3
4

stddev
stddev

5

stddev

Yearly (y)
stddev
max, extremum,
pct75, pct95,
IQR
stddev
stddev
max, extremum,
pct95

6
7
Waste
disposal
behaviors
(20)

𝑠𝑠𝑡𝑡5

weightn,
n = 11,15,16

8
9

Total (t)

IQR

stddev

pct75, IQR

stddev

10
11
12

stddev

13

stddev

stddev
stddev, max,
extremum, pct75,
pct95, IQR

14
Facility
usage
behaviors
(8)

15

stddev

16

stddev

17
18
19

stddev

20
21
Subtotal

stddev
stddev, max,
extremum, pct75,
pct95, IQR
stddev

1

2

9

31

1

920

Table A7. Average performance metrics of the ML models in 10-fold cross-validation, where
the best value in each row is in bold
No. of selected features
New building
Precision
Renovation
Overall
New building
Recall
Renovation
Overall
New building
F1 score
Renovation
Overall

Decision tree
45
0.34
0.98
0.66
0.86
0.81
0.84
0.49
0.89
0.74