You are on page 1of 15

Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Contents lists available at ScienceDirect

Journal of Infrastructure Intelligence and Resilience


journal homepage: www.journals.elsevier.com/journal-of-infrastructure-intelligence-and-resilience

Machine learning for all! Benchmarking automated, explainable, and


coding-free platforms on civil and environmental engineering problems
M.Z. Naser a, b, *
a
School of Civil & Environmental Engineering and Earth Sciences (SCEEES), Clemson University, USA
b
Artificial Intelligence Research Institute for Science and Engineering (AIRISE), Clemson University, USA

A R T I C L E I N F O A B S T R A C T

Keywords: One of the key challenges in fully embracing machine learning (ML) in civil and environmental engineering
Coding-free platforms revolves around the need for coding (or programming) experience and for acquiring ML-related infrastructure.
AutoML This barrier can be overcome through the availability of various platforms that provide automated and coding-
Civil and environmental engineering
free ML services, as well as ML infrastructure (in the form of a cloud service or software). Thus, engineers can
Machine learning
now adopt, create, and apply ML to tackle various problems with ease and little coding. From this view, this paper
presents a comparison of five automated and coding-free ML platforms: BigML, Dataiku, DataRobot, Exploratory,
and RapidMiner on civil and environmental engineering problems. This comparison shows that although these
platforms differ in their setup, services, and provided ML algorithms, all platforms performed adequately and
comparably well to each other and to coding-based ML analyses. These findings denote the usefulness of coding-
free ML platforms, which can lead to a brighter future for ML integration into our domain.

1. Introduction (Kuhn and Johnson, 2013). Without such familiarity, the use of ML be-
comes challenging – especially when an engineer needs to create new ML
Machine learning (ML) continues to grow at an astonishing rate. solutions or algorithms from scratch. In such an event, the lack of pro-
Recent advancements have opened the door for new and exciting op- gramming knowledge further complicates the adoption of ML.
portunities to integrate ML into civil and environmental engineering A look into our domain indicates that most civil and environmental
(Hsieh and Tsai, 2020). On the one hand, ML has been shown to actively engineering curricula lack emphasis on ML or coding/programming.
complement the more favored traditional approaches of conducting While programming is often taught in the first or second year to young
physical tests and numerical simulations. For example, recent works engineering students, many do not fully embrace it or utilize it in later
clearly demonstrate how ML can be meshed with physics (Vassallo et al., courses or throughout their careers (Naser, 2022a). This lack of pro-
2020), verify domain knowledge (al-Bashiti and Naser, 2022), and gramming or ML skills has become a growing challenge in integrating ML
bypass bottlenecks associated with theoretical derivation and complex into our domains (Xu and Ouyang, 2022).
numerical simulations (Ying et al., 2013). A similar observation, but to some extent, can be made with regard to
The traditional approach to ML is to collect data pertaining to a set of numerical models. For example, graduate engineering students are often
variables (also called features) and observations (outcomes). These are introduced to the principles of the finite element method (FEM) using 1D
then treated and input into an ML algorithm by means of coding (i.e., and 2D elements. While 3D elements are rarely covered in lecture halls,
programming). A code script lists a procedure for handling the input and the principles covered in lower-dimensional elements are often directly
one that is executed to train and verify the performance of an algorithm. transferable to higher-dimensional elements. Engineers then become
Finally, the results of the analysis are examined. familiar with the different types of elements – more so when using
As one can see, and aside from data collection, a major component of complex finite element software, many of which provide detailed de-
the development of ML algorithms revolves around coding. This effec- scriptions of elemental packages in their software help manuals. Navi-
tively implies that a user (e.g., an engineer) has some degree of famil- gating such manuals, given our engineers' background, is manageable
iarity and/or mastery of a programming language (i.e., Python, R, etc.) (Kurowski, 2011).

*School of Civil & Environmental Engineering and Earth Sciences (SCEEES), Clemson University, USA. Website: www.mznaser.com
E-mail address: mznaser@clemson.edu.

https://doi.org/10.1016/j.iintel.2023.100028
Received 2 December 2022; Received in revised form 28 December 2022; Accepted 7 January 2023
Available online xxxx
2772-9915/© 2023 The Author(s). Published by Elsevier Ltd on behalf of Zhejiang University and Zhejiang University Press Co., Ltd. This is an open access article
under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 1. Flowchart of a typical AutoML coding-free platform.

More practically, engineers are trained to operate such software. application of ML to data collection and model development problems
Their training starts by learning how to use a particular finite element once a problem and hypothesis are identified, wherein data pertaining to
software, its elements, solution criteria, settings required, etc. To become the engineering problem on hand are collected and input into an
proficient in such software, an engineer may attend a series of lectures/ already-built ML algorithm.
seminars or complete a certified program. The same tactic can also be This algorithm can become a model once it has been fully trained and
followed when learning statistical or specialized software. Evidently, an validated. A working model can then be used to address the civil and
engineer becomes accustomed to operating software with good knowl- environmental engineering problems. Suppose such a model, along with
edge of the fundamentals of such software and the finite element method its data and settings, is accessible to all involved stakeholders. In this
(e.g., the author suspects that few engineers are likely to solve a 3D FE case, the scalability of such a model can grow beyond the primary en-
model or to fit a nonlinear regression/classification model by hand or gineer or firm that creates such a model. Fortunately, the open literature
from first principles on a regular basis) (Bloomfield, 2014). homes a number of works that planted the successfully created auto-
The inadequate attention given to programming and ML severely limits mated ML pipelines, which can be found as standalone multicode scripts,
our engineers' ability to visualize and adopt ML in our area. Hence, at the software, or cloud services (Paszke et al., 2017; Naser, 2022b; Tensor-
moment, practicing engineers are less likely to utilize ML (whether Flow, 2022; Zhou et al., 2020; Xin et al., 2021).
through application or creation). This arises despite the recent efforts of This study compares five coding-free ML platforms, namely BigML,
civil and environmental engineering societies such as the American Soci- Dataiku, DataRobot, Exploratory, and RapidMiner, to show how such a
ety of Civil Engineers (ASCE), which recently identified “Innovative tech- platform would simplify the adoption and integration of ML into our
nologies that will transform the future of structural engineering include, but are domain. This comparison highlights the key characteristics of each
not limited to, artificial intelligence, high performance computing …” in its platform regarding the available algorithms, types of analysis, explain-
recent Task Committee Report to the Structural Engineering Institute ability measures, and other individual features. This paper then dives
Board of Governors (S. Engineering Institute of the American Society of into addressing pressing questions pertaining to the future of ML given
Civil Engineers, 2019). the rise of such coding-free platforms.
One possible approach to overcoming the limitations of ML adoption is
to follow a tactic similar to that adopted in modeling software. By 2. Description of AutoML and coding-free platforms
providing engineers with the fundamentals of ML and training them to
properly use coding-free ML platforms, then the need to learn to code is 2.1. An overview of AutoML
likely to lessen significantly – and may not be warranted to a fuller extent
when taking the above example of 3D FE models into account. The need for automated machine learning (AutoML) arises from the
This study advocates a smoother path to facilitate the adoption of ML. need to present an accessible means for users to create and apply ML with
By shifting the focus and efforts of our engineers from the narrow space of minimal need for coding or ML infrastructure. This accessibility is facili-
the need to learn ML by programming to learning ML by practice, the path tated by the presence of a graphical user interface that permits the user to
to adopting ML becomes a bit more direct. Equally, by shifting the need for easily modify algorithmic architecture and settings. The same also allows
engineers to develop and acquire specialized ML infrastructure, the use of users to explore a variety of algorithms and procedures instead of relying
existing infrastructure (i.e., automated platforms) can lower initial capital on a particular algorithm or programming language. In a way, a platform
investments – which in turn allows for a more accessible culture to ML.1 can be designed to house algorithms of different programming languages
There are two fronts of this advocacy: the first, and just as in expertise (such as Python, R, and others) into one space – thereby allowing an en-
in the finite element method, only a small percentage of engineers are gineer to maximize the use of the various available ML tools.
likely to become proficient ML coders/programmers. More importantly, In an AutoML approach, users have access to and use already built-in
the second lies in the fact that most of our problems likely to be tackled algorithms and ML recipes. In this approach, data are first input into a
via ML have the potential to benefit from existing and traditional ML platform. Then, the engineer can opt to utilize a specific ML algorithm or
algorithms with little tuning (Deka, 2019). This further simplifies the can simply ask the platform to recommend one/a group of algorithms. An
algorithm is then set for training, and the testing process starts and com-
pletes automatically. Subsequently, the user/engineer will have immediate
1
As expected, a shift of such magnitude is likely to be accompanied by new or access to a variety of ML findings in terms of model predictions and insights
unanticipated experiences – these are likely to be addressed in future works and regarding model performance. Evidently, an engineer can then deploy a
discussions. validated model to predict or examine a phenomenon of interest.

2
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

A typical workflow in an AutoML platform begins by applying simple Table 1


ML algorithms (see Fig. 1). Such algorithms often have a simple archi- General comparison between platforms.
tecture and are not only easy to use but have been well-vetted as well. BigML DataRobot Dataiku Exploratory RapidMiner
Adopting such models is often considered to satisfy Occam's razor, a
Analysis type S/U S/U S/U S/U S/U
implying that, with all things being equal, a simple model is preferred AutoML     
over a more complex one (Domingos, 1999). Some of such models can Explainability   
vary between decision trees, k-nearest neighbors, etc. measures
In lieu of simple models, complex models, namely layered algorithms Energy  
measures
such as deep learning or boosting/boosted trees (i.e., XGBoost, etc.), can also Code export  
be tried and examined when/if simple models fail to deliver an appropriate Cost S/Fb S/F S/F Pc S/F
model. Similarly, other types of more complex models can also be developed. Fairness 
Such models can be created by blending simple models into an ensemble Ease of use     
Data type Tabular Tabular Tabular Tabular Tabular
(Ganaie et al., 2021; Abdollahi-Arpanahi et al., 2020; Hamori et al., 2018).
and and
Such an ensemble is likely to combine the strengths and minimize the footage footage
weaknesses of each individual algorithm involved in such an ensemble. Programming Multiple Multiple Python, R Multiple
Whether simple, complex, or ensemble models are used, a platform language(s) R, Julia,
can automate such a process and compare and rank the performance of and
Scala
all the created models. This comparison is often displayed in a Cloud or Cloud Cloud Cloud Cloud/ Cloud/
leaderboard-type format, which presents the performance metrics ob- software software software
tained from each algorithm across its training and testing stages. The a
S/U: supervised and unsupervised learning.
best-ranking model is often nominated for ML analysis – equally, a few b
Free to students/faculty.
models can also be nominated to create an ensemble. The reader is c
Free to the public.
advised to review the following works for additional details on AutoML
(He et al., 2021; Guyon et al., 2019). based automated ML services. This platform contains a wide variety of
Given that most ML models are blackboxes, the need to explain model algorithms belonging to different programming languages and integrates
predictions is critical. Nowadays, a few measures can be employed to explainable ML, deployable ML, fair ML, and composable ML. DataRobot
explain ML models, i.e., feature importance and partial dependence plots. supports tabular and footage data.
Such measures articulate the reasoning behind model performance and
may lead to new insights that could pinpoint new research questions 2.2.4. Exploratory
(Naser, 2021a). Upon the selection of a model and explaining its per- Exploratory5 is based in California, USA (Exploratory, 2022). Un-
formance, such a model can be deployed and used. like the other platforms, Exploratory is available in a software manner
built using R. Exploratory provides free academic licenses to students
2.2. Selected platforms and faculty members. One of the features of Exploratory is its ability to
download and share projects with other users freely. The version used
To maintain a concise comparison, five such platforms are selected herein was 6.8.6.
herein, as they provide free access and academic support to students and
faculty. These platforms include BigML, Dataiku, DataRobot, Exploratory, 2.2.5. RapidMiner
and RapidMiner. While each platform is highlighted here, it should be noted RapidMiner6 was first released in 2006 by its two founders, Ingo
that these platforms were not reviewed indepth, as such a task would Mierswa and Ralf Klinkenberg. This platform has been recently acquired
exponentially enlarge this study. Thus, interested readers are invited to visit by Altair Engineering and continues to provide free academic licenses.
such platforms and review their complete and cited documentation. The version used herein was RapidMiner Studio 9.10.011.
It goes without saying that the presented analysis is not intended to
2.2.1. Big ML distinguish the best performing platform but rather to benchmark each
BigML2 was founded in 2011 and is headquartered in Corvallis, platform for civil and environmental engineering problems. Further-
Oregon, US. BigML's mission is to make ML easy and beautiful for everyone more, this comparison hopes to document the merits of utilizing the
(BigML, 2022). This platform has a strong academic program that sup- platforms listed above by engineers with little coding experience. A cross-
ports student education and research. Unlike other platforms, BigML al- comparison of all five platforms is presented in Table 1. Given the variety
lows users to export the created ML models in various languages, thus of available algorithms in each platform, we opted to showcase those
emphasizing the notion of portable ML. Similarly, Big ML supports found across multiple platforms in their default settings. This rationale
explainable ML. This platform primarily supports tabular data. serves two fronts: 1) to minimize the influence of user input and expe-
rience and 2) to allow easy reproduction of this work. The reader is to
2.2.2. Dataiku note that our testing examined all platforms in their academic versions as
Dataiku3 is a portmanteau for Data and Haiku – which is a short, these versions were available to us. Examining the full potential of each
structured Japanese poem (Dataiku, 2014). This platform originated in platform is better suited for a series of separate and future works.
2013 with its four founders, Florian Douetteau, Clement Stenac, Thomas
Cabrol, and Marc Batty. Dataiku has multiple offices worldwide and 3. Description of ML algorithms and measures
provides a wealth of information and documentation. This platform
supports tabular data and footage as well as explainable ML. The selected algorithms in this study are briefly described herein, and
their full description can be found in their respective references, as cited
2.2.3. DataRobot below, as well as our recent review (Tapeh and Naser, 2022). All algo-
DataRobot4 is a Boston-based data science firm that provides cloud- rithms were used in their default settings as available in each platform.
This decision stems from the fact that tuning these algorithms to be

2
https://bigml.com.
3 5
https://www.dataiku.com/. https://exploratory.io.
4 6
https://datarobot.com. https://rapidminer.com/.

3
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

consistent among all platforms is not possible since each platform opti- 3.7. Explainability via feature importance
mizes its algorithms differently (i.e., the same settings are not available
on each platform). Despite such a minor technical difficulty, the author As mentioned above, most ML models can be considered blackboxes
believes that using the same algorithms (with slightly varied architec- (Rudin, 2019). Recent research has shown that a model can potentially
ture) can still provide the readers of this paper with a general and concise be interpreted by understanding the influence of its own features, as such
view of these platforms' predictivity and ease of use. features drive the model's outcome.
From this perspective, a feature can be declared important if a given
3.1. Decision tree (DT) model relies heavily on such a feature to correctly predict the outcome.
This can be measured by evaluating a model's error increase after sys-
The decision tree (DT) is a simple algorithm that resembles a tree tematically permuting all features (Altmann et al., 2010). Thus, a feature
structure. This algorithm aims to minimize the amount of Gini impurity is declared “important” if permuting its values increases the model error,
(g, a measure of how often a randomly chosen data point would be effectively indicating that the model relied on this feature to arrive at a
incorrectly labeled if it was randomly labeled per the distribution of the good prediction. In this paper, given the different approaches that each
subset). The Gini impurity can be found in (Naser, 2021b). All the platform uses to calculate the importance of its features, we will present
selected platforms contain this algorithm. the relative importance of each feature to the most important feature. As
commonly noted, feature importance is often used to identify the relative
importance of features as opposed to interpreting the value of such
3.2. Random forest (RF) importance (see (Murdoch et al., 2019; Barredo Arrieta et al., 2020)).
The reader is to note that other explainability measures (i.e., partial
The Random Forest (RF) algorithm is an ensemble learner that forms dependence plots (Inouye et al., 2020) and individual conditional ex-
a series of decision trees (Liaw and Wiener, 2002). This algorithm then pectations (Goldstein et al., 2015)) also exist; however, such measures
averages the outcomes of all trees to arrive at a prediction. This algorithm are not available in all platforms (unlike the feature importance mea-
is available in all selected platforms. sure). Further information on the history and background of such mea-
sures can be found elsewhere (Molnar, 2019).
3.3. Gradient boosted trees (GBT)
4. Technical details and performance metrics

The Gradient Boosted Trees (GBT) is a simple algorithm with a tree-


This section documents the main technical details obtained during the
like structure (Natekin and Knoll, 2013). While training, this algorithm
ML analysis. In all platforms (except Exploratory, which does not support
builds additive models and uses the initial error to optimize arbitrary
this feature in default settings), a 10-fold cross-validation technique was
differentiable loss functions. This process reduces the errors in successive
adopted. In this cross-validation technique, a training set is split into ten
models and hence lowers the overall error in the model.
equal-sized subsets. In each sub-iteration, nine subsets are used for
training, and the remainder of these subsets are kept for validation until
3.4. Extreme gradient boosted trees (XGboost)7 all subsets are used (Pedregosa et al., 2011). For transparency, there was
no control for the same data to be split and applied across all the algo-
The XGboost was created by Chen and Guestrin (2016) as an rithms. However, the use of cross-validation is thought of as a means to
improvement on gradient boosted trees. This tree-based algorithm adopts a minimize such an issue, as noted in (Kohavi, 1995; Wong and Yang,
weighted quantile sketch (an approximation algorithm used to split the 2017), especially since the used datasets are not considered small and
candidates in a tree). The XGBoost also adopts a sparsity-aware split finding satisfy the three independent criteria to be discussed in Sec. 5.0.
to handle sparse data and missing values and contains a built-in pre-sorted The performance of the ML models created in all of the above five
and a histogram-based algorithm for computing the best split in each tree platforms is evaluated via a number of metrics. For regression problems, the
(Scikit sklearn.ensemble.GradientBoostingRegressor, 2020; XGBoost Py- metrics included the mean absolute error (MAE), root mean squared error
thon Package, 2020). All the selected platforms contain this algorithm. (RMSE), and coefficient of determination (R2). In general, lower values of
MAE and RMSE and values close to positive unity for R2 are favorable.
3.5. Light gradient boosting machine (LGBM)
Table 2
List of common performance metrics.
The Light Gradient Boosting Machine (LGBM) is another tree-based
Metric Formula
algorithm. This algorithm was created by Microsoft (Ke et al., 2017).
Like the XGBoost, the LGBM is also an improvement over the gradient Regression
Pn
i¼1 jPi  Ai j
Mean absolute error
boosted trees via two techniques, namely, the gradient-based one-side MAE ¼
(MAE) n
sampling (to identify the most informative data points) and exclusive sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Root mean squared Pn 2
feature bundling (to group data points in a near-lossless way). This al- i¼1 ðPi  Ai Þ
error (RMSE) RMSE ¼
n
gorithm is available from Dataiku and DataRobot. Coefficient of Pn P n
R2 ¼ 1  ðPi  Ai Þ2 = ðAi  Amean Þ2
determination (R2) i¼1 i¼1
A: actual measurements, P: predictions, n: number of data
3.6. Artificial neural network (ANN)
points.
Classification
A typical neural network comprises multiple layers, with each layer Accuracy TP þ TN
ACC ¼
containing a number of neurons. The first layer is fed the input data and PþN
P: predictions, N: number of real negatives, TP: number of
is connected to the second layer (called the hidden layers (Bishop,
true positives, TN: number of true negatives.
2007)). The hidden layer(s) have the ability to establish linear and/or Logloss error P
M
nonlinear relations and are connected to the last (output) layer, which LLE ¼  Ai logP
c¼1
displays the outcome of the ANN analysis. M: number of classes, c: class label, y: binary indicator (0 or
1) if c is the correct classification for a given observation.
Area under the ROC P1
N1
AUC ¼ ðFPiþ1  FPi ÞðTPiþ1  TPi Þ
7 curve i¼1 2
The reader should note that the terms gradient boosting trees and gradient
FP: number of false positives, FN: number of false negatives.
boosted trees are noted in the literature.

4
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Table 3 5. Description of datasets and benchmarking


Details of the adopted datasets.
Dataset No. of No. of Data typea Ref. This section describes the six datasets used in the study, as listed in
features data Table 3. The datasets cover supervised learning problems pertaining
Inputs Output
points to civil engineering. Three datasets were used in the regression, and
Regression the other three were used for classification. For completion, each of
Compressive 9 1030 N N Yeh (1998) the selected datasets was deemed to satisfy the recommendations of
strength of ultra three recent studies intended to ensure the useability and health of the
high
performance
data.
concrete
Axial strength of 5 1260 N N Thai et al.  van Smeden et al. (2018) – having a minimum set of 10 observations
CFST columns (2019a, 2019b) per feature.
Predicting CO2 4 945 M N Open Canada
 Riley et al. (2019) – having a minimum of 23 observations per feature.
emissions in (2022)
vehicles  Frank and Todeschini (1994) – maintaining a ratio of 3 and 5 between
Classification the number of observations to the number of features.
Concrete mixtures 12 646 C C Liu et al.
for fire-induced (2021a), Liu A brief description of each dataset is provided to present key statis-
spalling and Zhang
(2020)
tical insights. It is worth noting that most of these datasets were recently
Potability of 10 3277 N C Sadhwani benchmarked in independent studies; hence, this provides an opportu-
drinking water (2022) nity to cross-examine and compare the performance obtained from the
Failure in bridges 8 299 M C Abedi and selected platforms.
Naser (2021),
Kodur and
Naser (2013) 5.1. Dataset on the compressive strength of high-performance concrete
a
N: numerical data, C: categorical data, M: mixed (numerical and categorical)
data. 5.1.1. Description of the dataset
This dataset compiles over 1030 concrete mixtures taken from tests
In addition, other types of metrics were used for classification prob- that determined the compressive strength (fc) of ultra high performance
lems. These include accuracy, logloss error, and the area under the concrete (UHPC) (Yeh, 1998). Each mixture is specified in terms of its
receiver operating characteristic (ROC) curve (AUC). Naturally, higher content of Cement, C, Blast Furnace Slag, B, Fly Ash, F, Water, W,
values of accuracy and AUC and lower values of the logloss metrics are Superplasticizer, S, Coarse Aggregate, CA, Fine Aggregate, FA, and Age,
favorable. The above-selected metrics are commonly adopted and A. Fig. 2 displays statistical insights about this dataset in terms of mini-
accepted in structural engineering publications and are listed in Table 2 mum, maximum, median, mean and standard deviation. This is a
(Naser and Alavi, 2021). For the sake of uniformity, the average of regression case study in which the target is the compressive strength of
metrics between training and testing will be shown in each case study concrete. All features include numeric values without any missing data.
(since some platforms do not always report these metrics separately). This case study used the XGBoost algorithm.

Fig. 2. Insights into the dataset.

5
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Table 4 the effective length Le, tube thickness t, tube diameter D, yield stress fy,
Results of benchmarking analysis. compressive strength fc, and the outcome, i.e., axial strength N, were
Platform Metrics recorded. Fig. 4 shows the distribution of all the selected features. As one
can see, all features in this dataset encompass numeric values without
MAE RMSE R2
any missing data. For example, the range for the diameter and thickness
BigML 2.79 – 0.920 of the CFSTs varies between 44.45 and 1020 mm, and 0.52 and 16.54
Dataiku 3.69 4.791 0.924
DataRobot 3.12 4.363 0.937
mm, respectively. In this case study, the LGBM algorithm was used in
Exploratory – 4.350 0.925 DataRobot and Dataiku. This particular algorithm is not available on the
RapidMiner – 3.543 0.954 other platforms, and hence, a boosted tree-like version is used instead.

5.2.2. Benchmarking and comparison of feature importance


5.1.2. Benchmarking and comparison of feature importance
The results of the analysis are tabulated in Table 5. A holistic look at
The results of this analysis are listed in Table 4. As one can see, all
this performance shows that the five platforms realized comparable
platforms scored decently with an accuracy (in terms of R2) that exceeded
outcomes – with Exploratory and BigML scoring the highest R2 (and
92%. Collectively, the DataRobot platform scored the highest when
Exploratory scoring the lowest RMSE). Fig. 5 shows the feature impor-
comparing all metrics, while BigML also scored the best when only MAE
tance obtained for each platform. Herein, the agreement among all
was considered. RapidMiner scored the best on the metrics of RMSE and
platforms is clear and consistent, wherein the diameter of CFSTs is the
R2. On the feature importance front, all platforms identified the two
most important feature.
features with the highest importance (i.e., C and A). The importance of
The same dataset was used to train and build the LGBM algorithm.
the other features is much lower than that of C and A – see Fig. 3.
The benchmarking of this dataset resulted in an average of 158.57,
It is worth noting that this dataset first appears in Yeh (1998). In this
664.49, and 0.953 for MAE, RSME, and R2, respectively. Furthermore,
work, various neural networks were built. Yeh reported that one metric,
this dataset has also been examined in other studies, and models,
R2, of such networks varied between 0.855 and 0.945. A more recent
including Zarringol (Zarringol et al., 2021; Zarringol and Thai, 2022)
benchmarking study (Naser et al., 2021a) examined this dataset using a
used support vector machines and neural networks and reported a per-
tuned XGBoost algorithm as well. The resulting model scored an average
formance of R2 at 0.961 and 0.991. Genetic algorithms and genetic
of 2.95, 4.17, and 0.93 for MAE, RSME, and R2, respectively. As one can
programming were also used by (Naser et al., 2021b) and yielded 0.991
see, the performance obtained from the coding-free platforms is similar
and 0.958, respectively. It is clear that the performances obtained from
to those reported by the two benchmarking studies.
these studies and those obtained from coding-free platforms are quite
similar.
5.2. Dataset on axial strength of concrete-filled steel tubular (CFST)
columns 5.3. Dataset for predicting CO2 emissions in vehicles

5.2.1. Description of the dataset 5.3.1. Description of the dataset


This dataset contains 1260 observations from experiments on This regression dataset contains 945 observations, four features (en-
concentrically loaded circular concrete-filled steel tubular (CFST) col- gine size, number of cylinders, and fuel consumption in the city), and one
umns (Thai et al., 2019a, 2019b). In this dataset, five features, namely, target, CO2 emissions. The graphical distribution of all the features is

Fig. 3. Comparison of feature importance.

6
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 4. Insights into the dataset.

5.3.2. Benchmarking and comparison of feature importance


Table 5 The results of the analysis are listed in Table 6. All platforms scored
Results of benchmarking analysis. within a few percentages, thus emphasizing good attainment of accuracy.
Platform Metrics Fig. 7 also shows a similar degree of alignment between all platforms,
MAE RMSE R2 which identified fuel consumption as the primary feature of importance
tied to correctly predicting the generation of CO2 emissions. Given this
BigML 209.01 – 0.980
Dataikua 258.00 822.00 0.954
dataset's recent release, we could not find other studies that have
DataRobota 214.4 910.80 0.949 examined this dataset.
Exploratory – 471.10 0.990
RapidMiner – 647.41 0.981
5.4. Dataset on classifying concrete mixtures for fire-induced spalling
a
LGBM.
5.4.1. Description of the dataset
plotted in Fig. 6. These data were generated from the Canadian fuel This case study showcases a dataset containing 646 observations from
consumption rating dataset rating for vehicles produced in 2022 (Open fire tests (Liu et al., 2021a; Liu and Zhang, 2020). These data have 12
Canada, 2022). The analysis of this case study utilizes the features tied to influence fire-induced spalling and one dependent vari-
gradient-boosted tree algorithm. able (in two labels: no spalling or spalling). These variables are water/-
binder ratio (%), aggregate/binder ratio (%), sand/binder ratio (%),

7
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 5. Comparison of feature importance.

heating rate ( C/min), moisture content, maximum exposure tempera- calcium and magnesium salts), solids, chloramines, sulfate, conductivity,
ture ( C), silica fume/binder ratio (%), aggregate size (mm), organic carbon, trihalomethanes, and turbidity to predict the potability
ground-granulated blast-furnace slag (GGBS)/binder ratio (%), fly ash of water (i.e., bad quality water or good quality water) (Sadhwani, 2022).
(FA)/binder ratio (%), and polypropylene (PP) fiber quantity (kg/m3). Evidently, this is a binary classification case study. A decision tree al-
Fig. 8 displays the graphical distribution of all the selected numerical gorithm was used in this case study. Fig. 10 shows the distribution of
features and one categorical outcome. This case study aims to use default each selected feature, and as seen, almost all of the features have a uni-
neural networks from the selected platforms to classify concrete mixtures form distribution. One can also see that some features (i.e., pH, sulfate,
according to whether they would spall or not. and trihalomethanes) have missing values. Such missing data were kept
on purpose to evaluate the performance of the platforms when dealing
5.4.2. Benchmarking and comparison of feature importance with missing data. The default option in the selected platforms is to
Three of the selected platforms, BigML, DataRobot, and Dataiku, pro- ignore such data.8
vide options for neural networks. As such, results from these three plat-
forms are showcased herein (see Table 7). Overall, BigML and DataRobot 5.5.2. Benchmarking and comparison of feature importance
scored the highest. With regard to the feature importance analysis, one Overall, all platforms seem to perform similarly and achieve compa-
can see that the PP quantity and moisture content, as well as the heating rable performances (see Table 8). Comparing all metrics, it can be seen
rate and exposure temperature, are the most important features (see that Exploratory ranked highest in terms of accuracy. However, Dataiku
Fig. 9). The reader is to note that Dataiku does not provide feature seems to have produced the most consistent performance among the
importance for its default neural network. three selected metrics, followed by BigML in two metrics.
The spalling dataset was explored via the XGBoost model recently The results of the feature importance analysis are shown in Fig. 11. It
created by (al-Bashiti and Naser, 2022). In this study, the reported per- can be seen that there is good agreement between all the models, espe-
formances of this model in terms of accuracy and AUC were 92.5 and cially with regard to the top features. Despite this comparison, the
0.887, respectively. This dataset was also examined by the original re- outcome of the BigML model seems to score most features with lower
searchers (Liu et al., 2021b) using an artificial neural network, which importance, on average, when compared to the other platforms. This
yielded an accuracy between 90 and 100%. dataset was benchmarked by its originator with a metric reported to have
an accuracy of 0.59 (Sadhwani, 2022).

5.5. Dataset on classifying water potability given their chemical


composition

5.5.1. Description of the dataset


This classification dataset comprises 3276 observations in terms of 8
The reader is to note that other options exist such as (replacing missing
water pH value, hardness (a measure of mineral content, especially values with the average, etc.).

8
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 6. Insights into the dataset.

vulnerable bridges. Thus, in a recent survey (Abedi and Naser, 2021)


Table 6
reviewed 299 international bridge incidents (211 incidents covered
Results of benchmarking analysis.
bridges in the US, and 88 bridges were collected from around the world)
Platform Metrics during five different extreme events: fire, wind, scour/flood, earth-
MAE RMSE R2 quake, or impact/collision. In this survey, features pertaining to the
BigML 5.66 – 0.980
physical features of bridges (i.e., structural system, span, age, and
Dataiku 5.77 7.752 0.981 construction materials), traffic features served by the bridge
DataRobot 6.50 9.370 0.978 (geographical significance and number of lanes), and the damage level
Exploratory – 12.50 0.958 (collapse, major and minor damage) imposed upon the bridge were
RapidMiner – 11.41 0.960
collected. These features are shown in Fig. 12. It is quite clear that,
unlike the other datasets, this particular dataset contains a mixture of
features that vary in terms of numerics and categoricals, with the latter
5.6. Dataset on classifying bridge failures being the majority of the features. Thus, this dataset is representative of
a realistic scenario where an engineer has to tackle a mixture of data
5.6.1. Description of the dataset types. This multi-classification case study used the Random Forest
In the following works (Khan et al., 2021; Naser and Kodur, 2015), algorithm.
several features were identified to be of high merit to distinguish

9
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 7. Comparison of feature importance.

Fig. 8. Insights into the dataset.

5.6.2. Benchmarking and comparison of feature importance


Table 7 In this case study, the BigML model seems to climb the ranks with the
Results of benchmarking analysis. highest accuracy and AUC (see Table 9). For the feature importance front,
Platform Metrics as shown in Fig. 13 shows that all platforms agree on labeling Span and
Accuracy Logloss error AUC Age as the two main features with the highest importance. Beyond that,
the importance of the other features is somehow consistent only among
BigML 0.884 – 0.962
Dataiku 0.785 0.351 0.885 to 2–3 platforms. It is worth noting that Exploratory only identified three
DataRobot 0.891 0.261 0.947 important features (Location, Age and Lane). The original dataset first
Exploratory – – – appeared in a recent study (Abedi and Naser, 2021). It is worth noting
RapidMiner 0.610 – 0.518 that the overall accuracy reported by deep learning was between 0.700
and 0.896%.

10
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 9. Comparison of feature importance.

Fig. 10. Insights into the dataset.

Table 8
6.0. Discussion and results Results of benchmarking analysis.
Platform Metrics
The above analysis shows that the predictions from the selected
AutoML coding-free platforms are very similar to those from coding- Accuracy Logloss error AUC
based models. Overall, the difference between these models is suffi- BigML 0.617 – 0.610
ciently small (within a few percentages) when all are used in their default Dataiku 0.625 0.926 0.597
settings. A clear advantage in regression problems is noticed, as opposed DataRobot 0.412 0.679 0.575
Exploratory 0.681 – –
to platforms' performance on classification problems – wherein the RapidMiner 0.614 – 0.510

11
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Fig. 11. Comparison of feature importance.

Fig. 12. Insights into the dataset.

12
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

Table 9
Results of benchmarking analysis.
Platform Metrics

Accuracy Logloss error AUC

BigML 0.361 – 0.423


Dataiku 0.288 1.350 0.518
DataRobot 0.398 1.051 0.509
Exploratory 0.480 – –
RapidMiner 0.410 – –

Fig. 13. Comparison of feature importance.

observed performance seems to underperform against that seen in integration of ML into civil and environmental engineering areas.
coding-based models. However, from a practical point of view, all plat- Furthermore, this paper also presents elemental and pressing questions
forms perform adequately. pertaining to the future of ML in response to the rise of such coding-free
Collectively, this implies that using such platforms with default settings platforms. The following conclusions can also be drawn from the findings
will likely result in an adequate performance. Such modest performance, of this study.
when considering the ease of building models and the lack of specialized
programming, may present a case for an accessible and widespread ML. It  The integration of coding-free ML is likely to revive and ease the
is likely that tuning the examined models in the previous section can lead adoption of ML in the domain of civil and environmental engi-
to similar or possibly improved performance metrics compared to those neering. The coding-free option facilitates a more accessible use of
documented in the open literature – as pointed out by the reviewers of this ML for our engineers. On the other hand, the same may open the
work, especially with regard to hyperparameter optimization. The above door for an emphasis on the responsible and proper use of ML. Here
observation extends to charting a path for smoother facilitation and is where the role of education and practice becomes of utmost
adoption of ML. By leveraging existing platforms or possibly creating a importance.
unified platform for civil and environmental engineers, our engineers  The availability of coding-free ML presents a narrative for learning
could utilize ML via application in lieu of programming. ML by practice as opposed to learning ML by coding. This may lead to
a paradigm shift in engineering firms seeking to adopt ML in various
7.0. Conclusions projects and problems. Similarly, this will open the door for more
inclusive and modern curricula that can easily incorporate ML.
This paper presents a comparative study of five coding-free ML  Despite the different mechanisms built into the five examined plat-
platforms: BigML, Dataiku, DataRobot, Exploratory, and RapidMiner. The forms, all of them seem to perform equally well and comparatively
primary motivation behind this analysis is to demonstrate the merit of (e.g., within a few percentages) in terms of performance metrics
adopting such platforms to facilitate and simplify the adoption and across all datasets.

13
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

 While most platforms performed exceptionally well in their default Kodur, V.K.R., Naser, M.Z., 2013. Importance factor for design of bridges against fire
hazard. Eng. Struct. 54, 207–220. https://doi.org/10.1016/j.engstruct.2013.03.048.
settings, their performance was slightly outperformed by coding-
Kohavi, R., 1995. A study of cross-validation and bootstrap for accuracy estimation and
based models of the same nature. We anticipate that further tuning model selection. In: Proc. 14th Int. Jt. Conf. Artif. Intell., vol. 2.
of the platforms' model is likely to reduce such a gap. Kuhn, M., Johnson, K., 2013. Applied Predictive Modeling. https://doi.org/10.1007/978-
1-4614-6849-3.
Kurowski, P.M., 2011. Teaching finite element analysis for design engineers. In: Proc.
Data availability Can. Eng. Educ. Assoc.. https://doi.org/10.24908/pceea.v0i0.3839.
Liaw, A., Wiener, M., 2002. Classification and Regression by RandomForest. https://www
.researchgate.net/publication/228451484. (Accessed 8 April 2019).
Some or all data, models, or code that support the findings of this Liu, J.C., Zhang, Z., 2020. A machine learning approach to predict explosive spalling of
study are available from the corresponding author upon reasonable heated concrete. Arch. Civ. Mech. Eng. 20. https://doi.org/10.1007/s43452-020-
request. 00135-w.
Liu, J.-C.C., Huang, L., Tian, Z., Ye, H., 2021a. Knowledge-enhanced data-driven models
for quantifying the effectiveness of PP fibers in spalling prevention of ultra-high
performance concrete. Constr. Build. Mater. 299, 123946. https://doi.org/10.1016/
Declaration of competing interest j.conbuildmat.2021.123946.
Liu, J.C., Huang, L., Chen, Z., Ye, H., 2021b. A comparative study of artificial intelligent
The authors declare no conflict of interest. methods for explosive spalling diagnosis of hybrid fiber-reinforced ultra-high-
performance concrete. Int. J. Civ. Eng. https://doi.org/10.1007/s40999-021-00689-
7.
Acknowledgment Molnar, C., 2019. Interpretable Machine Learning. A Guide for Making Black Box Models
Explainable. Book.
Murdoch, W.J., Singh, C., Kumbier, K., Abbasi-Asl, R., et al., 2019. Definitions, methods,
We would like to thank BigML, Dataiku, DataRobot, Exploratory, and and applications in interpretable machine learning. Proc. Natl. Acad. Sci. U. S. A.
RapidMiner. All platforms were used in their default setting and in their https://doi.org/10.1073/pnas.1900654116.
academic editions. Naser, M.Z., 2021a. An engineer's guide to eXplainable artificial intelligence and
interpretable machine learning: navigating causality, forced goodness, and the false
perception of inference. Autom. Constr. 129, 103821. https://doi.org/10.1016/
References J.AUTCON.2021.103821.
Naser, M.Z., 2021b. Observational analysis of fire-induced spalling of concrete through
ensemble machine learning and surrogate modeling. J. Mater. Civ. Eng. 33,
Abdollahi-Arpanahi, R., Gianola, D., Pe~ nagaricano, F., 2020. Deep learning versus
04020428. https://doi.org/10.1061/(ASCE)MT.1943-5533.0003525.
parametric and ensemble methods for genomic prediction of complex phenotypes.
Naser, M.Z., 2022a. A faculty's perspective into infusing artificial intelligence to civil
Genet. Sel. Evol. https://doi.org/10.1186/s12711-020-00531-z.
engineering education. J. Civ. Eng. Educ. 148, 02522001. https://doi.org/10.1061/
Abedi, M., Naser, M.Z., 2021. RAI: rapid, autonomous and Intelligent machine learning
(ASCE)EI.2643-9115.0000065.
approach to identify fire-vulnerable bridges. Appl. Soft Comput. https://doi.org/
Naser, M.Z., 2022b. CLEMSON: an automated machine learning (AutoML) virtual
10.1016/j.asoc.2021.107896.
assistant for accelerated, simulation-free, transparent, reduced-order and inference-
al-Bashiti, M.K., Naser, M.Z., 2022. Verifying domain knowledge and theories on fire-
based reconstruction of fire response of structural members. ASCE J. Struct. Eng.
induced spalling of concrete through eXplainable artificial intelligence. Constr. Build.
https://doi.org/10.1061/(ASCE)ST.1943-541X.0003399.
Mater. 348, 128648. https://doi.org/10.1016/J.CONBUILDMAT.2022.128648.
Naser, M.Z., Kodur, V.K.R., 2015. A probabilistic assessment for classification of bridges
Altmann, A., Toloşi, L., Sander, O., Lengauer, T., 2010. Permutation importance: a
against fire hazard. Fire Saf. J. 76, 65–73. https://doi.org/10.1016/
corrected feature importance measure. Bioinformatics. https://doi.org/10.1093/
j.firesaf.2015.06.001.
bioinformatics/btq134.
Naser, M.Z., Alavi, Amir H., 2021. Error metrics and performance fitness indicators for
Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., et al., 2020. Explainable
artificial intelligence and machine learning in engineering and sciences. Archit.
explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and
Struct. Constr. 2021 (1), 1–19. https://doi.org/10.1007/S44150-021-00015-8.
challenges toward responsible AI. Inf. Fusion. https://doi.org/10.1016/
Naser, M.Z., Kodur, V., Thai, H.-T., Hawileh, R., Abdalla, J., Degtyarev, V.V., 2021a.
j.inffus.2019.12.012.
StructuresNet and FireNet: benchmarking databases and machine learning algorithms
BigML, 2022. About BigML.com. https://bigml.com/. (Accessed 6 July 2022).
in structural and fire engineering domains. J. Build. Eng., 102977 https://doi.org/
Bishop, C., 2007. Pattern recognition and machine learning. Technometrics. https://
10.1016/J.JOBE.2021.102977.
doi.org/10.1198/tech.2007.s518.
Naser, M.Z., Thai, S., Thai, H.-T., 2021b. Evaluating structural response of concrete-filled
Bloomfield, V.A., 2014. Using R for Numerical Analysis in Science and Engineering.
steel tubular columns through machine learning. J. Build. Eng., 101888 https://
https://doi.org/10.1201/b16841.
doi.org/10.1016/j.jobe.2020.101888.
Chen, T., Guestrin, C., 2016. XGBoost: a scalable tree boosting system. In: Proc. ACM
Natekin, A., Knoll, A., 2013. Gradient boosting machines, a tutorial. Front. Neurorobot. 7.
SIGKDD Int. Conf. Knowl. Discov. Data Min.. https://doi.org/10.1145/
https://doi.org/10.3389/fnbot.2013.00021.
2939672.2939785.
Open.Canada.ca, 2022. Fuel Consumption Ratings – Open Government Portal.
Dataiku, 2014. Pronouncing Dataiku. https://blog.dataiku.com/2014/08/07/pronouncin
https://open.canada.ca/data/en/dataset/98f1a
g-dataiku. (Accessed 27 July 2022).
129-f628-4ce4-b24d-6f16bf24dd64#wb-auto-6. (Accessed 27 July 2022).
Deka, P.C., 2019. A Primer on Machine Learning Applications in Civil Engineering.
Paszke, A., Gross, S., Chintala, S., Chanan, G., Desmaison, A., Antiga, L., Lerer, A., et al.,
https://doi.org/10.1201/9780429451423.
2017. Automatic differentiation in PyTorch. In: NIPS 2017 Autodiff Work. Futur.
Domingos, P., 1999. The role of Occam's Razor in knowledge discovery. Data Min. Knowl.
Gradient-Based MachLearn. Softw. Tech..
Discov. https://doi.org/10.1023/A:1009868929893.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., et al., 2011. Scikit-learn: machine
Exploratory, 2022. About. https://exploratory.io/about/ (Accessed 6 July 2022).
learning in Python. J. Mach. Learn. Res. 12, 2825–2830.
Frank, I., Todeschini, R., 1994. The Data Analysis Handbook. https://books.google.com/books
Riley, R.D., Snell, K.I.E., Ensor, J., Burke, D.L., et al., 2019. Minimum sample size for
?hl¼en&lr¼&id¼SXEpB0H6L3YC&oi¼fnd&pg¼PP1&ots
developing a multivariable prediction model: PART II – binary and time-to-event
¼zfmIRO_XO5&sig¼dSX6KJdkuav5zRNxaUdcftGSn2k. (Accessed 21 June 2019).
outcomes. Stat. Med. https://doi.org/10.1002/sim.7992.
Ganaie, M.A., Hu, M., Tanveer, M., Suganthan, P.N., 2021. Ensemble Deep Learning: A
Rudin, C., 2019. Stop explaining black box machine learning models for high stakes
Review.
decisions and use interpretable models instead. Nat. Mach. Intell. https://doi.org/
Goldstein, A., Kapelner, A., Bleich, J., Pitkin, E., 2015. Peeking inside the black box:
10.1038/s42256-019-0048-x.
visualizing statistical learning with plots of individual conditional expectation.
S. Engineering Institute of the American Society of Civil Engineers, 2019. A Vision for the
J. Comput. Graph. Stat. https://doi.org/10.1080/10618600.2014.907095.
Future of Structural Engineering and Structural Engineers. A Case for Change.
Guyon, I., Sun-Hosoya, L., Boulle, M., Escalante, H.J., et al., 2019. In: Analysis of the AutoML
Sadhwani, J., 2022. GitHub – JatinSadhwani02/Water-Quality-Prediction-using-ML-
Challenge Series 2015–2018. https://doi.org/10.1007/978-3-030-05318-5_10.
technique: in this post, I have implement a water quality prediction using machine
Hamori, S., Kawai, M., Kume, T., Murakami, Y., et al., 2018. Ensemble learning or deep
learning techniques. In: In This Technique, Our Model Predicts that the Water Is Safe
learning? Application to default risk analysis. J. Risk Financ. Manag. https://doi.org/
to Drink or Not Using Some Parameters like Ph Value, Conductivity, Hardness, Etc..
10.3390/jrfm11010012.
https://github.com/JatinSadhwani02/Water-Quality-Prediction-using-ML-technique
He, X., Zhao, K., Chu, X., 2021. AutoML: A Survey of the State-Of-The-Art, Knowledge-
. (Accessed 16 June 2022).
Based Syst. https://doi.org/10.1016/j.knosys.2020.106622.
Scikit, sklearn.ensemble.GradientBoostingRegressor — Scikit-Learn 0.24.1
Hsieh, Y.-A., Tsai, Y.J., 2020. Machine learning for crack detection: review and model
Documentation, 2020. https://scikit-learn.org/stable/modules/generated/sklearn.e
performance comparison. J. Comput. Civ. Eng. https://doi.org/10.1061/(asce)
nsemble.GradientBoostingRegressor.html. (Accessed 9 February 2021).
cp.1943-5487.0000918.
Tapeh, A., Naser, M.Z., 2022. Artificial intelligence, machine learning, and deep learning
Inouye, D.I., Leqi, L., Kim, J.S., Aragam, B., et al., 2020. Automated dependence plots. In:
in structural engineering: a scientometrics review of trends and best practices. Arch.
Proc. 36th Conf. Uncertain. Artif. Intell. UAI 2020.
Comput. Methods Eng. https://doi.org/10.1007/s11831-022-09793-w.
Ke, G., Meng, Q., Finley, T., Wang, T., et al., 2017. LightGBM: a highly efficient gradient
TensorFlow, 2022. A Neural Network Playground. https://playground.tensorflow.org/#a
boosting decision tree. Adv. Neural Inf. Process. Syst.
ctivation¼relu&batchSize¼10&dataset¼gauss&regDataset¼reg-plane
Khan, M.A., Khan, A.A., Anwar, G.A., Usmani, A., 2021. Framework for fire risk
&learningRate¼0.03&regularizationRate¼0&noise¼0&netw
assessment of bridges. Structures. https://doi.org/10.1016/j.istruc.2021.04.071.

14
M.Z. Naser Journal of Infrastructure Intelligence and Resilience 2 (2023) 100028

orkShape¼3,2,5,4,7,4&seed¼0.62778&showTestData¼true&discre Xin, D., Miao, H., Parameswaran, A., Polyzotis, N., 2021. Production
tize¼false&percTrainData¼50&x¼true&y¼true&xTimesY¼false machine learning pipelines: empirical analysis and optimization opportunities. In:
&xSquared¼false&ySquared¼false&cosX¼false&sinX¼false Proc. ACM SIGMOD Int. Conf. Manag. Data. https://doi.org/10.1145/
&cosY¼false&sinY¼false&collectStats¼false&problem¼regre 3448016.3457566.
ssion&initZero¼false&hideText¼false. (Accessed 4 July 2022). Xu, W., Ouyang, F., 2022. The application of AI technologies in STEM education: a
Thai, S., Thai, H.T., Uy, B., Ngo, T., 2019a. Concrete-filled steel tubular columns: test systematic review from 2011 to 2021. Int. J. STEM Educ. 9, 1–20. https://doi.org/
database, design and calibration. J. Constr. Steel Res. https://doi.org/10.1016/ 10.1186/S40594-022-00377-5/FIGURES/10.
j.jcsr.2019.02.024. Yeh, I.C., 1998. Modeling of strength of high-performance concrete using artificial neural
Thai, S., Thai, H., Uy, B., Ngo, T., Naser, M., 2019b. Test Database on Concrete-Filled networks. Cem. Concr. Res. 28, 1797–1808. https://doi.org/10.1016/S0008-
Steel Tubular Columns. https://doi.org/10.17632/3XKNB3SDB5.1. 8846(98)00165-3.
van Smeden, M., Moons, K.G., de Groot, J.A., Collins, G.S., et al., 2018. Sample size for Ying, Y., Garrett, J.H., Oppenheim, I.J., Soibelman, L., Harley, J.B., Shi, J., Jin, Y., 2013.
binary logistic prediction models: beyond events per variable criteria 28, 2455–2474. Toward data-driven structural health monitoring: application of machine learning
https://doi.org/10.1177/0962280218784726. and signal processing to damage detection. J. Comput. Civ. Eng. https://doi.org/
Vassallo, D., Krishnamurthy, R., Fernando, H.J.S., 2020. Utilizing physics-based input 10.1061/(asce)cp.1943-5487.0000258.
features within a machine learning model to predict wind speed forecasting error. Zarringol, M., Thai, H.T., 2022. Prediction of the load-shortening curve of CFST columns
Wind Energy Sci. https://doi.org/10.5194/wes-2020-61. using ANN-based models. J. Build. Eng. https://doi.org/10.1016/
Wong, T.T., Yang, N.Y., 2017. Dependency analysis of accuracy estimates in k-fold cross j.jobe.2022.104279.
validation. IEEE Trans. Knowl. Data Eng. https://doi.org/10.1109/ Zarringol, M., Thai, H.T., Naser, M.Z., 2021. Application of machine learning models for
TKDE.2017.2740926. designing CFCFST columns. J. Constr. Steel Res. https://doi.org/10.1016/
XGBoost Python Package, 2020. Python Package Introduction — Xgboost 1.4.0-SNAP- j.jcsr.2021.106856.
SHOT Documentation. https://xgboost.readthedocs.io/en/latest/python/pyth Zhou, Y., Yu, Y., Ding, B., 2020. Towards MLOps: a case study of ML pipeline platform. In:
on_intro.html#early-stopping. (Accessed 10 February 2021). Proc. 2020 Int. Conf. Artif. Intell. Comput. Eng. ICAICE 2020. https://doi.org/
10.1109/ICAICE51518.2020.00102.

15

You might also like