See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/385428445

A machine learning approach based on contract parameters for cost

forecasting in construction

Conference Paper · November 2024

CITATIONS READS

0 3

4 authors, including:

Maroua Ben Talha

National School of Mineral Industry
2 PUBLICATIONS 1 CITATION

SEE PROFILE

All content following this page was uploaded by Maroua Ben Talha on 01 November 2024.

The user has requested enhancement of the downloaded file.

 A machine learning approach based on contract
parameters for cost forecasting in construction

BEN TALHA Maroua1, HOUARI Fatima-Zahrae1, BENGHABRIT Asmaa2

and RHAZI Mohamed Sayf Eddine3
1 INDUSLAB, ONRST, ENSMR Rabat, Morocco
2 Laboratory of Applied Mathematics and Business Intelligence, ENSMR Rabat, Morocco
3 JACOBS Engineering S.A, Casablanca, Morocco

[email protected]

Abstract. The construction industry, characterized by its inherent complexity, is

a cornerstone of economic growth and development. In the era of Construction
4.0, accurate cost estimation within this sector is essential for successful project
planning and execution. This research aims to predict construction project costs
using real data from 203 projects, incorporating contract parameters provided by
an engineering consulting firm. Among the eleven machine learning algorithms
tested, the Partial Least Squares regression model emerged as the most effective.
It achieved the highest ranking in the TOPSIS analysis, demonstrating a coeffi-
cient of determination of 91% for the test set.

Keywords: Construction industry, Cost forecasting, Contract parameters, Ma-

chine learning, TOPSIS.

1 Introduction

Economic growth is heavily influenced by the construction industry, which involves a

diverse range of stakeholders such as contractors, clients, and consultants [1]. The in-
herent complexity and multifaceted nature of construction projects demand effective
collaboration and efficient management [2]. Success in this sector is often measured
through effective collaboration and efficient management [2], it is often measured
through the principles of the Iron Triangle [3], which emphasize the importance of stay-
ing within budget, meeting deadlines, and maintaining high-quality standards. The
complexity of construction projects makes variations in design or building processes
almost unavoidable. Variation orders, or change orders, are challenging elements that
require careful management to prevent disputes and claims between contractors and
owners, especially concerning contract costs and schedule delays [4]. These orders fre-
quently lead to higher costs and disruptions to ongoing work, ultimately resulting in
both cost and time overruns [5]. Therefore, developing robust strategies and approaches
to manage risks and uncertainties in construction projects is imperative [6], the con-
struction project management sector primarily consists of several key stages: the initial
design phase, detailed design phase, construction cost estimation, project bidding
2

phase, construction phase, and final delivery after completion [7]. Accurate cost esti-
mation is seen as essential for project success across all construction phases [8], it
significantly impacts the profitability of projects during the tender phase and is crucial
for the overall viability of the project [9]. This process involves the collaborative efforts
of a tender manager, who coordinates the cost estimation, and technical experts such as
engineers and project managers. Additionally, consultancy firms evaluate various pa-
rameters, including the type and scope of projects, to determine the necessary time and
effort for project preparation, i.e., consultancy services. This approach differs signifi-
cantly from cost estimation during construction, which primarily focuses on labor, ma-
terials, equipment, and other related expenses [10]. There are numerous estimation
methods and techniques available [11]. However, these conventional methods fail to
fully utilize the wealth of knowledge from past projects, leading to slow and incon-
sistent estimates. Advancements in computing power have led to the adoption of more
complex methods in recent cost estimation approaches. Artificial Intelligence (AI)
methods have been employed to explore multi-dimensional and non-linear relationships
between final costs and input variables [12]. In this context, Machine Learning (ML),
a subset of AI, has emerged as a promising tool for improving estimation accuracy.
Recent studies indicate that ML can be beneficial for predicting both duration and costs
[13]. Several researchers emphasize the importance of validating these machine learn-
ing techniques using real-world data to ensure their accuracy and reliability [14]. This
research focuses on developing a precise machine learning model that utilizes historical
data to estimate costs for engineering consultancy firm. To determine the most effective
model, we compared various ML regression techniques. Additionally, we used a
TOPSIS analysis (Technique for Order of Preference by Similarity to Ideal Solution)
to rank the algorithms based on four criteria: coefficient of determination (R-squared),
Root Mean Squared Error (RMSE), memory usage, and computational time. Out of the
eleven cost forecasting models evaluated, the PLS (Partial Least Squares regression)
model stood out for its robust learning capabilities and superior accuracy in cost esti-
mation, leading to its final deployment. In exploring these techniques, we address the
question of how ML models can be effectively leveraged within the context of Con-
struction to accurately predict costs using contract parameters in the early stages of
Project Management? The remainder of this paper is structured as follows: we begin
with a review of relevant literature. Next, we discuss the proposed methodology in de-
tail including data acquisition, description, pre-processing and implementation. This is
followed by analyzing the results of the cost estimates, and discuss the main findings.
We conclude our study by summarizing the main contributions, limitations, and future
research.

2 Literature Review

Previous research has explored various cost estimation techniques, dividing them into
qualitative and quantitative methods. Qualitative methods rely on the estimator’s
 3

knowledge of the project [15], and include expert judgment and heuristic rules. Expert
judgment is shaped by the success of prior estimations and often involves consulting
experienced professionals to enhance accuracy [16]. Heuristic rules are intuitive, using
rules of thumb based on comparisons with similar past projects to streamline the esti-
mation process. Quantitative methods rely on collecting and analyzing historical data,
and using quantitative models, techniques, and tools to estimate project costs. Among
these approaches, parametric and simulation methods are prominent. Parametric meth-
ods estimate project costs by identifying causal relationships with specific parameters,
resulting in a mathematical function of the related variables. This method is particularly
efficient during the early stages of a project when limited information is available [17].
Stochastic Budget Simulation (SBS) enhances cost estimation by combining Monte
Carlo simulation with the successive principle. It uses triple estimates (minimum, most
likely, maximum) to address uncertainties, improving reliability. This method produces
a probabilistic cost range, making it ideal for early-stage, large-scale projects [18].
Advanced methods utilizing AI and ML have also been employed for cost estima-
tion. Among these, Artificial Neural Networks (ANNs) are frequently used for estimat-
ing the duration and costs of construction projects during the preliminary design stage
[19]. The effectiveness of ANN and SVM (Support Vector Machine) models in esti-
mating the cost and duration of road projects was evaluated [20], with SVM models
outperforming ANN models in terms of accuracy and error reduction. Another study
highlighted the effectiveness of SVM in estimating the cost of bridge construction dur-
ing the initial stages of project development [21]. In [22], a comparison of twenty AI
techniques for estimating construction costs in field canal improvement projects re-
vealed that the XGBoost (eXtreme Gradient Boosting) algorithm provided the most
accurate predictions, achieving an R-squared value of 0.929. These models can signif-
icantly aid decision-makers in substation projects. The study in [23] applied three ma-
chine learning algorithms—MLP (Multi-Layer Perceptron), GRNN (General Regres-
sion Neural Network), and RBFNN (Radial Basis Function Neural Network) along with
a process-based method for early-stage cost prediction in project management. It con-
firmed that the GRNN algorithm produced superior results compared to other models,
assisting project managers in predicting construction costs during the contracting phase.
Our work builds on this foundation to forecast construction project costs by analyzing
contract parameters and using a quantitative approach based on machine learning tech-
niques. This method helps predict potential cost overruns early, offering insights for
more effective project management.
We overview the eleven regression algorithms used in our study, divided into linear
models (Ridge [24], Lasso [24], Elastic Net [25], PLS Regression [26]) and non-linear
models; Random Forest [27], ETR (Extra Tree Regression) [28], AdaBoost (Adaptive
Boosting) [29], CatBoost [30], LightGBM (Light Gradient Boosting Machine) [31],
Gradient Boosting [32], XGBoost [33]. This categorization stems from our literature
review, which shows these groups are commonly used in cost estimation for their spe-
cific benefits and applications. Linear methods are foundational in regression analysis
4

due to their simplicity and interpretability. These models assume a linear relationship
between input features and the target variable, making them effective for problems
where this assumption holds true. Ridge Regression adds an L2 penalty to the regres-
sion model, stabilizing it by shrinking the coefficients of correlated predictors. Lasso
Regression introduces an L1 penalty, which can shrink some coefficients to zero, ef-
fectively performing variable selection and simplifying the model [24]. Elastic Net Re-
gression, a hybrid of Lasso andRidge, combines their penalties, making it particularly
effective in the presence of highmulticollinearity among predictors [25]. PLS Regres-
sion reduces predictors to a smaller set of uncorrelated components while preserving as
muchvariance in the response variable as possible, making it useful in situations with
many correlated predictors [26]. The selected linear models offer substantial flexibility
in hyperparameter tuning. This flexibility is critical, enabling precise adjustments to the
model parameters, such as regularization strength in Ridge and Lasso, or the mixing
ratio in Elastic Net. These adjustments allow the models to better fit the specific char-
acteristics of the dataset, thereby enhancing both predictive accuracy and model gener-
alizability.
Non-linear methods are chosen for their ability to model complex, non-linear rela-
tionships between variables, capabilities that linear models lack. These methods do not
assume a linear relationship and can capture intricate patterns in the data. Random For-
est, introduced by Leo Breiman, constructs multiple decision trees using subsets of ran-
domly selected variables, improving predictive accuracy and reducing overfitting by av-
eraging these trees’ results [27]. ETR, proposed by Geurts et al., builds unpruned deci-
sion or regression trees,introducing more randomness in tree construction, which en-
hances its performance on complex datasets [28]. AdaBoost introduced by Freund and
Schapire, is an ensemble method that combines multiple weak learners, adjusting the
weights of incorrectly predicted instances to focus more on challenging cases [29]. Cat-
Boost, developed by Yandex, is a gradient boosting algorithm designed to handle cate-
gorical features efficiently without extensive preprocessing, significantly enhancing
model performance on categorical data [30]. LightGBM is a highly efficient gradient
boosting framework that uses a histogram-based approach to speed up training and re-
duce memory usage [31]. Gradient Boosting builds models sequentially, each correct-
ing the errors of its predecessor, thus improving predictive accuracy incrementally [32].
XGBoost, a scalable and high-performance implementation of gradient boosting, is de-
signed for speed and performance, making it highly effective for large datasets [33].

3 Proposed Methodology

3.1 Workflow

This research utilizes a systematic methodology to develop and evaluate machine learn-
ing models, as illustrated in Figure 1. The process begins with data acquisition, fol-
lowed by data preprocessing to organize and clean the data for effective modeling. Key
 5

predictors are then identified during the feature selection phase. Machine learning al-
gorithms are applied within a Jupyter Notebook, incorporating hyperparameter tuning
and testing, with Cross Validation to ensure robustness. The TOPSIS method ranks the
algorithms, selecting the highest-performing one based on predefined criteria. The final
step involves deploying the top model for practical use.

Fig. 1. Research Workflow.

3.2 Data acquisition and description

The dataset utilized in this research was furnished by the consulting firm JESA SA and
includes data on 203 real-world projects from sectors such as building, transportation,
and ports, covering a temporary margin from the year 2013, offering a longitudinal
perspective that enhances the analysis of trends and patterns over time. This dataset is
structured into 11 columns, divided into numerical and categorical variables:
• Numerical Variables: Contract Value (Revenue); indicates the financial mag-
nitude of the projects in MAD. Final Cost: represents the ultimate expenditure
incurred onthe projects in MAD.
• Categorical Variables: Project Number: a unique identifier for each project,
ProjectDescription: detailed narrative of the project scope, Project Size: cate-
gorized into Small, Medium, and Large, this variable assesses the scale of the
project, Business Area: encompasses 14 distinct sectors within the business,
Service Type: includes 12 different services offered, detailing the variety of
services engaged, Legal Entity:identifies one of four legal entities involved and
Geography: divided into Africa andNational (Morocco), indicating the geo-
graphical location of the project execution.

3.3 Data preprocessing

Data preprocessing is a critical phase in our research, essential for ensuring data quality
and integrity before applying machine learning models. The initial step in data cleaning
involved addressing missing values and managing outliers, both of which can signifi-
cantly impact the accuracy of machine learning models. For categorical variables such
as Business Area and Service Type, missing entries were imputed with the mode of the
6

respective variable, ensuring that the most frequently occurring category was used to
fill the gaps. Outliers, which can distort predictive modeling and lead to mis- leading
results, were addressed using the Interquartile Range (IQR) method. This method of
capping helps in reducing the skewness of the distribution and maintains the integrity
of the data set.

3.4 Feature selection

In the feature selection phase of our study, we have considered all available variables
within the dataset to forecast project costs accurately. Our aim is to establish a statisti-
cally significant relationship between these variables and the output cost, thereby vali-
dating their inclusion in the predictive model. This process begins with a comprehen-
sive evaluation of the independence of categorical variables through the Chi-square test
[34].

Table 1. Chi-square Test Results for Independence.

Variable 1 Variable 2 P-value

Contract Type Size 0.09990
Contract Type Geography 0.37234

Size Geography 0.89744

Table 1 shows Chi-square test results for independence between Contract Type and
both Size and Geography, and between Size and Geography. All pairs listed have p-
values above 0.05, indicating no significant associations and thus they are independent.
Other variable pairs not shown in the table have significant p-values below 0.05, indi-
cating strong dependencies between these variables. Following the assessment of cate-
gorical variables, we employ the Kruskal-Wallis test, that tests the null hypothesis that
the medians of all groups are equal. This allows us to determine if there are statistically
significant differences between the groups' medians [35].

Table 2. Kruskal-Wallis results.

Variable P-value
Size 0.0000001
Contract Type 0.0009505
Business Area 0.0000008
Service Type 0.0000431
Legal Entity 0.0090484
Geography 0.0175414
 7

The results from the Kruskal-Wallis tests, as summarized in Table 2, confirm the sta-
tistical significance of several variables influencing the output cost. The findings con-
firm that these factors significantly impact the dependent variable, thus justifying their
selection based on the rigorous criteria set by the Kruskal-Wallis test.

3.5 Algorithm implementation

In this section, we will detail the implementation of machine learning algorithms as

guided by the previously discussed flowchart Figure 1. Once the data is prepared, we
employ a grid search strategy to fine-tune the hyperparameters for each model, lever-
aging the GridSearchCV function. For instance, in the Ridge regression model, differ-
ent alpha values ranging from 0.01 to 100.0 were tested to identify the optimal setting.
Similarly, for models like Elastic Net, both alpha and l1_ratio parameters were explored
to balance the penalty terms, while the Random Forest model’s hyperparameters and
boosting models, including n_estimators, max_depth, and min_samples_split, were
tuned to enhance model performance. The machine learning models were implemented
and trained in a Jupyter Notebook environment (version 7.0.8) using Python 3.9, chosen
for its support of scientific computing libraries like Scikit-learn and Pandas. Following
the training phase, each model underwent rigorous evaluation using the designated test
set. Key performance metrics, including RMSE and R², were calculated to assess both
accuracy and generalizability, ensuring the models' applicability to real-world data.

4 Experimental results

4.1 Performance of ML techniques

In the results section of our study, we evaluated the performance of various machine
learning models, the evaluation focused on two key metrics: R² and RMSE. The varia-
tion in R² and RMSE scores across different models can be attributed to the distinct
algorithms they employ and their specific approaches to handling data complexities.
For example, Ridge Regression effectively manages multicollinearity, while PLS Re-
gression simplifies the dataset by reducing predictors to a smaller set of uncorrelated
components. These models have demonstrated strong performance, achieving an R² of
91% for PLS Regression and 90% for Ridge Regression, showcasing their aptness for
such data structures. To assess the robustness of our models, we employed c
Cross-Validation, analyzing the variability and consistency of model performance
through R² scores from various data splits. Figures 2 and 3 display the results, identify-
ing models with consistent performance and those prone to variations across different
splits. Specifically, models like Lasso and PLS Regression showed narrow distributions
in their R-squared scores, indicating stable and reliable performance. Conversely, mod-
els such as Gradient Boosting displayed wider score distributions, suggesting they may
be sensitive to specific data characteristics.
8

Fig. 2. Cross-validated R-squared Scores for train set.

Fig. 3. Cross-validated R-squared Scores for test set.

4.2 TOPSIS analysis

To rank the algorithms, we used the Technique for Order Preference by Similarity to
Ideal Solution (TOPSIS), developed by Hwang & Yoon [36]. This method evaluates
alternatives based on their proximity to an ideal solution, favoring those closest to the
positive ideal and farthest from the negative ideal. We considered four criteria: R²,
RMSE, Usage Memory, and Computational Time.

Table 3. Score Matrix for TOPSIS.

ML Technique R² RMSE Usage Memory Computational time

(MiB) (Seconds)
PLS Regression 0.9118800 309182.6 0.08593 0.3181
Ridge 0.8983940 332007.7 0.28515 0.2207
Elastic Net 0.8929560 340775.6 0.02343 0.3939
Lasso 0.8983090 332145.9 0.00000 0.2312
 9

Extra Trees Regressor 0.7515880 519128.4 0.43359 1.0290

Random Forest 0.8880313 348527.5 0.07812 16.283
Gradient Boosting 0.8403190 416212.8 0.09760 22.492
XGBoost 0.8426780 413127.0 0.52734 16.687
LightGBM 0.8021900 463245.4 0.24218 6.7239
CatBoost 0.6913820 578627.2 18.4179 44.553
AdaBoost 0.6616910 605822.0 0.00000 5.2532

In our evaluation methodology, we focused on optimizing the R-squared value to gauge

model accuracy and assigned lesser weights to RMSE, memory usage, and com- puta-
tional time (0.4 for R-squared and RMSE, 0.1 for memory and time). Sensitivity anal-
ysis validated this approach, consistently highlighting the PLS Regressor as the best
model. This confirms its effectiveness in various scenarios, leading to its imple- men-
tation in a Streamlit interface for dynamic user interaction.

Fig. 4. TOPSIS ranking of models.

5 Deployment

This section introduces the Cost Prediction Tool, a user-friendly interface built with
Streamlit that utilizes the PLS Regression model. Users can input key contract param-
eters like type, size, and geography to obtain precise cost estimates, demonstrating the
practical application of our research findings.
10

Fig. 5. Interface for cost predicting using contract parameters.

The following table provides a detailed specification sheet for the Cost As Sold Predic-
tion Tool, summarizing its key features.

Table 4. Specification sheet of cost prediction tool.

Specification Details
Web-based tool for predicting construction project costs us-
Overview
ing contract parameters.
Machine Learning Model: PLS Regression.
User Inputs: Contract Type, Size, Geography, Legal En-
Key features
tity, Business Area, Service Type, Contract Value.
Prediction Output: Estimated Cost in MAD.
Supported Platforms Web-Based: Chrome, Firefox, Edge.

6 Conclusion and discussion

Accurate construction cost estimation is crucial for ensuring the financial success and
operational efficiency of projects. Our research leveraged the PLS regression model,
which proved to be highly effective in forecasting costs based on contract parameters.
While the results are promising, it is important to acknowledge certain areas where
further refinement and exploration could enhance the robustness of our findings.
One consideration is the scope of the dataset, which could benefit from further ex-
pansion to include a broader range of project types and conditions. This would not only
enhance the model's generalizability but also its adaptability to different contexts within
the construction industry. Additionally, although the PLS regression model performed
exceptionally well, exploring more advanced modeling techniques, such as Artificial
 11

Neural Networks (ANNs), could further capture the complex, nonlinear relationships
that may exist within larger and more varied datasets.

References
1. Ingle, P. V., & Mahesh, G.: Construction project performance areas for Indian construction
projects. International Journal of Construction Management 22(8), 1443–1454 (2020).
2. Jing, W., Naji, H. I., Zehawi, R. N., Ali, Z. H., Al-Ansari, N., & Yaseen, Z. M.: System
dynamics modeling strategy for civil construction projects: the concept of successive legis-
lation periods. Symmetry 11(5), 677 (2019).
3. Pollack, J., Helm, J., & Adler, D.: What is the Iron Triangle, and how has it changed? Inter-
national Journal of Managing Projects in Business 11(2), 527– 547 (2018).
4. Charoenngam, C., Coquinco, S. T., & Hadikusumo, B. H. W.: Web-based ap- plication for
managing change orders in construction projects. Construction In- novation 3(4), 197-215
(2003).
5. Chan, A. P., & Y. C. M.: A comparison of strategies for reducing variations. Construction
Management and Economics 13(6), 467-473 (1995).
6. Jaafari, A.: Management of risks, uncertainties and opportunities on projects: time for a fun-
damental shift. International Journal of Project Management 19(2), 89–101 (2001).
7. Kim, S., Chang, S., & Castro-Lacouture, D.: Dynamic modeling for analyzing impacts of
skilled labor shortage on construction project management. Journal of Management in En-
gineering 36(1) (2020).
8. ElSawy, I., Hosny, H., & Razek, M. A.: A neural network model for construc- tion projects
site overhead cost estimating in Egypt. arXiv Prepr. (2011).
9. Ahn, J., Ji, S. H., Ahn, S. J., et al.: Performance evaluation of normalization- based CBR
models for improving construction cost estimation. Automation in Construction 119, Article
ID 103329 (2020).
10. Zwaving, J.: Probabilistic estimating of engineering costs. Delft University of Technology,
Delft, The Netherlands (2014).
11. Burke, R.: Project management - planning and control techniques. 4th edn. Wiley, India,
251–266 (2009).
12. Günaydin, H. M., & Doğan, S. Z.: A neural network approach for early cost estimation of
structural systems of buildings. International Journal of Project Management 22(7), 595–
602 (2004).
13. Pellerin, R., & Perrier, N.: A Review of Methods, Techniques and Tools for Project Planning
and Control. International Journal of Production Research 57(7), 2160–2178 (2018).
14. Aramali, V., Sanboskani, H., Gibson, G. E., Jr., El Asmar, M., & Cho, N.: For- ward-Look-
ing State-of-the-Art Review on Earned Value Management Systems: The Disconnect Be-
tween Academia and Industry. Journal of Management in Engineering 38(3) (2022).
15. Hashemi, S. T., Ebadati, O. M., & Kaur, H.: Cost estimation and prediction in construction
projects: a systematic review on machine learning techniques. SN Applied Sciences 2, Arti-
cle 1703 (2020).
16. PF: Project estimating and cost management. Management Concepts Incorpo- rated, Vir-
ginia (2002).
17. Hegazy, T., & Ayed, A.: Neural network model for parametric cost estimation of highway
projects. Journal of Construction Engineering and Management 124(3), 210-218 (1998).
 12

18. Elkjaer, M.: Stochastic budget simulation. International Journal of Project Man- agement
18(2), 139–147 (2000).
19. Lowe, D. J., Emsley, M. W., & Harding, A.: Predicting construction cost using multiple
regression techniques. Journal of Construction Engineering and Man- agement 132(7), 750–
758 (2006).
20. Peško, I., Mucenski, V., Seslija, M., et al.: Estimation of costs and durations of construction
of urban roads using ANN and SVM. Complexity 13 (2017).
21. Juszczyk, M.: On the search of models for early cost estimates of bridges: an SVM-based
approach. Buildings 10(1), 2 (2019).
22. Elmousalami, H. H.: Artificial intelligence and parametric construction cost es-timate mod-
eling: state-of-the-art review. Journal of Construction Engineering and Management 146(1)
(2020).
23. Car-Pusic, D., Petruseva, S., Zileska Pancovska, V., & Zafirovski, Z.: Neural network-based
model for predicting preliminary construction cost as part of cost predicting system. Ad-
vances in Civil Engineering (2020).
24. Melkumova, L. E., & Shatskikh, S. Ya.: Comparing Ridge and LASSO estimators for data
analysis. Procedia Engineering 201, 746–755 (2017).
25. Zou, H., & Hastie, T.: Regularization and variable selection via the elastic net. Journal of
the Royal Statistical Society: Series B (Statistical Methodology) 67(2), 301-320 (2005).
26. Chun, H., & Keles, S.: Sparse partial least squares regression for simultaneous dimension
reduction and variable selection. Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 72(1), 3–25 (2010).
27. Breiman, L.: Bagging Predictors. Machine learning 24(2), 123-140 (1996).
28. Lu, S., Wang, X., Zhang, G., & Zhou, X.: Effective algorithms of the Moore Penrose inverse
matrices for extreme learning machine. Intelligent Data Analysis 19(4), 743–760 (2015).
29. Shrestha, D. L., & Solomatine, D. P.: AdaBoost.RT: A boosting algorithm for regression
problems. Proceedings of the IEEE International Joint Conference on Neural Networks,
1163–1168 (2004).
30. Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., & Gulin, A.: CatBoost: unbi-
ased boosting with categorical features. arXiv preprint arXiv:1706.09516v5 (2019).
https://arxiv.org/abs/1706.09516
31. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu, T.- Y.: LightGBM:
A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Pro-
cessing Systems 30, 3149–3157 (2017).
32. Friedman, J. H.: Greedy function approximation: a gradient boosting machine. Annals of
Statistics, 1189–1232 (2001).
33. Chen, T., & Guestrin, C.: XGBoost: A Scalable Tree Boosting System. Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
785–794 (2016).
34. Chicco, D., Warrens, M. J., & Jurman, G.: The coefficient of determination R-squared is
more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis eval-
uation. PeerJ Computer Science (2021).
35. Rana, R., & Singhal, R.: Chi-square test and its application in hypothesis testing. Journal of
the Practice of Cardiovascular Sciences 1(1), 69-71 (2015).
36. Hwang, C. L., & Yoon, K. P.: Multiple attributes decision making methods and applications.
Springer-Verlag, Berlin (1981).

View publication stats