Determinants of Yearly CO₂ Emission Fluctuations: A Machine Learning Perspective to Unveil Dynamics

Abstract

In order to understand the dynamics in climate change, inform policy decisions and prompt timely action to mitigate its impact, this study provides a comprehensive analysis of the short-term trend of the year-on-year CO₂ emission changes across ten countries, considering a broad range of factors including socioeconomic factors, CO₂-related industry, and education. This study uniquely goes beyond the common country-based analysis, offering a broader understanding of the interconnected impact of CO₂ emissions across countries. Our preliminary regression analysis, using the ten most significant features, could only explain 66% of the variations in the target. To capture the emissions trend variation, we categorized countries by the change in CO₂ emission volatility (high, moderate, low with upward or downward trends), assessed using standard deviation. We employed machine learning techniques, including feature importance analysis, Partial Dependence Plots (PDPs), sensitivity analysis, and Pearson and Canonical correlation analyses, to identify influential factors driving these short-term changes. The Decision Tree Classifier was the most accurate model, with an accuracy of 96%. It revealed population size, CO₂ emissions from coal, the three-year average change in CO₂ emissions, GDP, CO₂ emissions from oil, education level (incomplete primary), and contribution to temperature rise as the most significant predictors, in order of importance. Furthermore, this study estimates the likelihood of a country transitioning to a higher emission category. Our findings provide valuable insights into the temporal dynamics of factors influencing CO₂ emissions changes, contributing to the global efforts to address climate change.

1. Introduction

Greenhouse gases (GHGs) are one of the main reasons behind natural disaster risks. Combined with socioeconomic conditions, governance, and conflict, these complex and dynamic phenomena are causing huge damage [1]. Extensive research has been conducted on natural disaster risks, related topics, and their possible implications, resulting in the establishment of several indicators suitable for explaining and quantifying their significance and possible impact [2,3,4]. While research has illuminated various aspects of these disaster risks, contributing thus to the progress achieved in this field, the current frameworks designed to reduce their impacts are often designed for long-term durations [3], which represents a major constraint, considering the capricious nature of these hazards and their escalating repercussions on human lives. Furthermore, even when those frameworks are implemented, the peril persists, thereby increasing the vulnerability of the nations categorized as least developed [5], limiting those nations in ameliorating their positions. As a fact, considering the WRI and its subcomponents, it is more likely for a country, either developed or not, to remain in its position of vulnerability and susceptibility within five consecutive years. Also, the least developed countries have a one percent probability of improving their position, but only after 5 years [6]. Among GHGs, carbon dioxide (CO₂) receives particular attention due to its high rate of production as a result of human activities and its negative environmental impacts, such as air pollution, temperature rise, etc.; this situation is alarming, since the least developed countries which pollute and emit less are more exposed to the disasters induced by the production of CO₂ compared to developed countries, which pollute and emit the most [7]. Thanks to technological advances, several studies have provided accurate forecasting and projections of CO₂ emissions_, as well as deep insight into the interplay of other components, such as the political, geographical, economic, environmental, and societal components in their production, thus enhancing our understanding on the subject, which supports current frameworks such as the Paris Agreement and other decarbonization pathways. Understanding the factors contributing to CO₂ emissions, whether they are direct (like burning fossil fuels) or indirect (like deforestation), can be relatively straightforward; however, explaining the changes in CO₂ emissions over a period is a more complex task because such a process involves not only understanding the factors contributing to emissions, but also understanding their dynamics over time. Such a process requires a deep understanding of a wide range of fields, including technological, economic, and policy changes, as well as changes in energy use, land use, and population growth. Moreover, the relationship between these factors and CO₂ emissions can be non-linear and involve complex feedback loops. For example, economic growth might lead to increased energy use and CO₂ emissions, but it could also drive technological innovation that reduces emissions [8,9]. This subject is even more complex considering the possible implication of decarbonization on the economy of nations which, in the majority, are sustained by high CO₂ emitters such as coal, oil, and cement. The limited or absent clear responses to explain the change in CO₂ emissions over time and on a global scale, coupled with the urgent need to provide a more inclusive response to address this threat, have prompted this research, which aims to answer these questions:

• Over time, how have economic factors, CO₂-related industries, educational levels and population dynamics interacted to influence the short-term change in the trend of CO₂ emissions across diverse countries with diverse characteristics, with respect to the factors mentioned?
• What insight in terms of the identification and quantification of the temporal dynamics and the influence of these factors can machine learning techniques highlight to deepen one’s understanding of this change over time on a global scale?

This study pioneers a holistic approach to understanding global CO₂ emissions, addressing a critical gap in existing research. Traditional studies often limit their focus to individual countries and a narrow set of factors, which impedes a comprehensive global understanding. Our research, however, broadens this perspective by analyzing an extensive set of factors across a diverse range of countries. This innovative approach illuminates the complex dynamics driving the changes in CO₂ emissions, providing crucial insights for formulating effective global responses to this pressing environmental issue. This research stands out for its use of a unique dataset that amalgamates data from 10 different countries, each with distinct characteristics. This allows us to acknowledge both the direct and indirect factors that have been identified by experts and researchers as potential influencers of changes in CO₂ emissions. On top of this, what sets our study apart is the application of advanced machine learning and statistical techniques, to understand the temporal dependency and the contribution of these factors regarding changes in CO₂ emissions. This approach enables us to provide a clear, quantifiable understanding of how these factors interplay and contribute to changes in CO₂ emissions on a global scale, replacing the blurred comprehension that currently exists. Following these steps, this study contributes to the field by demonstrating the generalizability of these results, given the diverse backgrounds of the selected countries. This not only enhances the applicability of our findings but also paves the way for future research in this area. To achieve this task, the remainder of the paper is organized as follows: after the Section 1 dedicated to the introduction, the Section 2 will discuss the materials and methods considered, followed by the Results and Discussion sections. Finally, the last section is for the conclusions.

2. Materials and Methods

2.1. Data Collection

A dataset comprising twenty-six distinct features, spanning the period from 1960 to 2022, was meticulously compiled from two esteemed sources: ourworldindata.org and the World Bank Open Data. The objective of this analysis was to yield a more generalizable outcome; hence, the selection of countries was based on a diverse range of factors including economic level, population dynamics, geographical location, education level, and their respective contributions to CO₂ emissions. The countries under consideration were the United States, United Kingdom, South Korea, China, India, France, Brazil, Democratic Republic of the Congo, Nigeria, and South Africa (refer to Appendix A for further details). The dataset, organized on an annual basis, encompasses five categories of features: population dynamics (2 features), economic factors (3 features), education (7 features), CO₂-related industry (5 features), and CO₂-related emissions and temperature (7 features). To handle the missing values in the dataset, the iterative imputation technique from the Fancy Impute library in Python was employed (for more information, refer to Appendix B). This approach ensures the integrity and completeness of the data, thereby enhancing the reliability of the subsequent analysis.

2.2. Data Preparation

Python 3.10 on the last distribution of anaconda was used. The data were prepared for the supervised machine learning techniques (both regression and classification). Considering the varying trend for each country for the absolute change in CO₂ emissions (Appendix D), which is the target variable, it was necessary to mitigate this considering its potential negative impact on the performance of the algorithms. To capture the short-term trend of the target, countries were grouped based on the percentile of volatility of the mean value for 3 years of the target. This data-driven approach instills generalizability in the findings and helped overcome the limitation of the regression technique which suffers from the extreme variability of the target.

Figure 1 explains the grouping process.

In order to capture the unique characteristics and conditions of each country, preserve the temporal order, avoid data leakage, and provide a realistic evaluation and understanding of the target’s temporal dynamics, we first grouped the data by country. Following this, we defined a rolling period of 3 years, to obtain the mean value for each country over this period, which served as the new target. By the same process of grouping data by country, the standard deviation was determined for each of them. The percentiles (25% as moderate low threshold and 75% as moderate high threshold) were used as thresholds to define the low, moderate and high volatility in the 3 years of mean values for each country. This stage helped to categorize each countries’ variation using data-driven approach that can be adapted to other datasets. Considering that the standard deviation is always positive but the variations sometimes take negative values, it was imperative to specify the direction of the variation as either negative or positive. Thus, after verification of the sign of the new target, that sign was picked and assigned to the level of volatility defined by the threshold. This process resulted in six classes of volatility: High positive and negative, Moderate positive and negative, and Low positive and negative. By doing so, not only is the volatility defined, but also the direction, making it more comprehensive for the interpretation of the prediction. The outcome of this process is presented in the Section 3.

2.3. Machine Learning Algorithms

For the regressions approach, the performance of the following regressors was compared: Linear Regression [10], Ridge Regression [11], Bagging Regressor [12], Random Forest Regressor [13], Gradient Boosting Regressor [14], XGBoost Regressor [15], AdaBoost Regressor [16] and KNeighbors Regressor [17]. Concerning classification, the performance of the following classifiers was compared: Logistic Regression (LogReg) [18], Decision Tree (DT) [19], Random Forest Classifier (RF) [20], XGBoost classifier (XGB) [21], Multi-Layer Perceptron classifier (MLP) [13], Bagging (BC) [22], AdaBoost (ABC) [23], Gradient Boosting (GB) [24], Support Vector (SVC) [25], and Gaussian Naïve Bayes (GNB) [26].

2.4. Metrics

Two rounds of evaluation were considered in the two approaches: the first consisting of the selection of the best performing model and the second of the final evaluation of the best performing model. To achieve this, the following metrics were considered:

o

For the selection of the best performing algorithm:

▪

Regression: Cross validation score [27], Mean squared error [28], Residuals [29], R-squared [30]

▪

Classification: Cross validation score [27], Accuracy score, Confusion matrix [31] and classification report [32].

o

For final evaluation:

▪

Regression: Mean squared error, Residuals, R-squared

▪

Classification: Accuracy score, Matthew correlation coefficient [33], Confusion matrix and classification report.

2.5. Explainable Machine Learning Techniques

To instill confidence in the prediction, the following explainable techniques were considered.

2.5.1. Partial Dependence Plots (PDPs)

It provides plots showing the marginal effect that features have on the predicted outcome of a machine learning model. A PDP can show whether the relationship between the feature and target is complex, monotonic or linear. It is an important technique since it has a causal interpretation, which means that is explains the outcome of a prediction [34,35].

It is defined as:

$$f_s\left(x_s\right)=\frac{1}{n}{\sum }_{i=1}^{n}f(x_s,x_c^{\left(i\right)})$$

(1)

where

x_s are the features for which the PDP is to be plotted,

x_c are the other features used in the machine learning model f,

$x_c^{\left(i\right)}$ are the actual features in the model which we are not interested, and

n is the number of instances in the dataset.

This analysis was achieved using the Partial Dependence Display package from the sklearn library. The outcome of this analysis is presented in the Section 3.

2.5.2. Sensitivity Analysis

Sensitivity analysis is a useful technique for understanding the impact of changes in the input features on the model outcome. By doing so, it provides insight into the most important features by quantifying the uncertainty in the model’s output [36]. It is often used to measure the correlation between changes in an input variable and the resulting changes in the output variable, and aims to study how the uncertainty in the output can be allocated to different sources of uncertainty in the inputs [37]. This process can be represented as follows:

Considering a model as a function $g:R^N\to R^M$, where N is the number of input variables and M the number of output variables. The input variables are represented as a vector $x=[x_1,x_2\dots x_N]$, and the output variables are represented as a vector $y=[y_1,y_2\dots y_M]$. The model maps the input variables to the output variables, for instance, y = g(x).

For a given input variable $x_i$, the sensitivity $S_i$ of the output variable $y_j$ with respect to $x_i$ can be calculated as follows:

$$S_i=\frac{\partial y_j}{{\partial x}_i}\dot{}\frac{x_i}{y_j}$$

(2)

Formula (2) represents the relative change in $y_j$ for a relative change in $x_i$.

This analysis was achieved using saltelli from the SALib package for sample generation, following the defined problem, and sobol from the same SALib, to get the first and total order sensitivity indices. The result of this process is provided in the Section 3.

2.5.3. Feature Analysis

A correlation analysis, especially the Pearson Correlation between features and the Canonical Correlation [38] among groups of features, was considered to evaluate the interplay of the features over time, to complete the one achieved using PDPs and sensitivity analysis.

2.6. Research Design

The process followed in this research is represented in Figure 2.

Once the data are collected from the different sources and for the considered country, they are merged to make a unique dataset. Missing values are imputed using the iterative imputation, then grouped by country before applying a temporal splitting of the train from 31 December 1960 to 31 December 2009 and the test from 31 December 2010 to 31 December 2022. Features were scaled using the Standard Scaler from the sklearn library. The Pearson and Canonical correlation analysis took place for the analysis of the interplay of features. The first approach of the modeling consisted in the regression technique to predict the average value of 3 years of the absolute change in CO₂ emissions, a variable which could capture the short-term trend of the target. Iteratively, the baseline modeling and grouping by volatility was considered for improvement, since the other did not improve it. The poor performance resulting from this approach led us to consider the classification technique. To achieve this, the process explained in Figure 1 was applied to the new target, and, to improve the performance of the classifiers, class balancing techniques such as Synthetic Minority Over -sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN) and class weight were considered. After this stage, the XAI is achieved and the result, including the one resulting from the correlation analysis, was analyzed and interpreted.

3. Results

3.1. Regression Analysis

In the process of selecting the best regressor, it appeared that the Gradient Boosting Regressor algorithm provided the best score (Figure 3, Appendix C.1 and Appendix C.2). To improve its performance, a reduction of dimensionality was applied using the PCA and feature selection techniques.

This general poor performance led us to consider the option of grouping the target based on each country’s volatility, to mitigate that difference, which tends to reduce the effectiveness of the model to capture hidden patterns during the training. With the Gradient Boosting Regressor as the best model, PCA (using n_components of 0.98) and a selection of the best 10 features using the Recursive Feature Elimination (RFE) [39] with the best model were applied. These best 10 features are: CO₂ from cement, CO₂ from gas, CO₂ from coal, share of cumulative CO₂ emissions, population—education: incomplete primary, population (number), GDP per capita, change in GDP, annual CO₂ emissions growth (%) and absolute CO₂ change. The summary of the performance of the best model based on the above mentioned techniques is presented in Figure 4.

This approach provided two groups: high volatility (China and United States) and low volatility (the remaining countries). When applying the same process of model selection, there is an observed improvement for countries in the low category, considering the result of the AdaBoost Regressor (70% r-squared), despite its inefficiency in generalizing the test set (cross validation score: −0.89).

Overall, using regression appeared not to be suitable, since it is not able to explain the target. This limitation justifies the need to find alternatives, among which we have the transformation of the target in classes following the process presented in Figure 2.

3.2. Classification Analysis

Grouping Target in Classes

The process explained in Figure 2 could provide six imbalanced classes, presented in

Table 1. Range of values by classes.

Category Names	Min Values	Max Values	Range	Countries
High positive	3.244667 × 103	6.769703 × 108	676,967,055.333	United States, China, India
High negative	−1.906534 × 108	−1.370680 × 105	190,516,332
Moderate positive	4.88107 × 105	3.741389 × 107	36,925,079.3	United Kingdom, France, South Korea, Brazil, India
Moderate negative	−3.209079 × 107	−4.587947 × 105	31,631,995.3
Low positive	2.683000 × 103	2.632321 × 107	26,320,527	South Africa, Nigeria, Democratic Republic of the Congo
Low negative	−2.045842 × 107	−8.609333 × 103	20,449,810.667

. A better understanding of this classification is provided in Figure 5, which depicts the temporal dynamics of the mean value for 3 years for each country.

The High (positive and negative) category represents the group of high polluting countries despite efforts to reduce the level of CO₂ emissions over time. The Moderate (positive and negative) category is the group of emitters whose level of CO₂ emissions is important but still tolerable compared to the previous group, And the Low category (positive and negative) represents those countries whose emission is quite good compared to the others. It appears, over time, that countries belonging to a given category remained in it but fluctuated in both positive and negative directions (

Table 2. Summary of classifiers’ performance.

Model	MCC	AUC	Precision	Recall	F1 Score	Mean CV
Logreg	0.67	0.73	0.76	0.73	0.68	0.74
DT	0.95	0.96	0.96	0.96	0.96	0.93
RF	0.87	0.89	0.91	0.89	0.89	0.86
XGB	0.83	0.85	0.89	0.85	0.84	0.91
GB	0.84	0.86	0.89	0.86	0.85	0.85
SVC	0.58	0.64	0.70	0.64	0.58	0.73
MLP	0.60	0.66	0.74	0.66	0.64	0.75
GNB	0.50	0.58	0.66	0.58	0.55	0.75

), except India, which in the early 2000s moved from the moderate to high volatility group (Figure 6).

Following the different thresholds considered in the process of transformation, a country could find itself in a different category over time. The general occurrence of classes over time is presented in Figure 7.

Figure 8 demonstrates that the positive trend occurred almost every year, making their number higher than the negatives in each category. This trend confirms the general increase in CO₂ emissions worldwide despite efforts to reduce it. With this categorization complete, the performance of the classifiers applied to these data is provided in

Table 3. Confusion matrix and classification report.

	Confusion Matrix						Precision	Recall	F1-Score	Support
High negative	12	0	0	0	0	0	0.71	1.00	0.83	12
High positive	0	27	0	0	0	0	1.00	1.00	1.00	27
Low negative	0	0	15	0	1	0	1.00	0.94	0.97	16
Low positive	0	0	0	23	0	0	0.96	1.00	0.98	23
Moderate negative	5	0	0	0	28	0	0.97	0.85	0.90	33
Moderate positive	0	0	0	1	0	18	1.00	0.95	0.97	19
	High negative	High positive	Low negative	Low positive	Moderate negative	Moderate positive

.

It appears that the Decision Tree model performed better compared to the other classifiers. Its flexibility and capacity to analyze data regardless of prior consideration regarding the distribution of the input data makes this model robust enough to capture hidden trends. Even after a search of the best parameters, the confusion matrix, as well as the classification report of the DT, present errors in the prediction of 1 instance out of 19 from the class of Moderate positive (providing a recall of 0.95, a precision of 1.00 and an F1 score of 0.97), which is predicted as Low positive, 5 instances out of 33 from the class of Moderate negative (having a recall of 0.85, a precision of 0.97 and an F1 score of 0.97), predicted as High negative, and 1 instance out of 16 from the class of High positive (displaying a recall of 0.94, a precision of 1.00 and an F1 score of 0.97), predicted as Moderate negative.

Based on this result, and despite the small misclassification, the model could capture the general short-term trend (3-year average) of the target. The features contributing to this prediction are used to understand their interplay and contribution over time on the target.

3.3. Feature Analysis

3.3.1. Feature Importance Analysis

Using the decision tree algorithm after a grid search of the best parameters, this analysis reveals that eight features contribute to the prediction (Figure 9).

From the present group of features, two features from the economic group (change in GDP and GDP per capita), one from the population group (population number), one from the education group (population—education: incomplete primary), one from CO₂-related industry (CO₂ from oil), and the remaining three from CO₂-related activity (GHG by world region, share of cumulative CO₂ emissions and the mean value of 3 years) contributed to the prediction. A summary of the PDPs (

Table 4. Summary of PDPs analysis.

Class	Increase	Decrease
High negative	Population number	3-year average
High positive	CO2 from coal Population number 3-year average	---
Low negative	---	Population number GDP 3-year average
Low positive	3-year average	CO2 from oil Population education: Incomplete Primary Population number Change in GDP
Moderate negative	GDP	Population number 3-year average
Moderate positive	CO2 from oil Population education: Incomplete Primary Change in GDP 3-year average	CO2 from coal Population number

) explains their contributions.

Evaluating each country’s trend over time for the features considered (Figure 10) provides a better understanding of the overall dynamics.

In all the categories, the increase or decrease in the 3-year average determines the sign (positive or negative) of the category (Figure 10j). Coupled with the fluctuation in the 3-year average, the dynamics in the population number influences the categorization of a country. High emitters have a large population compared to others (Figure 10a). Over time, countries that have an important variation in the change in GDP tend to emit less compared to those that have low variations (Figure 10f). The trend of the GDP per capita (Figure 10g) coupled with the GDP (Figure 10l) suggests that they cannot clearly explain the trend in CO₂ emissions, since some rich countries emit less compared to others. It also appears that the level of education (Figure 10d,i), more specifically, early access to education, could potentially explain the target. When it comes to the CO₂ related industry (Figure 10b,k), countries that emits the more (Figure 10e,h) appear to be the wealthy (Figure 10l), and have a lower variations in the GDP (Figure 10f) compared to the other group of countries. Furthermore, in the group of countries considered, the wealthier a country is, the higher the number of children that have early access to education (Figure 10d,i), result demonstrating the complexity of possible dynamics. This analysis could depicts the existing but complex interaction among features, which, once understood, could deepen our understanding of the present dynamics. To capture possible dependency between features, which could not be achieved using the feature importance, or PDPs which assumes independence among features, the sensitivity analysis was applied.

3.3.2. Sensitivity Analysis

Using 1000 samples with bounds between −5 and 9, it appears that seven variables, slightly different to those from the feature important analysis, with an impact on the performance of the model (Figure 11)

In capturing possible interactions among features, this analysis could identify and quantify four key groups that contribute to the short-term dynamics of year-on-year changes in CO₂ emissions. These groups are population, CO₂-related industries (including coal and oil), economic activity assessed through the GPD, the contribution to temperature rise, associated CO₂ emissions, and early access to education assessed by the population—education: incomplete primary. These features could be grouped into two: those having a direct impact (population, CO₂ from coil, oil, mean-3y), and those which explain it indirectly (GDP, contribution to temperature rise and population—education: incomplete primary).

3.3.3. Correlation Analysis

To deepen the understanding on the interaction of features, the Pearson and Canonical correlation was used. Figure 12 provides the correlation table of the features in the dataset.

In comparing the group of features, there is a strong positive correlation among variables in the CO₂ emissions and CO₂-related industry with the group related to the level of education, economic features, population number, but not with the population growth rate or GDP per capita. In more detail, there is a strong positive correlation (≥0.50) between the annual CO₂ emissions and GDP (0.84), population number (0.55), population—education: lower primary (0.50), population—education: lower secondary (0.71), population—education: upper secondary (0.82), population—education: post secondary (0.86), contribution to temperature rise (0.86), share of cumulative CO₂ emissions (0.71), cumulative CO₂ emissions (0.85), CO₂ from coal (0.92), CO₂ from oil (0.84), CO₂ from gas (0.73), CO₂ cement (0.77), GHG emissions by world region (0.98); there is also a strong negative correlation between this same variable and the population growth rate (−0.37). This result confirms the existing studies about the interplay of the considered variables to the emission of CO₂. For instance, rich countries are high polluters, and the concentration of population is one reason behind CO₂ emissions fluctuations. Also, education plays an important role in understanding and developing ways to mitigate CO₂ emissions [8]. While such observation seems straightforward, it is not the case for the absolute change in CO₂ emissions which is the target variable. Indeed, only population—education: primary (0.53/0.63), population—education: lower secondary (0.6/0.72), CO₂ from coal (0.57/0.70) and CO₂ from cement (0.56/0.69) have a strong positive correlation with the target. In the short-term, we can observe an average increase of 1.26% in the correlation coefficient between CO₂-related features with the target, as well as in the education features and population number, in comparison to what it was with the absolute change in CO₂ emissions. Thanks to the Canonical correlation, it is possible to deeply visualize this correlation direction. However, when it comes to the corresponding categories, this direction is still strong, but changes to a negative. This suggests that as the value increases, it is more likely for the target to be in the High positive category. To deepen understanding on the interplay of features already provided by the sensitivity analysis, the canonical correlation analysis allows us, through plotting, to visualize the direction of these features over time. To achieve this, after grouping features based on their groups, a comparison between them was achieved in the following order: population with CO₂ industry, population with CO₂ related emissions, population with education, population with economy, CO₂ industry with CO₂-related emissions, CO₂ industry with education, CO₂-related emissions with education, CO₂ industry with economy, CO₂-related emissions with economy, and education with economy. Figure 13 presents the result of this analysis.

Two trends are displayed from this analysis. Over time, there is a linear trend in the group of CO₂-related emissions and industry with education (Figure 10g and Figure 13f), and economy (Figure 13h,i), population with education (Figure 13c), and education with economy (Figure 13j), and a nonlinear one for population with CO₂-related emissions and industry (Figure 13a,b) and population with economy (Figure 13d). This result suggests a strong positive correlation between the group of features having a significant impact on the fluctuations of the CO₂ emissions overtime. In the group of countries considered, regardless of the decreasing trend in population growth rate of some countries, this analysis unveils that the dynamics in the population differently affect each country’s economy. As the economy grows, it tends to positively impact the CO₂ emissions, influencing the short-term variations of CO₂ emissions (Figure 13h,i). Early access to education displays a linear trend with the economy growth (Figure 13j), CO₂ emissions (Figure 13i) and population dynamics (Figure 13d), suggesting that the more a country emits, the more it becomes wealthy and is able to implement laws to support education. The level of education increases with the growth of the economy (Figure 13j) and CO₂ emissions (Figure 13g), suggesting the wealthier the country, access to education become easy overtime. The contribution to temperature rise does not solely depend on the emissions of CO₂. However, as for the early access to education, this indicator is meaningful in explaining the target of this analysis. These trends could help anticipate future dynamics in the monitoring of CO₂ emissions, resulting in the implementation of adequate policy to tackle this threat while maintaining a good level of economy.

3.4. Discussion

The significant improvement in model performance from an R-squared of 67% to an accuracy of 96% after applying the proposed data transformation technique indicates the effectiveness of this approach. Indeed, the increase in accuracy demonstrates that the transformation technique was successful in capturing the underlying patterns in the data. This suggests that CO₂ emissions changes are not just a function of the factors considered but also their volatility and direction. By categorizing the data into six classes of volatility (high positive and negative, moderate positive and negative, and low positive and negative), the model can capture more nuanced information. This approach recognizes that the rate and direction of change can be just as important as the magnitude of the change itself.

The use of percentiles to define thresholds for volatility is a data-driven approach that adapts to the specific characteristics of the dataset. This method is more flexible and potentially more accurate than using arbitrary or fixed thresholds.

Furthermore, the technique is designed to be adaptable to other datasets, enhancing its generalizability. This is a significant contribution, as it means the approach could be applied to other countries or variables, extending its usefulness beyond this specific study.

Some statistics of the categories provided in

Table 5. Summary of mean values by features in each category.

Features	High Negative	High Positive	Moderate Negative	Moderate Positive	Low Negative	Low Positive
Population (number)	4.81 × 108	8.07 × 108	7.18 × 107	7.91 × 107	6.74 × 107	6.74 × 107
CO2 from coal	1.45 × 109	1.76 × 109	1.15 × 108	1.24 × 108	1.04 × 108	8.21 × 107
Mean-3y	−7.31 × 107	1.07 × 108	−1.07 × 107	1.06 × 107	−3.02 × 106	4.76 × 106
GDP	9.99 × 1012	4.27 × 1012	1.96 × 1012	9.79 × 1011	1.72 × 1011	1.38 × 1011
CO2 from oil	1.58 × 109	8.85 × 108	2.12 × 108	1.71 × 108	2.43 × 107	2.31 × 107
Population—Education: Incomplete Primary	2.04 × 107	4.32 × 107	2.10 × 106	5.82 × 106	3.90 × 106	3.94 × 106
Contribution to temperature rise	0.173045	0.098758	0.029083	0.022507	0.010212	0.008529

, coupled with the sensitivity analysis result suggest that:

• High negative: Countries in this group have a high average population and GDP, a high CO₂ emission from both coal and oil. Despite their high GDP and CO₂ emissions, these countries have seen a decrease in CO₂ emissions over time. However, they also have a high contribution to temperature rise. These countries’ high contribution to global warming is due to their extensive use of fossil fuels for energy production, industrial processes, and transportation. Despite recent decreases in emissions, the cumulative effect of their past and present emissions continues to drive global temperature rise.
• High positive: Countries in this group also have a high average population, a slightly lower GDP compared to the High negative group, and they have seen an increase in CO₂ emissions over time. Surprisingly, they have a lower contribution to temperature rise compared to the High negative group. This could be due to their past contribution.
• Low negative: Countries in this group have a lower average population and GDP compared to the High groups. They have lower CO₂ emissions from both coal and oil and have seen a decrease in CO₂ emissions over time. Finally, they contribute less to temperature rise compared to the High groups.
• Low positive: Countries in this group have a similar average population to the Low negative group. They have a lower GDP compared to the Low negative group but they have seen an increase in CO₂ emissions over time. They also contribute less to temperature rise compared to the High groups.
• Moderate negative: Countries in this group have a moderate average population and GDP. They have moderate CO₂ emissions from both coal and oil, and they have seen a decrease in CO₂ emissions over time. They contributed moderately to temperature rise compared to the High groups.
• Moderate positive: Countries in this group have a similar average population to the Moderate negative group. They have a lower GDP compared to the Moderate negative group and they have seen an increase in CO₂ emissions over time. They have a lower contribution to temperature rise compared to the Moderate negative group.

A comparison of the two categories, High negative and High positive, suggests that:

• Population (number): The average population is higher in the High positive group (approximately 807 million) compared to the High negative group (approximately 481 million). This suggests that countries with larger populations tend to have increasing CO₂ emissions.
• CO₂ emissions from Coal: Both groups have high CO₂ emissions from coal, but the High positive group has slightly higher emissions on average (approximately 1.76 billion tonnes) compared to the High negative group (approximately 1.45 billion tonnes).
• CO₂ emissions from Oil: The High negative group has higher CO₂ emissions from oil (approximately 1.58 billion tonnes) compared to the High positive group (approximately 885 million tonnes).
• 3-Year Mean Change in CO₂ emissions (Mean-3y): The High negative group shows a decrease in CO₂ emissions over time (average change of −73 million tonnes), while the High positive group shows an increase (average change of 107 million tonnes).
• GDP: is higher on average in the High negative group (approximately 9.99 trillion USD) compared to the High positive group (approximately 4.27 trillion USD). This suggests that wealthier countries tend to have decreasing CO₂ emissions.
• Population with Incomplete Primary Education: The High positive group has a higher average population with incomplete primary education (approximately 43.2 million) compared to the High negative group (approximately 20.4 million).
• Contribution to Temperature Rise: The High negative group has a higher average contribution to temperature rise (0.173) compared to the High positive group (0.099).

The High negative group tends to have wealthier countries with larger CO₂ emissions from oil and a larger contribution to temperature rise. On the other hand, the High positive group tends to have countries with larger populations, higher CO₂ emissions from coal, and a larger population with incomplete primary education.

Considering the Low negative and Low positive groups:

• Population: The average population is approximately the same in both groups (approximately 67 million). This suggests that population size does not significantly differentiate these two groups.
• CO₂ emissions from Coal: The Low negative group has slightly higher CO₂ emissions from coal on average (approximately 104 million tonnes) compared to the Low positive group (approximately 82 million tonnes).
• 3-Year Mean Change in CO₂ emissions (Mean-3y): The Low negative group shows a decrease in CO₂ emissions over time (average change of −3 million tonnes), while the Low positive group shows an increase (average change of 4.8 million tonnes).
• GDP: The GDP is slightly higher on average in the Low negative group (approximately 171 billion USD) compared to the Low positive group (approximately 138 billion USD).
• CO₂ emissions from Oil: The Low negative group has slightly higher CO₂ emissions from oil (approximately 24 million tonnes) compared to the Low positive group (approximately 23 million tonnes).
• Population with Incomplete Primary Education: The Low positive group has a slightly higher average population with incomplete primary education (approximately 3.94 million) compared to the Low negative group (approximately 3.90 million).
• Contribution to Temperature Rise: The Low negative group has a slightly higher average contribution to temperature rise (0.0102) compared to the Low positive group (0.0085).

The Low negative group tends to have slightly higher CO₂ emissions from coal and oil, a higher GDP, and a higher contribution to temperature rise, but shows a decrease in CO₂ emissions over time. On the other hand, the Low positive group tends to have a slightly larger population with incomplete primary education and shows an increase in CO₂ emissions over time.

Finally, for the Moderate negative and Moderate positive groups:

• Population: The average population is slightly higher in the Moderate positive group (approximately 79 million) compared to the Moderate negative group (approximately 72 million).
• CO₂ emissions from Coal: The Moderate positive group has slightly higher CO₂ emissions from coal on average (approximately 124 million tonnes) compared to the Moderate negative group (approximately 115 million tonnes).
• 3-Year Mean Change in CO₂ emissions (Mean-3y): The Moderate negative group shows a decrease in CO₂ emissions over time (average change of −10.7 million tonnes), while the Moderate positive group shows an increase (average change of 10.6 million tonnes).
• GDP: The GDP is significantly higher on average in the Moderate negative group (approximately 1.96 trillion USD) compared to the Moderate positive group (approximately 979 billion USD).
• CO₂ Emissions from Oil: The Moderate negative group has higher CO₂ emissions from oil (approximately 212 million tonnes) compared to the Moderate positive group (approximately 171 million tonnes).
• Population with Incomplete Primary Education: The Moderate positive group has a higher average population with incomplete primary education (approximately 5.82 million) compared to the Moderate negative group (approximately 2.10 million).
• Contribution to Temperature Rise: The Moderate negative group has a higher average contribution to temperature rise (0.029) compared to the Moderate positive group (0.0225).

The Moderate negative group tends to have higher GDP, higher CO₂ emissions from oil, and a higher contribution to temperature rise, but shows a decrease in CO₂ emissions over time. On the other hand, the Moderate positive group tends to have a larger population, higher CO₂ emissions from coal, and a larger population with incomplete primary education, but shows an increase in CO₂ emissions over time. Countries with higher populations and GDPs tend to have higher CO₂ emissions and contribute more to temperature rise. However, some of these countries have seen a decrease in CO₂ emissions over time, suggesting that they may be taking steps to mitigate their impact on climate change. Countries with lower and moderate populations and GDPs show a diverse range of CO₂ emissions and contributions to temperature rise, and some of these countries are effectively managing their CO₂ emissions while others are still facing challenges. To provide a rough estimate of the shift from one category to another, we can consider the average values of the key variables for each category. For instance, the difference in average population between the High and Low categories is approximately 400 million. Therefore, an increase in population by this amount could potentially cause a shift from Low to High, or vice versa. For the CO₂ emissions from coal, the average difference the High and Low categories is approximately 1.3 billion tonnes. Therefore, an increase in CO₂ emissions from coal by this amount could potentially cause a shift from Low to High, or vice versa. The 3-Year Mean Change in CO₂ Emissions (Mean-3y) category indicates that the difference on average in the High and Low categories is approximately 80 million tonnes. Therefore, a change within the 3-year mean change in CO₂ emissions by this amount could potentially cause a shift from negative to positive, or vice versa. The difference in average GDP between the High and Low categories is approximately 9 trillion USD, suggesting that an increase in GDP by this amount could potentially cause a shift from Low to High, or vice versa. Concerning the Population with Incomplete Primary Education category, on average, the difference in between the High and Low categories is approximately 20 million, meaning that an increase in this population by this amount could potentially cause a shift from Low to High, or vice versa. Finally, in the Contribution to Temperature Rise category, the difference in average contribution to temperature rise between the High and Low categories is approximately 0.07; thus, an increase in the contribution to temperature rise by this amount could potentially cause a shift from Low to High, or vice versa. These estimates provide a rough idea of the magnitude of change in each variable that could potentially cause a shift from one category to another. However, it is important to note that these are just estimates, and the actual thresholds might be different due to the complex interactions among the variables.

Putting together the results of the feature importance analysis, PDPS and correlation analysis, this study could pinpoint the complexity of explaining the short-term trend in CO₂ emissions on a global scale. Indeed, it appears that no matter the country, the number of its inhabitants is the most important signal about future CO₂ emissions, and thus, its change over time. This is explained by the human impact on its direct environment in terms of construction, deforestation, etc. [40]. Fossil fuels remain a threat to the environment. This study demonstrates how particular attention should be paid to coal and oil production, since they can solely and in a very short time negatively impact the environment. This matter is quite complex because these two are strongly correlated with the wealth of countries, making it critical to find alternatives [41]. Indirectly, early access to education, similarly to the monitoring of the temperature rise appear to be among the game changers in this matter, suggesting a rapid possibility of improvement if properly used.

4. Conclusions

This study analyzed the short-term variations in CO₂ emissions across ten countries. By employing machine learning techniques on a unique dataset, we identified the key factors influencing these variations. Population growth, particularly population size, and the coal industry emerged as strong contributors. Early access to education and contribution to temperature rise, while less impactful, warrant further investigation. This result sheds light on critical factors and their contribution to the year-on-year change in CO₂ emissions over time and could potentially contribute in the implementation of policy that will address education and environment by promoting investment in early childhood education (The analysis suggests a link between early education and economic growth, potentially leading to higher CO₂ emissions later). By investing in early education, countries might be able to foster more sustainable education, focused on environmental awareness, population and economy, promoting family planning and economic incentives, and considering the complex relationship between population growth and economy as suggested in this analysis. When a country reaches a certain level of population and pollution, policies that encourage smaller families, coupled with economic incentives, could help manage population growth while maintaining economic stability. The level of granularity of the data, coupled with the considered group of features, represent the major limitations of this study. Indeed, a yearly analysis does not capture the monthly or weekly trend in the data, which could provide more insight in the interpretation of the result. Also, the group of features considered are not the only ones to explain the change in CO₂ emissions. Further analysis will consider reducing the level of granularity of the dataset and the adding other groups of features, which could potentially explain the target in different terms (mid or long-term analysis).