In this study, nine sampling sites in the CT groundwater of Abidjan were studied. Given the impact of the rainy season on CT groundwater recharge and imploring the representative sampling method, the National Public Hygiene Laboratory in Abidjan provided data (98 samples/site) from their sampling campaigns from May 2nd to September 1st, 2022. These samples were considered sufficiently representative to approximate the characteristics of the broader population from which they were extracted [46] . A total of twelve indicators have been assessed in compliance with World Health Organization (WHO) specifications, notably pH, electric conductivity (EC), total dissolved solids (TDS), chloride (Cl⁻), sulfate ( ${\text{SO}}_4^{2-}$), bicarbonate ( ${\text{HCO}}_3^{-}$), sodium (Na⁺), potassium (K⁺), magnesium (Mg²⁺), calcium (Ca²⁺), nitrate ( ${\text{NO}}_3^{-}$), and aluminum (Al³⁺). The methodology comprises two key phases designed to achieve the research objectives. Initially, the computation of the WQI was conducted using the Analytic Hierarchy Process (AHP) to determine the relative weight of parameters. Subsequently, the prediction of the AHP-based WQI was executed, applying the twelve physicochemical parameters as features and the WQI as label.

The Water Quality Index is a popular method with straightforward mathematical features and a user-friendly interface. This approach uses data or indicators from a water body to determine a unique number. The obtained number is then classified into a category that reflects the water quality status [19] . The analytical hierarchy process (AHP) is a multi-criteria decision analysis method (MCDM) developed by Saaty [47] . In recent years, numerous researchers have used AHP to calculate the relative weight of parameters for different groundwater studies [48] - [50] . This method has the advantage of limiting the possibility of errors and incorrect distribution of weight to a specific parameter [50] [51] . Additionally, the consistency of the approach can be verified after computation, and the pair-wise comparison matrix generated in the calculation is a promising approach for providing relative weights [20] .

The AHP-weighting was implemented through a four-step process.

Step 1: The parameters are hierarchically organized considering their nature or function.

Step 2: Scales of 1 to 9 ( Table 1 ) are attributed to the parameters based on their significance in the overall water quality.

Step 3: The pair-wise comparison matrix is constructed. Considering n evaluation factors and the importance intensity of factor i over j. The pair-wise comparison matrix A gives:

$A=\left[\begin{array}{cccc}{a}_{11}& a_{12}& \cdots & a_{1n}\\ a_{21}& a_{22}& \cdots & a_{2n}\\ ⋮& ⋮& \ddots & ⋮\\ a_{n1}& a_{n2}& \cdots & a_{nn}\end{array}\right]$(1)

a_ij = W_i/W_j ; where W_i and W_j represent the assigned scores to factors i and j, respectively. This value is supposed to approximate the relative importance of i to j.

The factors eigenvectors are then calculated using the average of the normalized column, presented as

$w_i=\frac{1}{n}{\displaystyle {\sum }_{j=1}^n\frac{a_{ij}}{{\displaystyle {\sum }_{i=1}^n{a}_{ij}}}}$; ( $i,j=1,2,\cdots ,n$)(2)

Step 4: The consistency index (CI) and the consistency ratio (CR) are calculated to validate the judgment of the different values assigned to parameters.

$CR=\frac{CI}{RCI}$(3)

where

$CI=\frac{{\lambda }_{\max }-n}{n-1}$(4)

With n, the pair-wise comparison matrix size, and ${\lambda }_{\max }$, the principal eigenvalue of the matrix A. The consistency is observed only when matrix A is a positive reciprocal one and ${\lambda }_{\max }$ is closer to n.

${\lambda }_{\max }={\displaystyle {\sum }_{i=1}^n\frac{{\left(A\stackrel{\to }{w}\right)}_i}{n{w}_i}}$(5)

where $A\stackrel{\to }{w}=A_1{W}_n$ is the product of the geometrical mean of the pair-wise comparison matrix rows and the normalized weight; $n{w}_i$ is the normalized weight of each variable $W_n$.

RCI is the average random consistency index, computed and tabulated as shown in the following Table 2 . If CR < 0.1, the numerical judgments will be considered acceptable.

The calculation is performed using the normalized rating scale (q_i), which is obtained by subtracting the measured value of a parameter from the acceptable level and then dividing it all by the permissible level. The equation is formulated as follows:

$q_i=\frac{A_l-V_m}{A_l}$(6)

where $A_l$ is the acceptable level set by the WHO [52] for each parameter in drinking water and $V_m$ is the measured value of this parameter.

For computing WQI, the sub-index SI for each parameter gives:

$S{I}_i=W_i\times q_i$(7)

where $W_i$ is the normalized weight of the i^th parameter.

Then, $\text{WQI}={\displaystyle {\sum }_{i=1}^{n}S{I}_i}$(8)

The computed WQI values are finally classified into five categories ( Table 3 ): excellent, good, poor, very poor, and unfit water for drinking purposes [53] .

Our study experiment was conducted in the Windows 11 system, using Python (version 3.11.0) as the programming language for code programming. The data reading and preprocessing via Pandas (version 2.1.4). Machine learning models were built through scikit-learn (version 1.2.2), and matplotlib (version 3.8.0) was used to draw graphics. First, the data preprocessing consisted of filling the missing values with 0 using the Pandas tool. The second step focused on feature scaling, with the min-max normalization technique commonly applied for standardization tasks [54] . These features range between 0 and 1, helping to improve the convergence speed and performance of specific machine-learning algorithms.

$x^{\prime }=\frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$(9)

Here, $x$ is the original data, $x^{\prime }$ represents the normalized data, $\max \left(x\right)$ represents the maximum value of the original data while $\min \left(x\right)$ represents the minimum value of the original data.

The third step encompasses data splitting into training and test sets. The training set allows ML algorithm data to learn how to generate predictions and uncover insights within the dataset, while the test set is used to execute the model. The current experiment implies time series data, where data before August 1, 2022, were used as the training set and those after August 1, 2022, as the test set. Based on the slide window projection, data features for seven consecutive days have been input in each model’s windows (window size = 10) to enable water quality predictions after seven days.

The basic principle of ridge regression analysis introduced by Hoerl and Kennard [55] , established research to estimate biases with a smaller Mean squared error (MSE). By adding a small constant ‘k’ to the diagonal of the correlation matrix of independent variables, the error variance can be considerably reduced to manage the overall instability of ordinary least square (OLS) estimates. The calculation formula gives:

${\hat{\beta }}_{\text{ridge}}=\underset{\beta }{\text{argim}}\left\{{\displaystyle {\sum }_{i=1}^N{\left(d_i-{\beta }_0-{\displaystyle {\sum }_{j=1}^p{x}_{ij}{\beta }_j}\right)}^2}+\lambda {\displaystyle {\sum }_{j=1}^p{\beta }_j^2}\right\}$(10)

With $\hat{\beta }$ as the ridge estimator, N samples, the constant ‘k,’ the p covariates, and $x_i={\left(x_{i1},x_{i2},\cdots ,x_{ip}\right)}^{\text{T}}$ the covariate vector, as well as $\lambda$ which refers to the shrinkage parameter indicating the larger the value of $\lambda$, the regression coefficients decline further as they approach zero. Before implementing Ridge regression, the inputs need to be standardized because the obtained solutions are not homogeneous when the input variables are scaled. Also, the intercept ${\beta }_0$ has not been penalized in the Ridge regression.

Lasso is a linear model with regularization, initially introduced in the context of least squares. Tibshirani [56] , presented the model as an innovative variable selection mechanism for regression, optimizing the residual sum of squares while constraining the total absolute coefficient magnitudes. Lasso is a well-known regression method that regularizes the parameter β under the sparse assumption [57] . Like ridge regression, the model considers N sample, p covariates, and a single outcome. Supposing $y_i$ is the response variable, and $x_i={\left(x_{i1},x_{i2},\cdots ,x_{ip}\right)}^{\text{T}}$ is the covariate vector for the i^th case $\beta =\left({\beta }_1,{\beta }_2,\cdots ,{\beta }_p\right)$, the following relation gives the mathematical expression:

${\displaystyle {\sum }_{i=1}^n{\left({\beta }_0+{\displaystyle {\sum }_{k=j}^K{\beta }_k{x}_{k,i}}-y_i\right)}^2}$(11)

In Lasso regression, λ subject is computed to minimize the residual sum of the square:

${\displaystyle {\sum }_{i=1}^n{\left({\beta }_0+{\displaystyle {\sum }_{k=j}^K{\beta }_k{x}_{k,i}}-y_i\right)}^2}+\lambda {\displaystyle {\sum }_{k=1}^k{\beta }_k}$(12)

The Gradient Boosting (GB) approach is an ensemble learning technique [27] that integrates multiple weak models to enhance overall performance. The GB algorithm consists of three key elements: the loss function, the weak learner, and the additive model. The loss function component must constantly be optimized to reduce the prediction error. Gradient Boosting is based on sequential learning using weak learners like the decision trees model to mitigate the losses without altering the existing tree. Instances mistakenly identified in one phase are assigned additional weight in the subsequent step. This way, each subsequent weak learner is designed to address and rectify the errors made by the previous models [58] [59] . Friedman provided the entire calculation process [60] , explained as follows:

Considering the predictors x, the targe y, and the number of observations $n=1,\cdots ,n$.

$F_0\left(x\right)=\underset{\gamma }{\arg \min }{\displaystyle {\sum }_{i=1}^{n}L\left(y_i,\gamma \right)}$(13)

where $F_0$ is the initial constant value prediction and L, the loss function.

Residuals computation: $r_{im}=-{\left[\frac{\partial L\left(y_i,F\left(x_i\right)\right)}{\partial F\left(x_i\right)}\right]}_{F\left(x\right)=F_{m-1}\left(x\right)}$(14)

for $i=1,\cdots ,n$ and $F_{m-1}$, previous prediction.

for $j=1,\cdots ,J_m$; with the features x against $\gamma$ $j$, the terminal or leave node and J, the total number of leaves.

Computation of ${\gamma }_{jm}=\underset{\gamma }{\arg \min }{\displaystyle {\sum }_{x_i\in R_{jm}}L\left(y_i,F_{m-1}\left(x_i\right)+\gamma \right)}$(15)

for $j=1,\cdots ,J_m$

Model update: $F_m\left(x\right)=F_{m-1}\left(x\right)+v{\displaystyle {\sum }_{j=1}^{J_m}{\gamma }_{jm}1\left(x\in R_{jm}\right)}$(16)

Verifying the accuracy of the constructed models is an essential phase in predictions; otherwise, the predictive model would lack scientific significance [61] . The performance assessment indicators for our study’s prediction include the mean square error (MSE), the root mean square error (RMSE), the Pearson, the Spearman, and the Kendall correlations. The Pearson, Spearman, and Kendall correlation coefficients represent the consistency evaluation metrics. These indicators are calculated as follows:

$\text{MSE}=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{\left(x_i-y_i\right)}^2}$(17)

$\text{RMSE}=\sqrt{\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{\left(x_i-y_i\right)}^2}}$(18)

where, $x_i$ is the i^th observed value, $y_i$ is the corresponding predicted value, and n the number of observations.

MSE and RMSE values approaching 0 imply higher prediction accuracy, whereas values approaching 1 reflect lower prediction performance [62] .

The correlation coefficients indicate the degree of linear association between two variables, with 1 indicating a positive relationship, -1 indicating a negative relationship, and 0 indicating no linear association [63] [64] . For variables a and b with n data sets (each data set can be labeled a_i and b_i, where $i=1,2,\cdots ,n$), the coefficient is used to estimate the correlation r between a and b.

Pearson: $r_p=\frac{{{\displaystyle \sum }}_{i=1}^n\left(a_i-\bar{a}\right)\left(b_i-\bar{b}\right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(a_i-\bar{a}\right)}^2}\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(b_i-\bar{b}\right)}^2}}$(19)

Spearman: $r_s=1-\frac{6\sum d_i^2}{n\left(n^2-1\right)}$(20)

Kendall: $\tau =\frac{n_c-n_d}{\frac{1}{2}n\left(n-1\right)}$(21)

where $\bar{a}$ and $\bar{b}$ are the means of a and b, respectively, $d_i=R\left(a_i\right)-R\left(b_i\right)$ is the difference between the two ranks of each variable and $n_c$ is the number of concordance and $n_d$ the number of discordances.

Table 4 presents the descriptive statistics of all the analytical data of the groundwater samples collected from the study area, majorly during the long rainy season. Most parameters were found within permissible limits of the WHO [52] , except for K⁺ and Al³⁺, which exceeded the permissible thresholds in 4.76% and 3.4% of samples, respectively. In the CT aquifer, water is particularly acidic, with pH values between 4 and 5.9. The primary factor predicting this pH is the acidity of the leached ferrallitic soil of the southern coastal plateau landscapes, which is 3.9 according to the soil record of the Côte d’Ivoire sedimentary basin [65] . Acidic rainwater recharge also influences the pH of the resource [66] . The observed EC varies from 4.1 to 469.0 µS/cm, indicating a low conductivity due to groundwater mixing, as the CT is an unconfined aquifer. After surface recharge enters the aquifer, a significant dilution occurs due to less mineralization and more acidic water [43] . The relatively high values correspond to areas that lack a functional sanitation infrastructure [67] . The observed TDS varies from 0.0629 mg/L to 214.315 mg/L, which ascertains the low salinity of the resource and the groundwater mixture.

The anions order in the overall samples is ${\text{HCO}}_3^{-}$ > Cl⁻ > ${\text{NO}}_3^{-}$ > ${\text{SO}}_4^{2-}$. The recorded ${\text{HCO}}_3^{-}$ level varies between 0 and 240.5 mg/L. In some samples, higher concentrations of HCO₃ can be justified by the weathering of silicates [68] , occurring when they react with CO₂ gas from root respiration and aqueous CO₂ from the water, leading to a potential release of Ca, Mg, and HCO₃ into the groundwater [69] . The measured concentration of Cl⁻ varies from 1.68 mg/L to 16.908 mg/L, while the nitrate ranges from 0 to 48.73 mg/L. The values of ${\text{NO}}_3^{-}$ close to permissible limits are symptomatic of sewage effluent drainage and decomposition of organic material with precipitation in the unconfined CT groundwater. The sulfate content ranges from 0 to 18.4 mg/L as the origins of ${\text{SO}}_4^{2-}$ in groundwater are mainly from the dissolution or oxidation of sulfate minerals and acidic rainwater recharge [70] .

The cationic order in the CT groundwater samples is Ca²⁺ > Na⁺ > Mg²⁺ > K⁺ >

Al³⁺. The Ca²⁺ value indicates variations from 0 to 51.3 mg/L. The primary factors influencing the presence of Ca²⁺ in groundwater are the calcium-rich silicate weathering and the process of ion exchange [71] . The Na⁺ concentration fluctuates between 1.056 mg/L and 22.4 mg/L, while the Mg²⁺ fluctuates between 0 and 17.07 mg/L. The recorded K⁺ varies between 0 and 33.61 mg/L. High levels of K⁺ observed in a few samples are due to alteration products such as kaolin in clay formations [72] , of the aquifer. For the Al³⁺ content, the measured concentration ranges between 0.001 mg/L and 4.209 mg/L and the highest Al³⁺ values potentially come from the geological formation of the CT aquifer [73] .

Applying the AHP methodology, the structure was organized into three hierarchies given major ions percentage in samples ( Figure 2 ): (1) physical parameters such as pH, EC, and TDS, (2) chemical parameters like Cl⁻, ${\text{SO}}_4^{2-}$, ${\text{HCO}}_3^{-}$, Na⁺, K⁺, Mg²⁺, and Ca²⁺, and (3) health-threatening chemical parameters such as ${\text{NO}}_3^{-}$ and Al³⁺. Scores of 5 and 9 were respectively given to the first and third structures. The physical variables are undoubtedly essential; however, the health-threatening parameters are paramount in groundwater. Additionally, structure (2) was divided into two subgroups, (a) and (b), with respective rates of 6 and 4 based on anions and cations orders in the samples. The consistency ratio calculation gives CR = 0.024 < 0.1 ( Table 5 ), validating the generated pair-wise matrix. This result of CR enables the application of normalized weights as relative weights for WQI determination.

The obtained plots ( Figure 4 ) suggested that the predicted WQI values vary within the same range as the measured values for locations AK, NN, NR, RC, ZO,

and ZN. Meanwhile other sites, DJI, ZE, and ZE, discrepancies are identified between the actual and predicted values at the same time. However, the performance metrics outlined in Table 9 indicate that the model has low performance (MSE and RMSE close to 1) at these sites, while the remaining sites showcase good performance (MSE and RMSE close to 0). Moreover, the indicators exhibit inconsistent model execution for most monitored sites like Ridge regression, with r, r_s, and τ close to 0.

Table 10 presents the results of the gradient-boosting regression performance evaluation for the CT groundwater quality prediction. The computed metrics at each site display satisfactory performance and consistency in the prediction. This is translated by the low error levels (close to 0) and high correlations (close to 1), with values ranging from 0.097 to 0.112 and 0.311 to 0.334 for MSE and RMSE, respectively. Also, r ranges from 0.439 to 0.966, r_s from 0.632 to 0.904, and τ from 0.836 to 0.945. The real and predicted WQI plots ( Figure 5 ) variations corroborate

the performance metrics results, showing fluctuation of predicted values in line with the trend of actual values in all sites. This confirms the accuracy of the Gradient-boosting model.

In the last decade, WQI models have been widely used to assess water quality for various contexts. Nonetheless, a major challenge faced while using WQIs is the uncertainties from incorrect weight assignment in the methodology [20] . Our study agrees that the AHP weight-based WQI model represents the ideal approach, highlighting its advantages over alternative models like weighted arithmetic WQI or entropy weight-based WQI. The AHP model allows for a systematic and structured approach to decision-making, ensuring that the most critical parameters are emphasized in the evaluation process [75] . This model focuses on the uncertainty in the data as well as considers the contextual significance of certain parameters that are better evaluated through expert input [20] [76] . Additionally, the AHP model’s flexibility allows for incorporating local knowledge and specific health impacts associated with groundwater quality parameters, making it more suitable for targeted assessments in diverse environments [77] .

Regarding the prediction phase, the three regression models (Ridge, Lasso, and Gradient Boosting) were executed. Choosing these models over others like random forest, artificial neural network, or deep learning can be attributed to different factors. First, the interpretability of ridge and lasso regression is significantly higher. For instance, Deep learning might give accurate predictions, however, hardly show how predictions are made, leading to its identification as “black boxes” [78] , posing limitations as researchers and stakeholders need to trust or understand the results. The ability to interpret prediction outcomes is essential for effective management and policymaking [29] . Next, these complex models typically require large datasets to train accurately and avoid overfitting. In the current framework, the available datasets do not effectively support such models. In contrast, traditional regression models can perform well even with smaller datasets, making them more practical for our scenario. Then, simpler models, such as ridge, lasso, and gradient boosting, not only require less computational power, but, they also allow for quicker deployment in real-time monitoring systems [38] .

To interpret the forecast results, the mean values of the performance metrics for each model were calculated ( Table 11 ). This enabled their comparison and the identification of the best model. The performance metrics lie in their ability to quantitatively assess prediction model accuracy and inform stakeholders about prediction reliability. In groundwater quality scenarios, a model with low error rates (MSE) and high correlations is reliable, providing a clear view of resource quality variation over time [79] . This is essential for understanding the strengths and limitations of different models, allowing their recalibration, the integration of multiple approaches, or additional monitoring to improve overall prediction robustness. These error metrics extend beyond academic interest as RMSE values can be used to establish confidence intervals around predictions [80] , while correlation coefficients serve as the measure of the strength and direction of the model. The correlations also capture real-world relationships in the groundwater quality parameters [79] . In practical decision-making, performance metrics are crucial indicators for resource allocation, pollution prevention, and the establishment of informed remediation strategies to inform plans for future water demands [81] . Ultimately, they develop evidence-based policies to justify investments and create sustainable decision-support systems.

Our findings indicate that the GB regression has the lowest values for MSE and RMSE (MSE = 0.097; RMSE = 0.300). Conversely, both Lasso and Ridge regression display significant errors, with Lasso regression recording the highest MSE and RMSE figures (148.921; 6.828). These metric outcomes suggest that GB potentially outperforms Lasso and Ridge regarding predictive accuracy. Nonetheless, obtaining smaller RMSE and MSE does not necessarily indicate excellent models due to their exclusive representation of the mean of squared variances between the forecast and observation failing to reveal the inherent logical bias [28] . The computation of Pearson, Spearman, and Kendall correlation coefficients was performed to identify the variance in model performance and ascertain their consistency [64] . In agreement with the performance metrics, the correlation coefficients reveal that GB demonstrates the most significant level of uniformity while Lasso indicates the lowest, highlighting the differences in predicted and actual WQI values between Lasso and Ridge compared to GB. This outcome was supported by Chen and Guestrin [82] , who stated that GB encapsulates complex patterns in a way that other models are unable to. Discrepancies between the performance rates and the calculated correlations were also observed. For instance, in Lasso regression prediction, the metrics suggested an excellent performance in the ZE site (MSE = 0.245 and RMSE = 0.485), while the consistency evaluation presented weak consistency for the same site (r = 0.359; r_s = 0.022; τ = 0.021). Thus, the Lasso and Ridge prediction potentially involves systematic biases.

Other researchers have argued that components such as model fitting can lead to systematic instability and bias [28] . An effective data-splitting approach should support the models in grasping the statistical patterns of the input and prevent the introduction of excessive bias. Our study divided sample data into training and testing subsets rather than training, testing, and validation subsets due to the limited sample size (4 months) and the significance of avoiding underestimation of the observed groundwater quality. As a result, the conventional training and testing window selection technique was employed for data splitting [83] . Therefore, three-quarters of the sample size was selected as the training set (data before August 1, 2022) and the rest as the testing set (data after August 1, 2022). The test set included daily sampling data for over one month, representing a short period that can hinder the ability of the model to capture the overall dynamic trend. However, identifying changes in groundwater quality over four months of observation under the same seasonal conditions was a complex challenge. Hence, both Lasso and Ridge regression experienced limitations due to the small sample size and the absence of diversity in the data categories, which used solely physicochemical parameters. Furthermore, Lasso and Ridge models exhibit limited capabilities in forecasting nonlinear occurrences compared to GB regression [28] . This outcome implores the hydrogeochemical process, which embodies nonlinear relationships and a diverse array of interactions within the subsurface environment [84] [85] . By iteratively fitting new models to the errors of the previous models, GB can learn complex patterns and interactions between variables [58] [59] , implying the model’s suitability for our study.