Classification of Salt Quality Based on the Content of Several Elements in the Salt Using Machine Learning

This study discusses the classification of salt quality using numerical data. Classification is one of the scientific disciplines in data mining which involves the process of extracting, identifying, and analyzing various information with the aim of discovering patterns within data using mathematical, statistical, artificial intelligence, and machine learning approaches [13, 14]. Data mining is known as the discovery of knowledge in databases. several types of data mining based on their function, including description, prediction, estimation, classification, clustering, and association [15]. Figure 1 shows the steps in data mining.

11.png

Figure 1. Data mining

Here is an explanation of the steps in the data mining process:

(1) Data Selection: remove irrelevant and inconsistent data.

(2) Data Cleaning: Involves merging and adding data that is relevant.

(3) Data Select: This involves choosing data to be used as the basis for analysis.

(4) Data Transformation: Before the mining process, the data is converted into a certain format.

(5) Data Mining: Involves searching for information using data mining methods.

(6) Evaluation: Involves identifying patterns from data obtained by the method and then evaluating the results and the hypothesis.

2.1 Data

The quality of salt is categorized into several classes based on its intended use. In all cases, it is important to understand the purpose of using salt and ensure that the salt used meets the required standards to maintain human health and product quality. In categorizing salt into classes K1, K2, K3, and K4, there are several features used in this study, which will be explained as follows:

(1) Water Content

Water content is one of the features used in this study to classify the quality of salt. This feature is employed because the moisture content in salt significantly impacts its quality. If the moisture content in salt is too high, the salt's durability will decrease [16].

(2) Not Dissolved

Salt is a compound made up of sodium, and almost all of it is soluble in water, including salt itself. Because of this feature, it is necessary to use it for classifying the quality of salt, as higher-quality salt tends to have a higher solubility level. Salt that readily dissolves is highly recommended for consumption as it is beneficial for health [17].

(3) Calcium

Calcium is one of the elements present in salt, thus the concentration of calcium also affects the quality of salt [18]. Salt that is suitable for consumption should have a maximum calcium content of 0.06% [17].

(4) Magnesium

Magnesium is also one of the elements present in salt, so the concentration of magnesium affects the quality of salt [18]. Salt that is suitable for consumption should have a maximum magnesium content of 0.06% [17].

(5) Sulfate

Sulfate is a compound that can decrease the NaCl content in salt, whereas good quality salt should have a minimum NaCl content of 97%. Therefore, if the sulfate content is detected to be high, it can degrade or lower the quality classification of the salt [17].

(6) NaCl (wb) and (db)

In this study, the features used are NaCl wet basis and dry basis to categorize salt data into classes based on their quality. Salt that meets the food grade standard must have a minimum NaCl content of 97% [17].

2.2 Data pre-processing

Data preprocessing is the initial step before entering the model training phase, aimed at organizing the input dataset into a structured format to facilitate the training process [19]. In this study, the data preprocessing stage involves data transformation to adapt the data format according to the requirements.

2.2.1 Data transformation

Data transformation involves modifying the scale of data into a different format to achieve the desired data distribution. Each data point undergoes similar mathematical operations as its original form [20]. One method of data transformation is data normalization, which aims to adjust several variables to have a uniform range of values to prevent overly large or small values that might affect analysis outcomes. The primary objective of altering all data is to maintain the relative differences between data points. If multiple variables are present, the transformation is applied to all variables to preserve the relationships between data points [21]. The process of data normalization is implemented using the Min-Max normalization technique, defined by Eq. (1) below:

$z=\frac{x-\min (x)}{\lceil\max (x)-\min (x)\rceil}$        (1)

where, z: normalization result, x: x value, min(x): minimal value of x, max(x): maximal value of x.

2.3 Data mining

Within this research, the data mining procedure includes dividing the data into training and testing datasets. The training dataset is used to train the model, whereas the testing dataset is employed for validation and assessment to gauge the effectiveness of the utilized techniques. The separation of the dataset is achieved through the implementation of the K-Fold Cross Validation approach, where the parameter "k" determines the dataset divisions. For instance, when using 5-Fold validation, the dataset is divided into four subsets for training and one subset for testing, with nearly equal sample sizes in each fold [22]. Subsequently, the classification process employs three different algorithms to construct a classification model capable of categorizing the quality of salt. These algorithms include K-Nearest Neighbor (KNN), Support Vector Machine and Naïve Bayes.

The advantage of the KNN method is that apart from being easy to implement and adapt, this method has few hyperparameters. For SVM, the advantage is that this method has two free parameters called upper bound and kernel parameters. SVM also produces unique and optimal solutions, and can implement the principle of structural risk minimization (SRM) which is known to have good general performance. Meanwhile, Naive Bayes has the advantage that this method is simple, fast and has high accuracy.

2.3.1 K-Nearest Neighbor (KNN)

K-Nearest Neighbor is an example of instance-based learning and is often used for classification tasks, where its objective is to classify unseen data based on the stored database. The new data point is classified based on its similarity to other data points stored in the model using various similarity metrics. The algorithm determines the class of the new point by selecting the K nearest points, also known as K neighbors, to the new data and choosing the most common class among the group of data points through majority vote as the class of the new point [23]. Several steps required for the implementation of this method can be outlined as follows:

(1) k values that have nearest neighbors must be determined at the beginning.

(2) Calculate the squared distance between the object point and the training data. In this study, the distance between object points is calculated using the Euclidean Distance method. Euclidean distance is a method of searching between two variable points, the closer and similar they are, the smaller the distance between the two points. Which is formulated in the Eq. (2):

$\operatorname{dist}(x, y)=\sqrt{\sum_{i=1}^n(X i-Y i)^2}$            (2)

where, d(x, y): Euclidian distance, X: Data 1, Y: Data 2, i: Attribute I, n: Number of attributes.

(1) After that, the results from the second step are sorted from the highest value to the lowest.

(2) Collect categories from the neighboring data based on the value of k.

(3) The final step, determines the majority category of nearest neighbors to be used to predict new data objects.

The K-NN method is often used for classifying data due to its simple and straightforward implementation, quick training process, and its applicability to data with noise. However, this method also has its drawbacks. It falls under the category of lazy algorithms, which can lead to slightly longer program execution times. It is highly sensitive when dealing with cases involving irrelevant features and requires memory storage for storing the training data records used in the process.

2.3.2 Naïve Bayes

Naïve Bayes is a simple probabilistic classification method that calculates various probability ranges by summing occurrences and combined values from the provided dataset. This algorithm employs Bayes’ theorem and assumes that all attributes are independent or non-interdependent based on the value of the class variable. An alternative description states that Naïve Bayes is a classification method that utilizes probability techniques and statistical insights developed by the British scientist, Thomas Bayes. It predicts future prospects using previous experiences as a reference foundation [10, 24].

This algorithm is a probability technique used to categorize classes in a given dataset. The basic outline of the Naïve Bayes method involves statistical analysis where initial probabilities (prior probabilities) are estimated from training data. The probabilities for each parameter are then calculated based on these initial probabilities. Its main characteristic is the strong assumption of the independence of certain phenomena or conditions. In the context of Bayes' theorem, when there are two distinct events, denoted as X and Y, Bayes' theorem can be expressed through Eq. (3) [24]:

$P(Z \mid Y)=\frac{P(Z)}{P(Y)} P(Y \mid Z)$            (3)

where, Z: New data record, Y: Hypothesis, P(Y|Z): Probability of Y toward Z hypothesis, P(Y): Probability hypothesis Y, P(Z|Y): Probability of Z based on the condition of Y, P(Z): Probability of Z.

The selection of this method is relatively straightforward as it doesn't involve matrix multiplication or numerical optimization. This method is more efficient when used for predicting large amounts of data and provides a relatively high level of accuracy in its prediction outcomes. However, this method cannot be applied to case studies that involve conditional probabilities with a value of zero, as the predicted probabilities will also be zero.

2.3.3 Support vector machine

The method of Support Vector Machine (SVM), introduced by Boser, Guyon, and Vapnik in 1992 at the Annual Workshop on Computational Learning Theory, serves as a machine learning technique applicable for both classification and prediction purposes. SVM's core concept in classification involves identifying an optimal separator, known as a hyperplane. The hyperplane is deemed optimal when it offers the largest margin, representing twice the distance between the hyperplane and support vectors. These vectors refer to the nearest data points to the hyperplane. SVM excels in managing high-dimensional data and limited training samples due to its adherence to the Structural Risk Minimization (SRM) principle, which maximizes margin and minimizes expected risk in the face of uncertainty [25].

2.png

Figure 2. Support vector machine

Figure 2 illustrates 2 classes, each marked with distinct patterns. In Figure 2, the two classes are isolated by a ran red line known as the hyperplane. In this algorithm, the hyperplane is changed in accordance with be ideal for isolating the two distinct classes. The best optimal the resulting hyperplane is, the lower the error rate can be in the classification system.

In addition to handling linear data issues, SVM is also capable of addressing problems with data that cannot be linearly separated, also known as non-linear data. Non-linear problems can be overcome by utilizing kernels in a higher-dimensional workspace [26, 27]. There are various variations within the SVM method, including:

(1) Kernel Linier

Linear kernel functions are used for linear data classification. Linear kernel is the simplest kernel function. Linear kernels are used when the data being analysed is linearly separated.

$K\left(x_i, x\right)=x_i^T x$        (4)

(2) Kernel Polynomial

The polynomial kernel function is a kernel function that is used when the data is not linearly separated.

Polynomial kernels are a more general form of linear kernels. In machine learning, a polynomial kernel is a kernel function suitable for use in SVMs and other kernelizations, where the kernel represents the similarity of training sample vectors in feature space. Polynomial kernels are also suitable for solving classification problems on normalized training datasets.

$K\left(x_i, x\right)=\left(\gamma x_i^T+r\right)^p, \gamma>0$          (5)

(3) Kernel Radial Basis Function

The Radial Basis Function (RBF) kernel function is used for non-linear data classification. The RBF kernel or also called the Gaussian kernel is the kernel concept that is most widely used to solve data classification problems that cannot be separated linearly. This kernel is known to have good performance with certain parameters, and the results of training have a small error value compared to other kernels.

$K\left(x_i, x\right)=\exp \left(-\gamma\left\|x-x_i\right\|^2\right)$          (6)

(4) Sigmoid Kernel

$K\left(x_i, x\right)=\tanh \left(\gamma x_i^T+\mathrm{r}\right)$          (7)

The following are the steps in carrying out classification using the SVM method, among others:

(1) The initial stage involves computing the Hessian matrix, which results from the multiplication of the kernel function by the values of $y$. The value of $y$ corresponds to the vector value, namely the values 1 and -1 . Calculation of the Hessian matrix using Eq. (8).

$D_{i j}=y_{i,} y_j\left(K\left(x_i, x_j\right)+\lambda^2\right)$             (8)

where,

_ij: The component of the Hessian matrix (i is row and j is column).

λ: The theoretical limits that will be derived.

y_I: The class of i data.

y_j: The class of j data.

(2) Next, the second phase entails evaluating the error value utilizing Eq. (9), computing delta alpha through Eq. (10), and ascertaining the updated alpha using Eq. (11), outlined in the subsequent manner.

$E_i=\sum_{i=1}^l a_i D_{i j}$             (9)

$\delta a_i=\min \left\{\max \left[\gamma\left(1-E_I\right),-a_i\right), C-a_i\right\}$             (10)

$a_i=a_i \delta a_i$             (11)

where,

$E_I$: error score

$a_i$: alpha

$\delta a_i$: delta alpha

C: Constanta

$b=-\frac{1}{2}\left(w \cdot x^{+}+w \cdot x^{-}\right)$             (12)

(4) The step four, calculating the dot product on training and testing data.

(5) The final step is to determine the class of the test data using the equation below:

$f(x)=\sum_{i=1}^l a_i y_i K\left(x_i, x\right)+b$             (13)

The SVM method is often used for in classifying data in previous studies, mainly due to its excellent accuracy and relatively easy training process. This method also adapts well to high-dimensional datasets. The balance between model complexity and error can be managed easily, and it can handle both continuous and categorical data. However, a notable drawback of this method is its difficulty in interpretation unless the features are easily understandable. Additionally, there's a lack of result transparency due to its non-parametric methodology.

2.4 Evaluation

To measure the accuracy level and performance outcomes of the algorithm, various methods can be employed. In this research, the Confusion Matrix method was used as a process for evaluation. Confusion Matrix is an evaluation method which calculates precision, accuracy, recall and F-measure of algorithm predictions based on test data [28]. Table 1 is a Confusion Matrix used as a reference for calculating accuracy, precision, recall, F1 Score and AUC values.

Table 1. Confusion matrix

Based on the confusion matrix Table 1, the values of accuracy, precision, recall, F1 Score, and AUC can be calculated using the equations below:

(1) Accuracy

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$            (14)

(2) Precision

Precision $=\frac{T P}{T P+E P} \times 100 \%$            (15)

(3) Recall

Recall $=\frac{T N}{T N+F N} \times 100 \%$            (16)

(4) F1 Score

$F-$ measure $=2 \times \frac{\text { Recall } x \text { Precision }}{\text { Recall }+ \text { Precision }} \times 100 \%$            (17)

(5) AUC

$A U C=\frac{1}{2} x\left(\frac{T P}{T P+F N}+\frac{T N}{T N+F P}\right)$            (18)

where, TP: Total number of true positive predictions, TN: Total number of data instances with actual positive class but predicted as negative, FN: Total number of true negative predictions, FP: Total number of data instances with actual negative class but predicted as positive.

4.1 K-Nearest Neighbor (K-NN)

At this stage the salt dataset is classified using the KNN method with k=5. In this classification process, 3 scenario trials are used, namely on folds 5, 10 and 20. The results obtained from the application of this method can be seen in the following Table 3.

Table 3. Evaluation result of K-NN

In Table 3, it is shown that the AUC value is highest at k=10 and k=20, with a value of 99.0%. Furthermore, the best evaluation in general results were obtained when k=10, with an accuracy of 91.7%, an F1 Score of 91.6%, precision of 91.9%, and recall of 91.7%.

4.2 Support vector machine

By applying the Support Vector Machine method to classify salt quality, good results were obtained as seen in Table 4. These evaluation results were achieved when using the parameters indicated in Table 4 and involving dataset splitting through k-fold values, namely 5, 10, and 20.

Table 4. Parameter SVM

Table 5. Evaluation result of SVM

From the three tests with varying k-fold values, the best performance was achieved when k=20. While the highest AUC value was observed at k=10, the classification accuracy, F1 Score, precision, and recall were better when k=20, as shown in Table 5. Specifically, this configuration resulted in an AUC value of 87.7%, an accuracy of 71.7%, an F1 Score of 71.2%, a precision of 72.6%, and a recall of 71.7%.

4.3 Naïve Bayes

The results obtained in this study, using the salt dataset and the Naïve Bayes algorithm, involved dataset division through k-folds values, namely 5, 10, and 20.

Among the three tests with different k-fold values, the best performance was achieved when k=10, as indicated in Table 6. This resulted in an AUC value of 78.55%, an accuracy of 55.7%, an F1 Score of 56.0%, a precision of 56.9%, and a recall of 55.7%.

Table 6. Evaluation result of Naive Bayes

The results of the tests that have been carried out are comparing the three methods, which one is more accurate. The comparison of the performance of each algorithm model can be seen in Table 7.

Table 7. Comparison evaluation results of AUC values for the KNN, SVM and Naïve Bayes methods with fold 10

It can be seen that the KNN algorithm is superior to the SVM and Naïve Bayes algorithms for multivariate data types with an accuracy value of 96%. The K-NN method is very good in predictions, but when used on the multivariate data type in this study. This shows that the classification and prediction algorithm models are getting better too. Based on the results of the accuracy, precision, recall and F1-Score values, it can be concluded that the KNN algorithm has better performance than the SVM and Naive Bayes algorithms in classifying salt quality.