Machine learning and artificial intelligence are crucial tools in the vast majority of different fields, but mainly in computer science and technology. Classifiers play a vital role in this field, especially in predicting the class membership of a sample under consideration. An example of the practical use of classifiers is the spam filter of email messages. The following paper aims to determine the most efficient classifier from selected: kNN, Soft set, and Naive Bayes on Iris database. Different versions of each of the classifiers have been considered. For kNN, the performance of various metrics was compared, for the Soft set, two approaches for establishing intervals during the classification, and for Naive Bayes, the normal and triangular distributions were compared. The most effective versions of the classifiers have been selected for the final comparison.
This is actually a small fraction of the immeasurable possibilities in this field of study. Machine learning models are used to learn the patterns in data. Machine learning algorithms can be used for example to gather information about data, split data for two parts and try to identify unknown sample as a data element. The methods which determine this, are called classifiers. W e have various types of classifiers applicable to diferent task. In machine learning models neural networks based ideas are very eficient in complex data analysis. In [4] was presented how to use them in low-dimensional data feature learning. The idea presented in [8] proposed neural network for analytical purposes of data recorded form high-speed train. There are also very eficient, however simple in construction, classifiers based on approaches sourced in Before we started working on our base, we made sure that our Iris base didn’t have null or NaN values. We have helped us to shufle, normalize and split database by dividing it into 70% as a training set and 30% as a validation set. As we can see, the setos class is significantly separated from the rest of the classes. This will result in a high proportion of correctly recognized objects for this class. For all classifiers we were normalizing a given database by taking all values from a specific column, determining of the lowest and highest value in the specific column and changing all values in the column according to the formula which is given below: [][] = [][] − − (1) a current row a current column a minimal value in the current column a maximal value in the current column
3.1. kNN
In order to function properly, the kNN algorithm needs functions that calculate the distance of the object for which we are looking for a class to the objects of classes already known to us. It is on the basis of this distance that the kNN algorithm decides to which class a given object may belong. There are many ways to calculate distances, each with its pros and cons. When calculating distances, we can, for example, use one of the known metrics, e.g. Euclid: || − ||2 = ∑︁( − )
2
(, ) = (∑︁ | − |) 1 =1
The kNN (k-Nearest Neighbors) classifier is one of the most important non-parametric classification methods. The kNN algorithm does not create an internal representation of the training data, but looks for a solution only when the testing pattern appears. It consists in assigning an object to a given class by checking to which representatives a given object has the shortest distance. The algorithm works as follows.
for a given sample, the distance to each object in the test set is calculated. Then list is created containing the given test file object and the distance to the sample which then is sorted from shortest distance to longest.
After that from this list, the k objects in the shortest test set were calculated. On their basis, the middle values distance from the sample are analyzed. At the end the of species range values were determined. sample is assigned to the class with the most objects. The classifier considers samples from the validation set together with the selected characteristic weight.
while < ( ) do while < ( ) do
distance of test object j to the sample and add the result to the list of distances. + +
Return from dictionary variety with the highest probability 3.2. Soft sets
There are a few ways in which soft set may be implemented, for example: including weight or not. In this case weight was included. Pearson correlation coeficients were calculated to properly choose the most appropriate weight values for particular characteristics.
(∑︀( − ¯)( − ¯) = √︀∑︀( − ¯)2√︀∑︀( − ¯)2 (4) characteristic value for i = 0,1,...,n mean value for particular characteristic value of compared characteristic for i = 0,1,...,n mean value for compared characteristic
lfing, normalizing, and splitting the database into a test set and validation set in the ratio of 70 to 30.
← ℎ for < ( ) do
and maximal values for each iris type in test set. Creates nested list with iris type name, centre value for each iris type based on minimal and maximal values. ← ← ← 0 ← 0 for do ◁ Add all trait values for sample + = * ℎ[] + = 1 for do ◁ Calculate distances = [0] = | − | Chooses minimal distance and corresponding iris type. Returns classified type.
model that’s used for classification task. At the beginning it reduces database by splitting an Iris database to three smaller databases according to their variety. After that classifier assigns the initial probability of a given species appearing in the database. Next, it takes a sample and counts a probability for each reduced database. It uses one of two considered distribution formulas. Subsequently, it multiplies the initial probability with all partial probabilities (with all attributes that a reduced database has). And at the end it compares which probability of three possible is the highest.
Normal distribution: (| ) = √ 1 22 (− ( − )2 22 ) ⎧0, ⎪⎪⎪⎨ 6−2 + √16 , − √6 ≤ ≤ ⎪⎪− 6−2 + √16 , ≤ ≤ + √6) ⎩⎪0 > + √6) < − √6 a current attribute in reduced database a current attribute of a sample a standard deviation of an attribute a mean of an attribute in reduced database current reduced variety database
4 ( ) = () * ∏︁ (| ) =1
databases according to their varieties.
counting the number of elements in reduced database divided by the number of all elements in the main database. Next, for a given sample it calculates a partial probability for each attributes in each reduced database. 4.1. kNN + + + + (8)
very similar to each other. This may be due to the fact 4. Experiments that the Setos class is significantly distant from the other two classes, which means that there is a very high probIn order to properly analyze individual classifiers, we ability that the objects closest to the sample will also be will perform a series of tests that will allow us to select Stetosa. Below are the results for k equal to 1 2 3 and the best classifier using the confusion matrix. Confu- 4, respectively. sion matrix is used in assessing the quality of a binary classification. It describes how well the classifier classi- Table 1 fied given samples. It also gives us information about Table for k = 1 several things about the classifier such as:
k 1 4.2. Naive Bayes
4.3. Soft sets
for each characteristic of iris flowers allowed choosing the weight of the features for the optimal classifier accuracy. An association between individual features was considered and their influence on the classifier eficiency.
Considering obtained correlation values, it was concluded that the most important characteristics are the following in descending order: petal-length, petal-width, sepal-length, and sepal-width. According to these observations, successively assigning diferent weight values, the best results were observed with weight = [0.1, 0, 0.5, 0.4].
results were obtained for individual types of iris flowers.
pared implementations were compared on the same partition of the Iris database. In comparison, the following results were statistically calculated.
Through analysis of attained results for all classes of all classifiers, it can be noted that the level of accuracy is the highest for the soft set classifier. T he v alues of [1] Akram, M., Ali, G., Butt, M. A., Alcantud, J. C. R. other statistically obtained characteristics also reach the (2021). Novel MCGDM analysis under m-polar fuzzy highest levels for the previously mentioned classifier. It soft expert sets. Neural Computing and Applications, is worth mentioning that the obtained results are similar 33(18), 12051-12071. for particular characteristics of the kNN and Naive Bayes [2] Chen, W., Zhou, Y., Zhou, E., Xiang, Z., Zhou, W., Lu, classifiers. J. (2021). Wildfire risk assessment of transmission-line
Based on the obtained results, it can be concluded corridors based on Naïve Bayes network and remote that the soft set classifier implementation classifies most sensing data. Sensors, 21(2), 634. efectively. All of the implemented classifiers have been [3] Dong, W., Wozniak, M., Wu, J., Li, W., Bai, Z. (2022). properly implemented. The results of the best classifier De-Noising Aggregation of Graph Neural Networks difer only by a few percentage points from each other. by Using Principal Component Analysis. IEEE Trans
The work and efort that was applied to completing actions on Industrial Informatics. this article are practical and applicable. This research [4] Dong, W., Wu, J., Bai, Z., Hu, Y., Li, W., Qiao, W., ofered an opportunity to learn and expand knowledge Woźniak, M. (2021). MobileGCN applied to lowabout the diferent approaches to assessing and teaching dimensional node feature learning. Pattern Recognichosen classifiers as well as through the process of iden- tion, 112, 107788. tifying the best solution. The analysis allowed acquiring [5] Rani, P., Verma, S., Kaur, N., Wozniak, M., Shafi, J., practical experience in implementing machine learning Ijaz, M. F. (2021). Robust and secure data transmisalgorithms. sion using artificial intelligence techniques in ad-hoc
In the future, the project could be extended and fol- networks. Sensors, 22(1), 251. lowed with further analysis of other classifiers, for in- [6] Ruan, S., Chen, B., Song, K., Li, H. (2022). Weighted stance through rebuilding the current classifiers in a more Naïve Bayes text classification algorithm based on imadvanced way and selecting even more eficient solutions. proved distance correlation coefficient. Neural Computing and Applications, 34(4), 2729-2738.
Table 28 [7] Shokrzade, A., Ramezani, M., Tab, F. A., Mohammad, Results M. A. (2021). A novel extreme learning machine based kNN classification method for dealing with big data.
Classifier Expert Systems with Applications, 183, 115293. [8] Siłka, J., Wieczorek, M., oWzniak, M. (2022).Recurrent neural network model for high-speed train vibration prediction from time series. Neural ComputkNN ing and Applications, 1-14.
[9] https://c3.ai/glossary/data-science/classifier/ [10] https://www.sas.com/en_th/insights/articles/bigdata/artificial-intelligence-machine-learning-deeplearning-and-beyond.html [11] https://www.sciencedirect.com/topics/computer
science/machine-learning [12] https://www.sciencedirect.com/science/article/ Naive pii/S0898122199000565 Bayes [13] https://towardsdatascience.com/machinelearning-classifiers-a5cc4e1b0623 [14] https://monkeylearn.com/blog/what-is-aclassifier/
Setosa 0.96 0.87 1.00 0.93 1.00 13 0 30 2 0.96 0.87 1.00 0.93 1.00 13 0 30 2 0.98 0.93 1.00 0.96 1.00 13 0 31 1