Comparative Analysis of SVM and Perceptron Algorithms in Classiﬁcation of Work Programs

Government agencies are required to mobilize every aspect of publication, which is carried out every year and must be accounted for and also carried out for each device that receives it, such as assisted villages by utilizing available APBD funds in maximizing work programs designed so that they can be implemented optimally and effectively. Getting the best from all aspects of the work program implementation, of course, there are important points in designing an annual work program without exception. Data mining itself can help the department of population, family planning, women’s empowerment, and child protection in analyzing each work program design from before it is implemented onwards to look at various aspects of past data whose grouping is in the form of classiﬁcation. This study aimed to build a classiﬁcation model by adding a sigmoid activation function that used SVM and perceptron to obtain a comparison value for the accuracy of the algorithm used to obtain the best working program design. The classiﬁcation results were used to get the best value for classifying the best P2KBP3A work program dataset, where it can be seen that the average accuracy value was 87.5%, the f1 value was 82.2%, the precision value was 80.2%, and the recall value is 87.5% so that the ﬁnal result of the research results obtained a good accuracy value.


2.
RESEARCH METHOD The classification system for the architectural level of research is designed to look for the best work programs of case studies P2KBP3A, which is where the course of this system goes according to the procedures addressed in Figure 1.

Model Dataset
In conducting research, datasets are used in the form of a range from 2018 to 2021 in the sense of using data that has been grouped for three years, thus testing validation data from every three years to find out from the value of the determination of accuracy every year after testing based on the year of testing. From the dataset itself, it has more than 600 data on annual performance work program plans where the average annual performance plan range is 40% to 28%, which means that from one year, the work plan from P2KBP3A is around 180 to 260 work plan data made and must be carried out annually. This study applied the sigmoid activation function to the SVM and perceptron algorithms to classify the best work program at P2KBP3A Deli Serdang based on pre-determined categories. The results of the classification with the output of target achievement and realization of work programs with low, medium, and high categories are then analyzed using cross-validation with a value of K = 10 to obtain the accuracy, precision, recall, and fmeasure values for each algorithm. These values are then compared to determine which algorithm, between SVM and perceptron, the sigmoid activation function results in a better classification of the best working program problem.
This study compares the support vector machine and perceptron, which consists of the sigmoid function, to the best work program dataset in the P2KBP3A case study using the Orange 3.30 application, which uses different annual calculations from 2018 to 2021, which is carried out annually in the form of a model (as shown in Figure 2). This study also has details to identify the classification of the models used from SVM and perceptron to activate the sigmoid function with the caption Figure 3 SVM Model and Perceptron model.

Work Program Plan Data
Data from the work program plan from P2KBP3A, which is used as reference material to determine the feasibility of the work program, from the work program plan that is used as data taken with a range of years starting from 2018 to 2021 as test material. From each work program data, each year has a different range of work programs with the following specifications: 1. Data for 2018 with a total number of work program plans of as many as 262 implementers. 2. Data for 2019 with a total number of work program plans of as many as 189 implementers. 3. Data for 2020 with a total number of work program plans of as many as 174 implementers. 4. Data for 2021 with a total number of work program plans of as many as 161 implementers. The result is that each year has a reduced work program from the initial year, which has filtered the feasibility of the work program but does not demand the possibility of increasing each work program. Therefore, the features used in determining the criteria for the work program plan are 1 and 1 as the target. The following features are used, and the explanations (are shown in Table 1). The description of the features and targets in Table 1 used is as follows: 1. The number of assisted villages participating explains how many villages have participated in the implementation of the work program; the more assisted villages that participate in these activities, the better. 2. Activity duration explains how long it takes to carry out activities from the implementation of the work program, where the longer the implementation, the better. 3. Sources of budget funds, explaining the sources of funds available for the initial design of the work program to be implemented. 4. Budget explains the budget issued by the government for the implementation of work programs where the more budget spent, the better. 5. The infrastructure work program describes the location of the infrastructure available from the implementation of the work program. 6. The number of participants describes the participants who took part in activities from each village; the more participants, the better. 7. The number of sub-district activities for one month explains the number of work program activities that have been implemented for one month, and the more work programs implemented, the better. 8. Remaining budget, explaining the remaining budget after carrying out a work program where the less remaining budget, the better. 9. Performance achievement targets for one month, explaining the achievements of the work programs that have been implemented where the more they are implemented, the better. 10. Acceptance of implementation results explains the achievement value of what has been implemented in the work program where the value range is between 0 and 100. 11. Activity level describes the activity level of the city, district, sub-district, and village. 12. The remaining funds are indicative of the initial budget, explaining the difference between budgeted costs and the remaining budgeted funds; the smaller the difference, the better. 13. The year of work plan describes the range of years the work program has been implemented. 14. Target achievement describes the achievement target of the work program obtained, namely feasible or not feasible.

ISSN: 2476-9843
The following is a sample of the data for the work program plan carried out starting in 2018 (as seen in Table 2). Coordinating meetings and consultations outside and within the region 2 Provision of office administration services 3 Provision of Water Installation Equipment 4 Provision of Equipment and Work Equipment 5 Provision of office publication services 6 Development of Women's Organizations 7 Education and Training Activities to Increase Participation and Gender Equality 8 Counseling Activities for Housewives in Building a Prosperous Family 9 Business Management Guidance Activities for Women in Business Management 10 Exhibition of Women's Work in the Development Sector The following is a sample of the data for the work program plan carried out starting in 2019 (as seen in Table 3). Provision of Office Administration Services 5 Provision of work tools and equipment 6 Construction of the Office House 7 Office Building Construction 8 Procurement of Service/Operational Vehicles 9 Procurement of furniture 10 Routine / Periodic Maintenance of Office Buildings The following is a sample of the data for the work program plan carried out starting from 2020 in Table 4. The following is a sample of the data for the work program plan carried out starting from 2021 in Table 5. Explanations from Tables 2, 3, 4, and 5 are sample work program data that have been implemented in 2018, 2019, 2020, and 2021.
As for the split of the scale of each feature into a range of numbers starting from 1 5 on each type of feature for the extraction of the initial data value and also for the target used in the target name, the achievement is called the feasible and unfeasible extraction result where each target category This achievement is given a score for worthy of being given a value of 1 while it is not worthy of being given a value of 0, the following is an explanation of the results of the features and extraction targets (Show Table 6, Table 7,  Table 8, Table 9, Table 10, Table 11, Table 12, Table 13, Table 14, Table 15, Table 16, Table 17, and Table 18).        Explanations from Tables 6,7,8,9,10,11,12,13,14,15,16,17,and 18 are data extraction values from descriptions of features and targets from Table 1 by using work program data in Table 2, 3, 4 and 5 with work program data consisting of 2018, 2019, 2020 and 2021 that have been implemented.
The classification process with the SVM and perceptron algorithms uses the number of folds for cross-validation of 10 and the sigmoid activation function.

ISSN: 2476-9843
Regarding the evaluation of the results used, one of which is a confusion matrix which is obtained from the results of accuracy, precision, and recall as well as from the ROC curve to measure the AUC value. That way, the larger the area under the curve (AUC), the better the prediction results. The following is Table 19 of the confusion matrix [29].

RESULT AND ANALYSIS
The results obtained in testing the data from the total dataset used are 754 data from 2018, 2019, 2020, and 2021, with detailed data as much as the work program in 2018 there are 262 data, in 2019 there are 189 data, in 2020 there are 174 data. In 2021 there are 161 there, and overall the data for very good results by carrying out tests and a similar level on the results of work program contributions where the existing results for carrying out the results of the classification experiment between SVM and perceptron are very good and more dominant in SVM as at 89, 2% from [30] then there is 80.5% from [31] as well as from the perceptron with a value of 96.2%, with a value of 85.0% from [14].
The Each classification model from Figure 3 is inputted into the SVM and perceptron models in Figure 2 with a maximum number of epochs of 100. Accuracy, F1, precision, and recall values are obtained from the classification results using variations of the sigmoid function using 10-Fold Cross Validation (shown in Table 21).  Table 21 can be seen from the results that used the sigmoid activation function in the evaluation with 10-fold cross-validation resulting in the highest accuracy value of 88.8% in 2021 for the SVM model, the highest F1 value of 83.6% in 2021 for SVM the SVM model, the highest precision value is 85.9% in 2018 for the perceptron model, and the highest recall value is 88.8% in 2021 for the SVM model.
For the results of using the algorithm with the SVM model and perceptron, which uses the sigmoid function with the use of work program data that has been implemented, the best on average is obtained every year from 2018, 2019, 2020 up to 2021 for an accuracy value of 87.5%, for the F1 value of 82.2%, for the precision value of 78.7%, for the recall value of 87.5%. Thus, the results in Table 21 get good results for 2021 which are already in the work program plan and are also higher than the previous year's average results in 2018, 2019, and 2020.
The results that occur for classification can be seen in the comparison of research results from the differences in accuracy values (shown in The following are the results of the classification that has been done, which can be seen in the following confusion matrix display (Shown in Figure 4, Figure 5, Figure 6, and Figure 7).

CONCLUSION
From the evaluation results obtained in finding the best implementation of the work program from government agencies in P2KBP3A itself by using a comparison of the SVM and perceptron algorithms that use the sigmoid function and also the results of the first test for the dataset tested, it can be concluded from the results of the research stated that the SVM algorithm model is more dominant in the new P2KBP3A data collection from 2018 to 2021 which has the highest level of accuracy in the field of the model. On this occasion, in conducting research, the data used is original data and only tested for new cases. It is also hoped that further research from this study can test using other than k-fold 10 and also adopting other algorithm models or can also use function models other than sigmoid. For further research, it is recommended to use the model with other models or make comparisons with other model algorithms and also use cross-validation using different numbers such as K-5 or K-20. Then, the level of distribution of training data and test data can be used with a comparison of 70% and 30% or other because, in this study, 80% and 20% were used. With backtesting, you can also use the model by combining the existing activation functions or also using several models that are more than the existing ones.

ACKNOWLEDGEMENTS
Thank you to Mrs. Dr. Rika Rosnelly, S.Kom., M.Kom. and Mr. Dr. Zakarias Situmorang, MT., for finishing this scientific work to the best of his ability.

AUTHOR CONTIBUTION
The first author is me, Jaka Tirta Samudra, S.Kom. as the author for each data collection and data extraction on the dataset; the second author is Mrs. Dr. Rika Rosnelly, S.Kom., M.Kom. as analytical research on every aspect of existing data and also conducting research in research methodology. The third author is Dr. Zakarias Situmorang, MT. as a helper and correction of the results of each result of data mining object and also checks the results of the classification.

FUNDING STATEMENT
Statements regarding financing in conducting research are the personal property of authors 1, 2, and 3.

COMPETING INTEREST
I have no declaration under financial, general, and institutional competing interests.