Learning Vector Quantization 3 (LVQ3) Usage To Determine Recipients of the Family Hope Program (Case Study: Tanjung Lubuk District)

The problem of poverty is a dilemma that the Government must solve. One of the Government's programs is the welfare program for the Family Hope Program (PKH). Tanjung Lubuk district, implementing the Family Hope Program experienced several obstacles in identifying PKH recipients, one of which was selection, limited, and close to officers so that it could lead to the provision of PKH assistance on target. Another problem is that the recipients of the data used are still using old data that has not been updated regularly, so many people who deserve assistance do not receive assistance. The research variables used were 35 variables. The output categories were entitled to receive and not entitled to receive PKH. The research method uses Learning Vector Quantization (LVQ) 3. The data are from 654 low-income families in Tanjung Lubuk District. The data used are 90:10 for practice data and 80:10 for test data. The learning rate values are 0.1, 0.3, 0.5, 0.7, and 0.9, while the learning rate reduction is 0.1, the minimum learning rate is 0.01, the window is 0.1, 0.5, and the m value is 0.1, 0.5. The accuracy obtained is 94.4%.


INTRODUCTION
The Family Hope Program or known as Program Keluarga Harapan (PKH) is one of the social protection in Indonesia in aid format. According to the Statistic Centre Bureau (BPS). The definition of poverty is the person or someone who cannot afford or cannot fulfill their basic needs, such as food and clothes, measured by their monthly expenses. These limited expenses are called Poverty Line (PL/GK) [1]. PKH aid has also become a work program in Tanjung Lubuk District. However, several obstacles are still found, especially in identifying and determining a family who has entitled to receive the aid, besides that there are still susceptibilities, such as subjective addressee and relative personal closeness with PKH officers, which can make the PKH aid distribution are misleading, also the non-update periodic data are still implemented. The effect is that many entitled recipients do not get aid distribution.

M. Deni Alfani, Ilman Zuhri Yadi | 1031
Another problem is that in some implantation, some district officers are still manually surveying data manually and registering every requirement indicator of the PKH agent by visiting every head of the family. Data also can be manipulated by an unauthorized person. So it requires a system that can classify the PKH distribution correctly and quickly. Based on the problem above, there are some research can be related to the PKH aid determination among others research ( [2], [3], [4], [5], [6], [7] in this research, the PKH determination are still using manual implementation which the officers surveying family then registering criteria, the results get 0,70 of learning rate grade, 2 epoch grade, 0,01 Min Alpha grade, 0,03 Dec Alpha grade, related research with PKH determination are implemented by [8] have found a problem with is the PKH distribution are misspointed. Researcher determined the recipient by using C4,5 algorithm. Research [9], [10], [11], [12] use the LVQ algorithm research to determine the classify of a low-income family, examining problems to determine low-income family, determining categories are: type 1 very low-income family, type 2 lowincome family include, type 3 nearly low-income family include. LVQ is the algorithm used to classify and duplicate determined input-output. This research uses 70 data sheet, 10 neuron input, 3 neuron output, 100 MaxEpoh and 0,05 learning rate (α) in prediction. The testing is conducted 5 times so that the accuracy level and error rate shall be directly compared to the amount determined of training data and testing data. The next research [13] is where the research objects are a set of data from heads of families from Mlandingan's district, Situbondo. The collected data contains 7 poverty parameters such as age, number of family members, income, expenses, house condition, house status, and recent education. This research used 5 testing scenario, which delivers grade 0,1 for learning rate recommendation, 0,1 for learning rate subtraction, 30% for practice data, 0,01 MinAlpha, and 2 iterate maximum, so that can get 98% accuracy. Other researchers who use LVQ3 to classify the research from (Jasril, 2018) [6], this research use learning rate grades of 0,0001, 0,01, 0,1, 0,4, 0,7, 0,9, and 0,0001, 0,4, 0,7 window grade. Tested and for practice, used data are 90:10%. Maximum epoch uses are 1000 iterations. Based on the test result, the highest accuracy is 91,67%.
Based on research [14], the vector quantization method (LVQ) 3 was chosen because this method has the advantage of being able to find the closest distance. Besides that, during learning, the output units are positioned by adjusting and updating the weights through supervised learning to estimate classification decisions. In LVQ 3, 2 vectors are updated if some conditions are met. Developing a decision support system process to determine the social assistance program for the Family Hope Program (PKH) in Tanjung Lubuk District uses LVQ 3; namely, if the input has the same estimated distance as the winner and runner-up vectors, then each Vector must learn. The variables used in this study were 35 data, namely: the number of family members, building/house status, land status, floor of house area, house floor type, house wall type, house roof type, number of house rooms, drinking/fresh water source, cooking fuel, toiletry, septic tank, wall of toilet room condition, the roof of toilet room condition, light source, power source, closet, another house, any existing 5 Kg/more cooking gas, existing refrigerator, air conditioner, heater, static phone, TV, and jewelry/savings, existing laptop/computers, bicycle, motorcycle, car, boat, yacht, static asset and business status from all family members, amount of collected data are 564 data.

Data Research
The numbers of data used are 654, then processed based on requirement so that it can be used as tested data and for practice data to facilitate and simplify the author/writer to design a PKH recipients determination system.

RESULTS AND DISCUSSION
The data analysis process and input variable can be seen in Table 1. Determination targeting/classification used in the LVQ method has already been determined first. Targeting/classification used to specify the PKH recipients are shown in Table 2. An architectural drawing of an LVQ3 artificial neural network system that will achieve based on inputed variable data and the grade who wants to reach to classify PKH recipients determination is shown in Figure 2. Variable normalization or inputted data are conducted to get the smallest data value between 0 to 1, which represents the exact value, so it can not relieve the value of accurate data. It has been explained that the use of LVQ3 depends on the range between the input vector with the quality of each grade and inputted vector GP into the grade with the closest range. Hence, to get recognized by the LVQ network, the existing data in the inputed variable data should be first modified into numeric form. The Diagram LVQ3 learning process can be seen in Figure 3.  (2) w2,t+1 = w2,t+ αt (x-w2,t) (3) 7. If the search result uses the window (ε) equation to have a false value, the conducted formula is (Lee, 1993): w1,t+1 = w1,t-β(t) (x-w1,t) (4) w2,t+1 = w2,t+β(t) (x-w2,t) (5) 8. Learning score β(t) is the multiplication from mα(t), equation mα(t)=εα(t).
The multiplication value is to gain the grade update while the window has a false value (m) (Fausett, 1994) between 0,1 and 0,5, the equation: β(t) = mα(t), dengan 0,1 < m < 0,5 After the learning process, they will get the final grade (W). This grade value will be used in the testing process. The design of the research application for the accustomed data menu and testing data of PKH recipient candidates are shown in Figures 4 and 5.