ANN: Prediction of Per Capita Income Rural Community on Poverty Line Based on Province

The problem of poverty is a fundamental problem that is of concern to every country. The Indonesian state has established a poverty reduction program as the main program. Poverty occurs in urban and rural communities. This research raises the problem of poverty in rural communities. The contribution of research to the government is to predict the per capita income of rural communities according to the poverty line based on the future provinces. The data used is data from the National Statistics Agency. These data are 2015 semester 1 data up to 2018 semester 1. The algorithm for its completion uses the artificial neural network backpropagation method. Input data is 2015 data for the 1 to 2017 semester 2. The training and testing architecture model is 4, namely 6-2-1, 6-3-1, 6-2-3-1, and 6-3-2-1. Target data is 2018 semester 1 data. The best architectural model is 6-2-1 with 79 epoch, MSE 0,004801 and 100% accuracy rate. From this model, a prediction of rural income per capita in the poverty line is based on the provinces of each province in Indonesia.


Introduction
Poverty is a fundamental problem that is of concern to the Indonesian government. Every new government always brings the idea of change in the economic field. The government needs to predict rural income per capita in the poverty line based on the province as a material consideration to take policy on poverty alleviation programs in rural areas. The method used in this study is an artificial neural network (ANN) [1] with a backpropagation algorithm [2], [3]. Backpropogation algorithms are able to generalize and extract from a certain pattern and create new patterns through the process of learning ability (self organizing) [4]- [6]. In addition, ANN can also process input data without having to have a target [3] [7]. Based on these advantages, this technique is able to predict rural income per capita in the poverty line by province. The results of this study provide input for the government to take policies in poverty alleviation programs in rural areas by making activities that support increased rural income so that income in rural areas continues to increase.

Methodology
Data on rural per capita income in the poverty line based on the province will be processed by the Artificial Neural Network with the backpropogation method. Data is represented in numerical form between 0 and 1 because the network uses the binary sigmoid activation function (logsig) whose range is from 0 to 1. The values used are obtained based on the categories of each variable as well as to make it easier to remember in defining it. Input data is per capita income data of rural communities in the 2015 poverty line semester 1 to 2017 semester 2. Input data is obtained from the National Statistics Agency's website on the per capita income of rural communities on the poverty line by province. This data will be transformed into a data between 0 to 1 before training and testing using artificial neural network backpropagation method with the formula: The target data is data on rural per capita income in the poverty line based on the province in 2018 Semester 1.

Artificial Intelligence
Artificial intelligence [8] is a branch of computer science in representing knowledge through symbols and processing information based on a number of rules [9]- [11] [12] [13]. The process of knowledge gained through education and experience organized with each other that can be applied to decision making and problem solving [14] [15] [16].

Backpropogation Neural Network
The Backpropagation method [17] [18] is one of the supervised learning methods that minimizes errors in the output generated by the network originating from the network against the network weights that can be modified [19]- [21]. This error is named Error gradient which is used to find the weight value that will minimize error [22][23].

Data processing
Data processing is carried out with the help of Matlab 6.1 software applications. The sample data is the per capita income of rural people on the poverty line by province. This data will be used in training data and test data. Samples of data that have been processed and transformed are as follows.

Backpropagation Architecture Design
The network used to predict rural per capita income in the poverty line based on the province by backpropogation with feedforward learning steps. The parameters in the formation of a backpropagation network use 6 input variables, 1 or more hidden layers and 1 output layer. The architectural model used to get the best architecture is 6-2-1, 6-3-1, 6-2-3-1 and 6-3-2-1. The neural network that will be built is a backpropagation algorithm with the Sigmoid activation function. The activation function in the artificial neural network is used to calculate the actual value of the output in the hidden layer and calculate the actual value of the output in the output layer.

Defining Output
The category for output is determined by the minimum error level of the target. The categories are listed in the following table:

Design of Artificial Neural Network architecture
The following stages will be performed in the user of the back propagation algorithm with the sigmoid activation function. The steps that must be done are as follows: 1. Initialization, is the stage where the value variables will be set or defined in advance, for example: input data value, weight, expected output value, learning rate and other data values. 2. Activation, is the process of calculating the actual value of the output in the hidden layer and calculating the actual output value in the output layer 3. Weight Training, is the process of calculating the gradient error value in the output layer and calculating the value of the gradient error in the hidden layer. 4. 4. Iteration, is the final stage in testing, where if there is still a minimum error that is expected not to be found then return to the activation stage.

Selection of the best architectural neural network
The results of the Matlab 6.1 application software used for the 6-2-1 architectural model, 6-3-1 architecture, 6-2-3-1 architecture and 6-3-2-1 architecture are obtaining the best architectural patterns. From this pattern, it will be used to predict rural income per capita in the poverty line by province. The best assessment of architectural models is seen from several aspects such as epoch, minimum error and accuracy of truth. For more details can be seen in the following: From table 3 it can be seen that the best architectural model that will be used to make predictions from a series of model trials is 6-2-1 with an 79 epoch, MSE 0,004801 and 100% accuracy rate.

Conclusion
Based on the results and discussion above, the writer can draw the following conclusions: a) The Artificial Neural Network model used is 6-2-1 architecture, 6-3-1 model, 6-2-3-1 model and 6-3-2-1 model, can get the best results by looking at the smallest MSE Testing is 6-2-1.
b) With a 6-2-1 architectural model, it can predict urban per capita income at the poverty line by province by showing 100% performance.