This project demonstrates how a small Deep Neural Network (DNN) can be used to predict election results based on two factors: population size and delegate probability. The project is divided into two parts: training on a small extract of the prediction graph (less than 1000 samples) and extrapolation through the trained model to predict a specific threshold.
The model architecture used in this project is simple yet effective for extrapolation. It consists of the following layers:
-
Input Layer: The input layers are splitted to give the delegate probability more room in the DNN.
-
Hidden Layers: The three hidden layers consist of Rectified Linear Units.
-
Output Layer: The output layer produces the election result from 0 to 1 and therefore is a sigmoid layer.
A figure of the model layout is included at the end.
Install the necessary dependencies: Tensorflow; Keras; Pandas; Numpy
Due to the diminutive differences in the training data, the Hyperparameters have to be tuned carefully. The included ones + automatic adjustement have been found to fit all current known files.
Also the training runs differ a bit. On account of this, all training runs are repeated several times to be sure to get the correct result.
Enjoy predicting election results with the power of deep learning!
