Simple project using dimensionality reduction and machine learning to classify handwritten digit images.
The MNIST dataset is a publicly available dataset that is very often used for simple classification algorithms. This is a short project that loads in the MNIST dataset, uses Principle Component Analysis (PCA) to reduce the dimensionality of the data, and determines a simple machine learning technique for modeling of the data. The techniques that are applied in this project are: KNN and Random Forest.
In order to run the script MNIST_main.R, first the 4 data files necessary for this project need to be downloaded and stored in an approriate folder. The file data_README.txt contains the instructions to do so.
Once the data is stored under the "data" folder, then the working directory must be set in the script MNIST_main.R. This is to ensure that the file will look in the correct location for the data necessary for the script to run.
Principle component analysis (PCA) was performed on the MNIST dataset. Results suggest that ~2% of the features can accurately represent 98% of the variance.
PCA representation of the images:
Random Forest number of trees, accuracy and error by class:
