wwwidonja/ml-showcase

A showcase of my own implementations of ML approaches. Created for the MLDS course @ DS Masters, FRI UL.

Machine Learning Implementations

This repository serves as a showcase of the machine learning approaches studied and implemented as part of the Machine Learning for Data Science course at the Data Science Masters' course at DF@FRI UL. All work is my own.

It is advised to look at individual .pdf reports in each folder for better understanding of the provided code.

Contents:

HW1: Trees

Own implementation of classification trees, bagging, and random forests on the housing dataset, with use of Gini Index for best split optimization. Reported misclassification rates for predefined hyperparameters.

HW2: Logistic Regression

Own implementation of multinomial and ordinal logistic regression with MLE. Proposal of dataset on which the ordinal approach works notably better. Interpretation of model coefficients on a practical application, with quantification of uncertainty.

HW3: Kernelized Ridge Regression

NOTE: The solution in this report contains faulty selection of columns (only first two) on the Housing Dataset. The study of the application of kernels: notably ridge regression. Application of an own implementation to both a toy and practical dataset.

HW4: Support Vector Regression

Own implementation of SVR, on toy and practical dataset., with two kernels. Comparison to results from HW3 and a comment on differences and similarities.

HW5: Loss estimation

Own implementation of holdout estimation and cross validation, with several demonstration on selection methods' effect on model risk estimation and split/test data variability.

HW6: Artificial Neural Network

Own implementation of a multi layer, fully connected ANN, with numerical verification of the gradient and application to a classification and regression dataset. Qualitative reasoning for performance differences on different hidden layer configurations is also provided.