This repository contains a few machine learning projects, that do not have a deep-learning component. Decision Trees, SVMs, Naïve Bayes, SVMs, dimensionality reduction, clustering (Gaussian Mixture models, k-means).

Other elements in the thought process, such as exploratory data analysis, cross validation, feature importance, data normalization, appropriate visualization, choosing the complexity of the model, evaluation of the model are considered.

This is a classification problem on the census income data set from the UCI Machine Learning repository. The problem is to look at various characteristics of a person such as the number of years of formal education, age, gender, profession etc, to try and classsify their income as being below or above $50,000.

After preprocessing of the data, various supervised models (DT, SVM, kNN) are trained on it. Then, I pick one of them, a simple decision tree model, and use it as a weak-learning base for ensemble methods like AdaBoost. I then compare with random forests as well. Feature importance is considered and it can be seen that if chosen carefully, a small feature subset can be enough to get almost as good a performance as the performance from having the full feature set.

This famous dataset contains housing data collected back in 1978. It has the prices of a small number of houses - only 506. There are 14 features available for each house. The problem is a regression problem. Is it possible to use features such as the number of rooms, the neighborhood in which it is located, to learn what the selling price would be?

The problem here is an exploration of decision tree regressors, with all their limitations and strengths, a look at bias-variance trade-off and the use of cross-validation.

This directory has a self-contained script to compress any image to a few colors without losing a lot of the information. We should expect to be able to compress the image to about a third of its size by this alone. This is done by k-means clustering the pixels in color-space.

Here, we segment the costumers (retailers) of a goods distribution company based on the orders they place for various items. Other matters explored are EDA, PCA for dimensionality reduction.

Here, the dataset is the [MovieLens User Ratings Dataset] (https://grouplens.org/datasets/movielens/). We have about 100,000 ratings for 9,125 movies from 183 users. k-means can be used to cluster them into groups of users of movies. The Silhouette score is used to explore the number of groups to clusters to use.

We can use the information from these clusters to predict ratings that a particular user may give to a certain movie if that (user, movie) pair has no rating recorded.

The other is a project for fraud detection by looking at emails from the Enron email dataset. The task is to identify the people who were likely to have been involved in the Enron scandal by looking at a number of features, most of the relevant ones being related to their compensation details and the degree of communication they had with others who were already known to have been involved. Please see README within the directory for more.

One of the items is simply a description of the GPU-enabled machine to run my projects as an alternative to running them on the cloud.