Skip to content

Latest commit

 

History

History
72 lines (58 loc) · 3.66 KB

README.md

File metadata and controls

72 lines (58 loc) · 3.66 KB
Raw

ML-snippets

Snippets of machine learning code for educational porpoises

When I write or find a very short piece of code for some machine learning task that looks simple enough to be educational I put it in here and refer my students to it.

Hello!

Contents:

  • DBOCC: density-based one-class classification.

    One-class classification is also known as anomaly detection. It's about training a classifier when you only have one set of data -- no labels. We can do it be learning the density (or similar concepts) of the data, and then applying a threshold -- new points are classed as anomalies if they lie in regions where the training data is of low density. This code implements several related approaches to OCC, and is built in Python using Numpy, Scipy, and Scikit-learn components.

  • Liver: simple contingency table, predict majority, and logistic regression on the well-known BUPA Liver Disorders dataset. There has been a widespread misconception that the final variable indicates presence or absence of a liver disorder in the subjects. In fact, the final variable is just a train/test selector. Richard Forsyth and I wrote an article discussing the issue. This (very simple) code is part of that.

  • Uniform In Hypersphere: generate vectors uniformly distributed in a hypersphere of given dimension. Two methods are supplied, which are similar but not identical. One, from Tax and Duin 2001, corrects what I think is an error in the original.

  • RecSys: a collaborative filtering recommender system, using singular value decomposition by manual gradient descent. Based on Funk's and Paterek's work towards the Netflix prize.

  • SVM: I think a lot of SVM tutorials give all the details about the maximum margin separating hyperplane, the quadratic programming and support vectors, and radial kernels, but don't give a good intuition on one important part of the big picture. When the kernel does its implicit mapping from the original feature space to a new feature space, what does that new feature space look like? What do the features in that new space mean? This short notebook tries to fill in the missing link.

  • Representations for AI: This notebook/presentation is about learning representations (embeddings) for data using classic machine learning methods, and also using a modern method (a Siamese neural network with triplet loss). It is part of the Atlantec 2019 AI Tools and Techniques session.

  • Kernel Regression Association: This notebook presents a novel approach to measuring the strength of association between two real variables. Correlation measures strength of linear relationship, while rank-based correlation such as Kendall's tau measures strength of non-linear monotonic relationship. Mutual information measures strength of non-linear, non-monotonic relationship, but depends on an arbitrary binning of the data into discrete bins. Kernel Regression Association aims to measure the strength of non-linear, non-monotonic relationship without binning, relying on the training error of kernel regression of one variable against the other (and vice versa).

  • Metric Distortion: This .py file presents an approach to measuring metric distortion. Suppose we have metric spaces X and Y and a mapping f: X -> Y.The metric distortion is a measure of how well f preserves structure, in particular how well it preserves distances. If d_X(x1, x2) = d_Y(f(x1), f(x2)) for all x1, x2 in X, for some distance d, then f preserves structure perfectly so there is low metric distortion.