This is a machine learning library, made from scratch.

It uses:

• numpy: for handling matrices/vectors
• scipy: for various mathematical operations
• cvxopt: for convex optimization
• networkx: for handling graphs in decision trees

It contains the following functionality:

• Supervised Learning:
  • Linear and Logistic regression
    • Regularization
    • Solvers
      • Gradient descent
      • Steepest descent
      • Newton's method
      • SGD
      • Backtracking line search
      • Closed form solutions
  • Support Vector Machines
    • Soft and hard margins
    • Kernels
  • Tree Methods
    • CART (classificiation and regression)
    • PRIM
    • AdaBoost
    • Gradient Boost
    • Random Forests
  • Kernel Smoothing Methods
    • Nadaraya average
    • Local linear regression
    • Local logistic regression
    • Kernel density classification
  • Discriminant Analysis
    • LDA, QDA, RDA
  • Naive Bayes Classification
    • Gaussian
    • Bernoulli
  • Prototype Methods
    • KNN
    • LVQ
    • DANN
  • Perceptron
• Unsupervised Learning
  • K means/mediods clustering
  • PCA
  • Gaussian Mixtures
• Model Selection and Validation

Examples are shown in two dimensions for visualisation purposes, however, all methods can handle high dimensional data.

• Linear and logistic regression with regularization. Closed form, gradient descent, and SGD solvers.

• Support vector machines maximize the margins between classes

• Using kernels, support vector machines can produce non-linear decision boundries. The RBF kernel is shown below

• An alternative learning algorithm, the perceptron, can linearly separate classes. It does not maximize the margin, and is severely limited.

• The library contains a large collection of tree methods, the basis of which are decision trees for classification and regression

These decision trees can be aggregated and the library supports the following ensemble methods:

• AdaBoosting
• Gradient Boosting
• Random Forests

Kernel methods estimate the target function by fitting seperate functions at each point using local smoothing of training data

• Nadaraya–Watson estimation uses a local weighted average

• Local linear regression uses weighted least squares to locally fit an affine function to the data

• The library also supports kernel density estimation (KDE) of data which is used for kernel density classification

• Linear Discriminant Analysis creates decision boundries by assuming classes have the same covariance matrix.
• LDA can only form linear boundries

• Quadratic Discriminant Analysis creates deicion boundries by assuming classes have indepdent covariance matrices.
• QDA can form non-linear boundries.

• Regularized Discriminant Analysis uses a combination of pooled and class covariance matrices to determine decision boundries.

• K-nearest neighbors determines target values by averaging the k-nearest data points. The library supports both regression and classification.

• Learning vector quantization is a prototype method where prototypes are iteratively repeled by out-of-class data, and attracted to in-class data

• Discriminant Adaptive Nearest Neighbors (DANN). DANN adaptively elongates neighborhoods along boundry regions.
• Useful for high dimensional data.

• K means and K mediods clustering. Partitions data into K clusters.

• Gaussian Mixture Models. Assumes data are generated from a mixture of Gaussians and estimates those Gaussians via the EM algorithm. The decision boundry between two estimated Gaussians is shown below.

• Principal Component Analysis (PCA) Transforms given data set into orthonormal basis, maximizing variance.