tilitools

tiny little machine learning toolbox

Tilitools is a collection of (non-mainstream) machine learning model and tools with a special focus on anomaly detection, one-class learning, and structured data. Author: Nico Goernitz

Examples, description, and lectures examples can be found in the notebooks/ directory.

Test data was collected from ODDS (outlier detection database).

Currently available models:

• Bayesian data description
• support vector data description:
  1. dual qp
  2. latent
  3. cluster
  4. multiple kernel learning
• one-class support vector machine:
  1. huber loss primal
  2. latent
  3. primal lp-norm w/ SGD solver
  4. dual qp
  5. convex semi-supervised
  6. multiple kernel learning
• lp-norm mkl wrapper
• structured output support vector machine (primal)
• latent variable principle components analysis
• LODA

structured objects:

• multi-class
• hidden markov model

supported kernels and features:

• rbf, linear kernel
• histogram intersection kernel
• histogram features
• hog features

lectures: Lectures contains exercise and solution notebooks for various topics. Right now the following are available:

• introduction to anomaly detection
• optimization I + II
• learning with kernels I + II
• geoscience project grainstones: detecting fossils in microscopic images

notebooks: Contains notebooks to various topics related to machine learning, anomaly detection, and structured output learning. Currently, the following are available:

• high-dimensional outlier detection
• the anomaly detection setting
• Bayesian data description
• introduction to SVDD and OCSVM

data: The data sub-directory contains well-known benchmark datasets which are modified to fit in the anomaly detection setting. These modified datasets can be downloaded from ODDS website (outlier detection database).

installation: There are two packages that need to be installed separately

• pyOD (for LODA)
• pytorch / torchvision