| Feature Selection Toolbox | |
| Developer: | UTIA, Czech Academy of Sciences |
| Latest Release Version: | 3.1.1 |
| Programming Language: | C++ |
| Operating System: | Cross-platform (v3) |
| Genre: | Machine learning, pattern recognition |
| License: | Free for non-commercial use |
Feature Selection Toolbox (FST) is software primarily for feature selection in the machine learning domain,[1] written in C++, developed at the Institute of Information Theory and Automation (UTIA), of the Czech Academy of Sciences.
The first generation of Feature Selection Toolbox (FST1) was a Windows application with user interface allowing users to apply several sub-optimal, optimal and mixture-based feature selection methods on data stored in a trivial proprietary textual flat file format.[2]
The third generation of Feature Selection Toolbox (FST3) was a library without user interface, written to be more efficient and versatile than the original FST1.[3]
FST3 supports several standard data mining tasks, more specifically, data preprocessing and classification, but its main focus is on feature selection. In feature selection context, it implements several common as well as less usual techniques, with particular emphasis put on threaded implementation of various sequential search methods (a form of hill-climbing). Implemented methods include individual feature ranking, floating search, oscillating search (suitable for very high-dimension problems) in randomized or deterministic form, optimal methods of branch and bound type, probabilistic class distance criteria, various classifier accuracy estimators, feature subset size optimization, feature selection with pre-specified feature weights, criteria ensembles, hybrid methods, detection of all equivalent solutions, or two-criterion optimization. FST3 is more narrowly specialized than popular software like the Waikato Environment for Knowledge Analysis Weka, RapidMiner or PRTools.[4]
By default, techniques implemented in the toolbox are predicated on the assumption that the data is available as a single flat file in a simple proprietary format or in Weka format ARFF, where each data point is described by a fixed number of numeric attributes. FST3 is provided without user interface, and is meant to be used by users familiar both with machine learning and C++ programming. The older FST1 software is more suitable for simple experimenting or educational purposes because it can be used with no need to code in C++.