Data Miner

Features

Three classifiers

TKNearestNeighbors, with its choice of distance model and number of neighbors; TNaiveBayes, with prior probabilities; and TLinearClassifier, a completely new algorithm. Each classifies a new case, returns its class response, and can be given a reject probability. All three derive from TClassifier, the abstract class you derive from to add an algorithm of your own, which TDBClassifier then drives like the others.

Real and discrete data

Attributes can be real or discrete, and missing values in either are handled. Attributes can be enabled or disabled one by one, and their quality is measured by entropy and information gain.

Incremental learning

Every classifier learns incrementally, example by example or from a whole data set, and can store the examples for later use. A learned model saves to and loads from a file or a stream.

Validation and pruning

A classification test measures the accuracy on your own test data, and TKNearestNeighbors can leave one stored example out for a leave-one-out test. Pre-pruning, post-pruning and Prune disable the attributes that only add noise and keep those that give the best accuracy.

Database classification

TDBClassifier connects a classifier to any TDataSet descendant: it learns from the dataset or the current record, classifies records, and runs classification tests over splits you define into learn and test sets. It maps fields to attributes automatically or field by field, ignores the fields you name, tells discrete fields from already indexed ones, indexes discrete values itself, takes the value you define as missing, and reads string, Boolean, integer, floating-point, currency, BCD, date and time fields, each field type mapped to a discrete or a real attribute as you choose.

Reference results

The package includes a demo with reference results on 20 standard data sets, Iris among them, examples of usage and performance tests, and a 9-page introduction to the key characteristics of classification algorithms.