Result matrix – Kofax Getting Started with Ascent Xtrata Pro User Manual

Page 112

Advertising
background image

Classification

Ascent Xtrata Pro User's Guide

93

Result Matrix

To provide a complete and detailed analysis of classification quality, Project Builder
allows you to test the classification against a reference set.

A reference set is a set of test documents (different from the training set documents)
that has a directory structure identical to the class hierarchy. If the reference set does
not have the same structure as the class hierarchy, then you will not get useful
statistical results.

The Result Matrix tool generates a matrix all of the correctly and incorrectly classified
documents, and provides statistics for recall and precision (or further details see
Thresholds, Precision, and Recall).

For calculating the result matrix, you need a reference set, or you can use the training
set.

You can generate the result matrix for the reference set using hierarchical rules, for
details see Hierarchical Evaluation and Other Classification Rules. You can also generate
the result matrix without using hierarchical rules to optimize precision and recall.

You can generate the result matrix for the training set to see if the training set
contains documents that are classified incorreclty. If so, you can easily check and
optimize the training set by creating a new class and moving the incorrectly
classified documents to a training set for the new class.

X

To calculate the result matrix using the training set

1

From the main menu bar, select Tools | Calculate Result Matrix for | Training
Set. A progress indicator displays while all the documents in the training set
are classified. When classification is finished, the result matrix displays the
calculated results.

Alternatively you can click Result Matrix from the toolbar. The Result Matrix
panel is displayed. Select the training set that you want to test from the
Selected Training Set drop down list and click Calculate Training Set to
perform classification.

Advertising