Browse Prior Art Database

Method and System for Building a Classifier on Objects with Fuzzy Labels

IP.com Disclosure Number: IPCOM000248655D
Publication Date: 2016-Dec-22
Document File: 3 page(s) / 67K

Publishing Venue

The IP.com Prior Art Database

Abstract

A method and system is disclosed for building a classifier on objects with fuzzy labels. The method and system trains the classifier on the objects with the fuzzy labels/non-exclusive labels in order to obtain higher performance objectives while performing classification.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

1

Method and System for Building a Classifier on Objects with Fuzzy Labels

Disclosed is a method and system for building a classifier on objects with fuzzy labels. The method and system builds the classifier on objects with the fuzzy labels/non-exclusive labels in order to obtain higher performance objectives while performing classification.

The method and system trains a classifier on objects by setting classification objectives and implicitly specifying misclassification costs. Then, the method and system derives a result of classification and measures performance of the classifier.

FIG.1 illustrates a flow of the method and system for building a classifier on objects with fuzzy labels.

Figure 1

As illustrated in FIG.1, the method and system creates a tree of classes corresponding to labels with objects. The classes include specific classes, general classes and artificial classes. The specific classes are located under the general classes. The method and system may not represent the artificial classes with no objects. If a set of subclasses for the classes does not cover all possible options for the classes, then the method and system introduces a new subclass as others.

Then, the method and system adjusts the tree according to a desired loss structure. According to the desired loss structures, a loss value depends on a node height. The node height D is multiplied on \gamma^D, where 0 < \gamma < 1.

Based on the desired loss structure, for the classes with a low desired misclassification cost, a common node is placed on bottom levels of the tree by introducing artificial node above the common node. For the classes with high desired misclassification cost, the common node is placed on top levels of tree by specifying a highly important class as a separate class on an upper level.

Once the tree is adjusted, the method and system chooses objectives for a parent node. The objectives represent accuracy and minimal sensitivity.

2

Based on the objectives, the method and system computes cost coefficients for the classes and multiplies the loss value on the cost coefficients.

For instance, if an objective for a parent node is to maximize accuracy, then the method and system calculates cost coefficients as 1 (C_y=1) for children classes. The children classes are not balanced if the objective of the parent node is to maximize accuracy. If an objective for the parent node is to maximize a minimal sensitivity of children, then the method and system calculates cost coefficients as C_y=S/(N*M_y), where N is a number of children, M_y is a number of objects for a current node, S=\sum_y M_y is a total number of objects for all children of the parent node. The children classes are balanced if the objective for the parent node is to maximize the minimal sensitivity of children. Also, if appropriate, the method and system performs balancing by removing oversampled class objects from a dataset, or using a combination of calculating cost coefficient and removing o...