Multi-class Classifier Category Compression

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing multi-class classification systems face challenges in efficiently training accurate models when dealing with high-cardinality and unevenly-distributed target variable categories, leading to resource inefficiencies and accuracy degradation due to informational noise from infrequently-seen categories.

Innovation Solution

The proposed solution involves identifying infrequent and non-separable target variable categories within the training dataset, encoding these categories into a general 'others' category, and then training a multi-class classifier using the encoded dataset, thereby reducing resource utilization and improving model accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If high-cardinality target variable categories are used in multi-class classification, then the model can capture more detailed category information, but resource utilization (learning time, memory footprint, CPU cycles) increases

Engineering Contradiction:

Improvecategory information detailVSAvoidlearning time

Core Design Contradiction:

Loss of informationVSLoss of time

Solution Approach 1:

The patent extracts and removes infrequent categories from the training dataset, retaining only frequent categories for model training. This extraction reduces the cardinality of the target variable while preserving the most informative categories, thereby reducing learning time and resource utilization while maintaining predictive accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent discards infrequent categories that contribute informational noise and increase resource consumption, while recovering the essential category information through the remaining frequent categories. This selective discarding approach reduces training complexity without losing critical classification capabilities.

Inventive Principle:

Principle #34Discarding and recovering

2Loss of information

If high-cardinality target variable categories are used in multi-class classification, then the model can capture more detailed category information, but memory footprint increases

Engineering Contradiction:

Improvecategory information detailVSAvoidmemory footprint

Core Design Contradiction:

Loss of informationVSVolume of stationary object

Solution Approach 1:

The patent extracts and removes infrequent categories from the training dataset, retaining only frequent categories for model training. This extraction reduces the cardinality of the target variable while preserving the most informative categories, thereby reducing memory footprint while maintaining predictive accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent discards infrequent categories that contribute informational noise and increase memory consumption, while recovering the essential category information through the remaining frequent categories. This selective discarding approach reduces training complexity without losing critical classification capabilities.

Inventive Principle:

Principle #34Discarding and recovering

3Adaptability or versatility

If infrequently-seen categories are included in the training dataset, then the model can learn from all possible categories, but model accuracy decreases due to informational noise

Engineering Contradiction:

Improvecategory coverageVSAvoidmodel accuracy

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent converts the harmful effect of infrequent categories (informational noise reducing accuracy) into a benefit by systematically identifying and removing them. The harm of excluding rare categories is mitigated by using frequency thresholds that preserve enough category diversity for accurate classification while eliminating noisy infrequent categories.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

Data Source

PatentUS12326879B2Compression of target variable for multi-class classification

Publication Date: 2025.06.10 SAP SE

AI summary

Systems and methods include determination of a dataset comprising a plurality of instances, each instance comprising a value of each of a plurality of input variables and of a target variable, where the values of the target variable comprise a plurality of categories, determination of two or more infrequent categories of the plurality of categories from the dataset, determination of two or more non-separable categories from the two or more infrequent categories based on the dataset, changing of occurrences of the two or more non-separable categories within the dataset to a single category to generate a modified dataset, and training of a classifier to output a value of the target variable based on the modified dataset.