Multi-Label Classification via Model Merging and Data Transformation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multi-label classification approaches are inefficient due to the need for multiple binary classification models, which are resource-intensive and impractical for large-scale applications, especially in quality management systems where scalability and processing efficiency are critical.

Innovation Solution

Transforming the original multi-label classification task into a multi-class classification task using a predefined protocol, training a multi-class classification model on the adapted data set, and then converting multi-class output data back into multi-label output data using a label combiner operator, thereby simplifying the classification process and reducing resource requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If multiple binary classification models are used for multi-label classification, then classification accuracy can be maintained, but system complexity and resource consumption increase significantly

Engineering Contradiction:
Improveclassification accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent combines multiple binary classification models into a single multi-class classification model. Instead of training separate binary models for each label, the invention creates one unified model that directly predicts multiple class labels simultaneously, thereby reducing system complexity while maintaining classification accuracy through the adapted training data transformation approach

Inventive Principle:
Principle #5Merging (Combining)

2Adaptability or versatility

If multiple binary classification models are deployed for multi-label classification, then comprehensive label coverage is achieved, but processing time and computational resources increase

Engineering Contradiction:
Improvelabel coverageVSAvoidprocessing efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The invention merges multiple binary classification tasks into a single multi-class classification operation. By transforming the training data into adapted training data with combined classification labels and using a label combiner operator during prediction, the system achieves comprehensive label coverage through one model inference instead of multiple sequential binary classifications, significantly improving processing efficiency

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent performs preliminary transformation of the training data before model training. By pre-processing the training data to create adapted training data where multiple labels are combined into single classification labels, and pre-defining the label combiner operator, the system prepares the classification task in advance to enable efficient single-pass prediction that covers all labels

Inventive Principle:
Principle #10Preliminary action

3Reliability

If multiple binary classification models are trained for multi-label classification, then each label can be optimized independently, but implementation complexity and resource requirements become impractical for large-scale applications

Engineering Contradiction:
Improvelabel-wise optimizationVSAvoidimplementation feasibility
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The invention merges the training process of multiple binary models into a single multi-class model training operation. The adapted training data transformation allows the model to learn all label relationships simultaneously through unified loss optimization, making implementation feasible for large-scale applications while maintaining the ability to optimize label predictions through the label combiner operator during inference

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4478265A1Apparatus, method, and computer program product for multi-label classification using adapted multi-class classification model
Publication Date: 2024.12.18 HONEYWELL INTERNATIONAL INC
  • EP4478265A1 patent drawingFigure 1
  • EP4478265A1 patent drawingFigure 2
  • EP4478265A1 patent drawingFigure 3

AI summary

A multi-label classification system performs multi-label classification using an adapted multi-class classification model in conjunction with pre-classification and post-classification processing and data transformation. An original, multi-label training data set in which objects are classified into at least one but possibly multiple original classification sets is transformed into an adapted, multi-class training data set in which the same objects are each classified into a combined classification set representing a combination of all of the individual, original classification sets into which the object is classified in the original training data set. The adapted training data set is used to train a multi-class classification model, which trained model is used to generate a multi-class classification of objects with respect to an input data set. The multi-class classification is transformed into a multi-label classification.