Taxonomy Tree Classification Using Link Prediction and Beam Search

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Material classification in large corporations with numerous engineering departments is time-consuming and prone to errors due to the complexity of aligning purchasing requests for similar materials across different departments, leading to inefficiencies and increased costs.

Innovation Solution

A computer-implemented method using a link prediction model with a taxonomy tree structure, incorporating tree-weighted soft labels, negative sampling, and beam search to reduce computational effort and improve classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual classification methods are used to assign materials to taxonomy classes, then classification accuracy can be maintained through expert knowledge, but the classification process becomes time-consuming and labor-intensive

Engineering Contradiction:
Improveclassification accuracyVSAvoidclassification time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces manual mechanical classification processes with an automated machine learning system. The link prediction model automatically assigns materials to taxonomy classes by learning from training data, substituting human expert manual classification with an automated computational system that maintains accuracy while dramatically reducing time consumption.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent implements preliminary action by pre-training the link prediction model on a comprehensive dataset of materials and taxonomy classes before actual classification tasks. This pre-training phase allows the system to learn patterns and relationships in advance, enabling rapid and accurate classification during operational use without requiring manual intervention for each new material.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a detailed taxonomy tree with multiple hierarchical levels is implemented to improve material classification precision, then the complexity of the classification system increases, making it more difficult to manage and compute

Engineering Contradiction:
Improveclassification precisionVSAvoidtaxonomy system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces complex manual taxonomy management with an automated link prediction model that handles the complexity of multi-level hierarchical taxonomies. The model automatically navigates the taxonomy structure, learns relationships between materials and classes at different hierarchical levels, and performs classification without requiring manual system management, thereby managing complexity through automation rather than simplification.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Ease of manufacture

If traditional machine learning models are used for material classification, then implementation is straightforward, but computational efficiency decreases when handling large-scale taxonomy trees with many classes

Engineering Contradiction:
Improveimplementation easeVSAvoidcomputational efficiency
Core Design Contradiction:
Ease of manufactureVSProductivity

Solution Approach 1:

The patent applies segmentation by dividing the classification task into distinct components: embedding generation for materials and taxonomy classes, link prediction modeling, and hierarchical navigation. This segmentation allows each component to be optimized independently, improving overall computational efficiency while maintaining implementation feasibility through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements parameter changes by transforming the classification problem into a link prediction framework with specific parameter optimizations including embedding dimensions, loss function parameters, and beam search width. These parameter adjustments enable the system to efficiently handle large-scale taxonomy trees while maintaining accuracy, improving computational productivity without sacrificing implementation ease.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4523158B1Classification method and device
Publication Date: 2026.03.11 SIEMENS AG
  • EP4523158B1 patent drawingFigure 1
  • EP4523158B1 patent drawingFigure 2
  • EP4523158B1 patent drawingFigure 3

AI summary

Computer implemented method for classification of an object having a certain set of properties into a class containing objects with at least one common property, said classes being arranged in a taxonomy tree being denoted by a class label and said classification being performed by using a trained link prediction model which provides, for a textual description of an object, a prediction of a class label of said object, wherein for the link prediction model the textual description of an object and the taxonomy tree (T ) is used as input and wherein the taxonomy tree comprises a set of nodes and a set of directed edges leading from a root node across nodes in at least one intermediate level, said nodes representing intermediate class labels to leaf nodes at a final level, said leaf nodes not having child nodes and denoting a class label and said taxonomy tree containing a structure information defined by a branching ratio at each node and a number of levels, wherein for the application of the link prediction model, structure information is used so that the number of computations to obtain a class label for a given textual descriptions reduced by pursuing only selected paths from the root node to leaf nodes.