Temporal Feature-Dynamics Models for Covariate-Shift Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine-learning models experience classification errors due to covariate shift in feature distributions, particularly in applications involving temporal data streams, leading to inaccurate classifications and potential safety issues in automated systems.

Innovation Solution

A computer-implemented method that utilizes a feature dynamics model to track the evolution of feature densities over time, adapting the machine-learning model's weights to compensate for distribution shifts, thereby improving classification performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a pre-trained machine-learning model is used for classification, then classification speed is improved, but classification accuracy deteriorates due to covariate shift in feature distributions

Engineering Contradiction:
Improveclassification speedVSAvoidclassification accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent implements a feature dynamics model that dynamically adapts the pre-trained machine-learning model to temporal changes in feature distributions. The model continuously updates its parameters based on evolving feature densities over time, allowing it to maintain high classification speed while compensating for covariate shift and preserving accuracy in non-stationary environments.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the pre-trained model by introducing a feature dynamics model that models the evolution of feature density over time. This allows the model to adapt to distribution shifts by adjusting its internal parameters dynamically, thereby maintaining classification accuracy despite changes in the underlying data distribution.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If the machine-learning model is adapted to track distribution shifts, then classification accuracy is improved, but model complexity increases

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

Solution Approach 1:

The patent segments the classification system into two distinct components: a pre-trained machine-learning model for fast classification and a feature dynamics model for tracking distribution shifts. This segmentation allows each component to specialize in its strength while keeping the overall system manageable and interpretable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The feature dynamics model serves as an intermediary between the pre-trained model and the changing data distribution. It mediates the adaptation process by modeling feature density evolution and guiding parameter updates, thereby simplifying the complexity of directly adapting the main model to distribution shifts.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Speed

If classification is performed on temporal data streams, then real-time processing capability is improved, but classification error increases due to continuous distribution shift

Engineering Contradiction:
Improvereal-time processing capabilityVSAvoidclassification error rate
Core Design Contradiction:
SpeedVSReliability

Solution Approach 1:

The patent implements a feedback mechanism where the feature dynamics model continuously monitors the evolution of feature densities in temporal data streams and uses this information to adjust the pre-trained model's parameters. This closed-loop feedback allows the system to maintain reliable classification in real-time despite continuous distribution shifts.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent ensures continuous adaptation of the model to distribution shifts by continuously tracking feature density evolution in temporal streams. This continuous useful action prevents accumulation of classification errors that would occur with periodic or batch-based adaptation approaches.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4645118A1Computer-implemented method for classifying data elements of a data set using a machine-learning model
Publication Date: 2025.11.05 ROBERT BOSCH GMBH
  • EP4645118A1 patent drawingFigure 1
  • EP4645118A1 patent drawingFigure 2
  • EP4645118A1 patent drawingFigure 3

AI summary

A computer-implemented method (200) for classifying data elements of a data set using a machine-learning model includes: classifying the data elements of the data set associated with a time sequence, wherein each data element is associated with a corresponding time step of the time sequence, wherein the data elements are classified by inputting them into the machine-learning model one after the other according to their temporal order, wherein classifying a respective data element of the data elements includes: determining features of the respective data element using a feature extractor of the machine-learning model; determining, using the features of the respective data element and the features of one or more other data elements temporally preceding the respective data element, parameters of a feature-dynamics-model which represents an evolution of a feature density of the features over time; and determining a class associated with the respective data element using the feature-dynamics-model.