Fractal Dimension Estimation for Robust Machine Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models struggle to accurately estimate the latent feature dimensionality of high-dimensional datasets, leading to potential learning of spurious or non-causal relationships due to noise in the data.

Innovation Solution

The method involves monitoring N observed variables of a training dataset, translating them into equisized bin indexes to generate hypercells, and computing the fractal dimension based on data coverage to estimate the fundamental dimensionality of the dataset.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning models use high-dimensional datasets with many observed variables, then the model can capture more potential relationships and improve prediction accuracy, but the model may learn spurious or non-causal relationships due to noise and over-specification

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel stability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent extracts and removes noise and spurious relationships from high-dimensional datasets by computing the fundamental dimensionality of latent features. This extraction process separates meaningful signal from noise, allowing the model to focus on genuine causal relationships while discarding irrelevant dimensions that contribute to instability and spurious correlations.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of dimensionality from the observed high-dimensional space to the fundamental latent dimensionality computed through fractal analysis. By transforming the dimensionality parameter from N observed variables to d fundamental latent features, the model achieves both high prediction accuracy and stability by operating in the true underlying dimension space rather than the noisy observed space.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If the model complexity is increased to capture nonlinear relationships, then the model can better fit the data, but the model becomes more sensitive to noise and less robust

Engineering Contradiction:
Improvedata fitting accuracyVSAvoidmodel robustness
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent performs preliminary action by computing the fundamental dimensionality and determining the appropriate model complexity before training the machine learning model. This advance determination of the fundamental dimension d guides the selection of model architecture and complexity, ensuring that the model is neither under-specified nor over-specified, thereby achieving robust nonlinear fitting without excessive sensitivity to noise.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes the complexity parameter from being arbitrarily high to being precisely matched to the fundamental dimensionality d of the latent features. By adjusting the model complexity parameter to align with the computed fundamental dimension, the model achieves optimal balance between fitting accuracy and robustness, capturing nonlinear relationships without overfitting to noise.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If dimensionality reduction techniques are applied to reduce noise, then model stability improves, but the risk of losing important information increases

Engineering Contradiction:
Improvemodel stabilityVSAvoidinformation preservation
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent performs preliminary action by computing the fundamental dimensionality d before applying dimensionality reduction. This advance knowledge of the true latent dimension guides the reduction process, ensuring that exactly d dimensions are retained while reducing from N observed dimensions. This preliminary determination prevents both under-reduction (retaining noise) and over-reduction (losing information), achieving stable models with complete information preservation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent replaces conventional dimensionality reduction techniques (which are mechanical and arbitrary) with a physics-based fractal dimension computation that naturally identifies the fundamental dimension d. This substitution replaces arbitrary dimension selection with a scientifically grounded measurement, ensuring that dimensionality reduction preserves all meaningful information while eliminating only the noise dimensions.

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

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach allows for the estimation of robust latent feature dimensionality, improving the stability and accuracy of machine learning models by avoiding over-specification and capturing the correct nonlinear relationships.

Implementation Method 1

computing a fractal dimension of the equisized hypercube phase space based on count of cells with data coverage of at least one data point

Methodology Applied
Scientific EffectFractal dimension: Fractal Forms

Data Source

PatentEP4553712A1Latent feature dimensionality bounds for robust machine learning on high dimensional datasets
Publication Date: 2025.05.14 FAIR ISAAC & CO INC
  • EP4553712A1 patent drawingFigure 1
  • EP4553712A1 patent drawingFigure 2(a)~2(d)
  • EP4553712A1 patent drawingFigure 3

AI summary

Computer-implemented methods and systems for quantifying appropriate machine learning model complexity corresponding to training dataset are provided. The method comprises monitoring, using one or more processors, N observed variables, vi through vN, of a training dataset for a machine learning model; translating the N observed variables into m equisized bin indexes which generate mN possible equisized hypercells to estimate a fundamental dimensionality for the dataset; generating one or more samples by assigning a record in the dataset with numbers j through k as set id; generating a merged sample Si, for one or more values of the set id i, where i goes from j to k; and computing a fractal dimension of the equisized hypercube phase space based on count of cells with data coverage of at least one data point.