Time Series Model Explanation via First-Order Difference Perturbation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for explaining time series machine learning models, such as LIME, struggle with data that is not independent and identically distributed (IID), which is common in time series data.

Innovation Solution

The proposed solution involves determining the first order difference in time series input data and historical training data, performing perturbation of the input data based on these differences, and generating a uniform random sample to determine other input values for the time series machine learning model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing explanation methods like LIME are used on time series data, then model prediction explanation is attempted, but the methods fail to handle non-IID data properly

Engineering Contradiction:

Improveexplanation reliabilityVSAvoiddata distribution adaptability

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent transforms the time series data by computing first-order differences, converting the original non-IID time series into a new representation that better satisfies IID assumptions. This parameter transformation (from raw values to differences) enables standard explanation methods to work effectively on time series data while preserving the explanatory power for the original predictions

Inventive Principle:

Principle #35Parameter changes

2Ease of operation

If perturbation is applied to time series input data, then explanation generation is enabled, but maintaining data dependencies becomes difficult

Engineering Contradiction:

Improveexplanation generationVSAvoiddata dependency structure

Core Design Contradiction:

Ease of operationVSStability of the object's composition

Solution Approach 1:

The patent performs preliminary transformation of the time series data into first-order differences before applying perturbation. By pre-processing the data to remove temporal dependencies, the method enables independent perturbation of each data point while still preserving the ability to explain the original model predictions, thus facilitating explanation generation without violating data dependencies

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12210939B2Explaining machine learning based time series models

Publication Date: 2025.01.28 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

A method, computer system, and computer program product for explaining time series machine learning model are provided. The embodiment may include determining a first order difference in time series input data and historical training data. The embodiment may also include performing perturbation of time series input data based on the determined first order difference and the determined historical training data. The embodiment may further include computing closeness of the determined first order difference in the historical training data to the determined first order difference in the time series input data. The embodiment may also include generating a uniform random sample of first value input to a time series machine learning model. The embodiment may further include determining values of other inputs to the time series machine learning model based on the generated random sample and a random sample from the historical training data first order differences.