Autorelevance function and other feature relevance measures for univariate time series

2026-07-02 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A new model-agnostic methodology is introduced for measuring lag relevance in machine learning forecasting models applied to univariate time series. This approach leverages Ghost variables, Shapley values, and additive importance measures to define "auto-relevance" and "partial auto-relevance" functions, which quantify lag importance. The authors also propose a novel technique for coalition-based methods, replacing absent features with a one-step forecast derived from the same model. This combined framework is specifically designed for time series analysis. The methodology's effectiveness was evaluated through various simulations and real-world data cases, utilizing models from the seasonal ARMA family and recurrent neural networks. The relevance measures consistently demonstrated the expected lag structure across nearly all test scenarios.

Key takeaway

For Machine Learning Engineers developing univariate time series forecasting models, understanding lag relevance is crucial for model interpretability and feature engineering. You should consider implementing the proposed auto-relevance and partial auto-relevance functions to quantify the importance of historical data points. This approach offers a model-agnostic way to identify significant lags, potentially improving feature selection and model performance, especially when working with complex models like recurrent neural networks.

Key insights

A model-agnostic framework quantifies lag relevance in univariate time series forecasting using novel auto-relevance functions.

Principles

• Lag relevance can be measured model-agnostically.
• Coalition-based methods benefit from forecast-based feature replacement.
• Ghost variables and Shapley values inform lag importance.

Method

The methodology combines Ghost variables, Shapley values, and additive importance measures to define auto-relevance and partial auto-relevance functions for lag importance. Absent features in coalition methods are replaced with a one-step forecast from the same model.

In practice

• Apply auto-relevance to understand time series lag structures.
• Use one-step forecasts for missing features in coalition methods.
• Evaluate lag importance with seasonal ARMA or RNN models.

Topics

• Time Series Forecasting
• Lag Relevance
• Shapley Values
• Ghost Variables
• Feature Importance
• Recurrent Neural Networks
• ARMA Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Spectral Retrieval-Augmented Time-Series Forecasting — SpecReTF enhances time series forecasting by integrating spectral analysis and temporal recency into its retrieval mechanism.
• Autoregressive Models: Predicting the Future Using the Past — Autoregressive models predict future values by leveraging patterns observed in their own historical data.
• From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting — Importance-aware compression and PRM-guided retrieval improve news-enhanced time series forecasting by managing context and selection.
• Product Forecasting through Time Series Analysis (Modelling) — Effective product sales forecasting requires a hierarchical, multi-model approach combining ML, deep learning, and statistical methods.
• Exploring Novel Uncertainty Quantification through Forward Intensity Function Modeling — A new framework uses forward intensity functions to quantify uncertainty in time-to-event predictions with dynamic stochastic predictors.
• PeakFocus: Bridging Peak Localization and Intensity Regression via a Unified Multi-Scale Framework for Electricity Load Forecasting — PeakFocus unifies peak timing localization and intensity regression to overcome smoothing and misalignment in electricity load forecasting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.