Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Event Detection via Probability Density Function Regression

Clark Peng, Tolga Dinçer

TL;DR

The paper tackles the challenge of accurately detecting time-interval events in time series, where segmentation-based methods often struggle with onset/offset precision and class imbalance. It proposes a generalized regression framework that predicts probability density functions at event locations (two outputs for onset and offset) and can retrofit existing segmentation models by swapping the final layer; targets are stabilized via normalization. Across sleep and seizure datasets, pdf regression matches or exceeds segmentation baselines, with tuned models achieving notable EDAP gains and demonstrating robustness across architectures. The approach is model-agnostic, extensible to online detection and ensembling, and holds promise for broader applications including change-point detection, making it a versatile tool for precise temporal event localization in complex time-series data.

Abstract

In the domain of time series analysis, particularly in event detection tasks, current methodologies predominantly rely on segmentation-based approaches, which predict the class label for each individual timesteps and use the changepoints of these labels to detect events. However, these approaches may not effectively detect the precise onset and offset of events within the data and suffer from class imbalance problems. This study introduces a generalized regression-based approach to reframe the time-interval-defined event detection problem. Inspired by heatmap regression techniques from computer vision, our approach aims to predict probability densities at event locations rather than class labels across the entire time series. The primary aim of this approach is to improve the accuracy of event detection methods, particularly for long-duration events where identifying the onset and offset is more critical than classifying individual event states. We demonstrate that regression-based approaches outperform segmentation-based methods across various state-of-the-art baseline networks and datasets, offering a more effective solution for specific event detection tasks.

Event Detection via Probability Density Function Regression

TL;DR

The paper tackles the challenge of accurately detecting time-interval events in time series, where segmentation-based methods often struggle with onset/offset precision and class imbalance. It proposes a generalized regression framework that predicts probability density functions at event locations (two outputs for onset and offset) and can retrofit existing segmentation models by swapping the final layer; targets are stabilized via normalization. Across sleep and seizure datasets, pdf regression matches or exceeds segmentation baselines, with tuned models achieving notable EDAP gains and demonstrating robustness across architectures. The approach is model-agnostic, extensible to online detection and ensembling, and holds promise for broader applications including change-point detection, making it a versatile tool for precise temporal event localization in complex time-series data.

Abstract

In the domain of time series analysis, particularly in event detection tasks, current methodologies predominantly rely on segmentation-based approaches, which predict the class label for each individual timesteps and use the changepoints of these labels to detect events. However, these approaches may not effectively detect the precise onset and offset of events within the data and suffer from class imbalance problems. This study introduces a generalized regression-based approach to reframe the time-interval-defined event detection problem. Inspired by heatmap regression techniques from computer vision, our approach aims to predict probability densities at event locations rather than class labels across the entire time series. The primary aim of this approach is to improve the accuracy of event detection methods, particularly for long-duration events where identifying the onset and offset is more critical than classifying individual event states. We demonstrate that regression-based approaches outperform segmentation-based methods across various state-of-the-art baseline networks and datasets, offering a more effective solution for specific event detection tasks.
Paper Structure (26 sections, 3 equations, 7 figures, 3 tables, 3 algorithms)

This paper contains 26 sections, 3 equations, 7 figures, 3 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background and Problem Statement
  3. Previous Work
  4. Regression Methodology
  5. Experiments and Evaluation
  6. Data Description
  7. Evaluation
  8. Objective Distributions
  9. Segmentation
  10. Regression
  11. Preprocessing
  12. Modeling and Training Parameters
  13. Postprocessing
  14. Regression Postprocessing
  15. Segmentation Postprocessing
  16. ...and 11 more sections

Figures (7)

  • Figure 1: Example of Segmentation vs PDF Regression Target
  • Figure 2: Example regarding a convolved gaussian pdf over a binary event detection problem. The bottom graph shows two time series, corresponding to the onset and offset of the event respectively.
  • Figure 3: Custom-designed Model Architectures, a Bidirectional RNN model, a 1D UNet Model and a 1D UNet model with an Attention layer perslev2019
  • Figure 4: All post-processing procedures step-by-step. Including both forms of segmentation processing and the regression post-processing results
  • Figure 5: Chains used to caculate timeseries features using a bipolar montage
  • ...and 2 more figures