跳转至

文章背景与核心概要

自监督学习（SSL）在从光电容积脉搏波（PPG）等充满噪声的无约束生理信号中提取通用表征方面表现卓越，然而其在高度主观的任务中的有效性仍存疑问。本文评估了基于 PPG 的 SSL 在检测真实生活强烈情绪方面的效能，研究人员在无约束数据上预训练了一个真实生活 PPG 编码器（RL-PPG）。

研究发现，在物理活动识别任务中，这些表征迁移效果极佳，在留一主体交叉验证（LOSO）评估中带来了近 5 倍的性能提升；然而，当应用于相同 LOSO 协议下的主观真实生活情绪检测时，这些通用表征未能超越朴素基线。关键结论表明，通过采用跨时间（Across-Time）验证策略，微调时纳入个人的私有数据是驱动预测性能的核心因素——其重要性远远超过了群体层级的预训练。最终，个性化被证明是实现现实世界情感推理的基本要求。

Take it Personally: The Limits of General SSL Representations for Real-Life PPG Emotion Detection

Take it Personally: The Limits of General SSL Representations for Real-Life PPG Emotion Detection

arXiv: 2608.14675 [cs.LG]
Submitted: August 3, 2026
Authors: Dominika Kunc, Przemysław Kazienko, Stanisław Saganowski
Conference: Accepted at the 14th International Conference on Affective Computing and Intelligent Interaction (ACII 2026)

arXiv: 2608.14675 [cs.LG]
Submitted: August 3, 2026
Authors: Dominika Kunc, Przemysław Kazienko, Stanisław Saganowski
Conference: Accepted at the 14th International Conference on Affective Computing and Intelligent Interaction (ACII 2026)

Summary

Summary

虽然自监督学习（SSL）在从光电容积脉搏波（PPG）等充满噪声、无约束的生理信号中提取通用表征方面表现出色，但其在高度主观任务中的有效性仍值得怀疑。

While Self-Supervised Learning (SSL) excels at extracting general representations from noisy, unconstrained physiological signals like photoplethysmography (PPG), its effectiveness for highly subjective tasks remains questionable.

本文评估了基于 PPG 的 SSL 在检测真实生活强烈情绪方面的效能。研究人员在无约束数据上预训练了一个真实生活 PPG 编码器（RL-PPG）： * 物理活动识别： 作为健全性检查（sanity check），这些表征的迁移效果非常好，在留一主体交叉验证（LOSO）评估中，性能提升了近 5 倍。 * 情绪检测： 当在相同的 LOSO 协议下应用于主观的真实生活情绪检测时，这些通用表征未能超越朴素基线。 * 核心发现： 通过使用跨时间（Across-Time）验证策略，作者证实了在微调过程中融入个人的私有数据是驱动预测性能的主要因素——这远远超过了群体层级的预训练。最终，研究表明个性化是现实世界情感推理的基本要求。

This paper evaluates the efficacy of PPG-based SSL for detecting real-life intense emotions. The researchers pretrained a Real-Life PPG encoder (RL-PPG) on unconstrained data: * Physical Activity Recognition: As a sanity check, the representations transferred exceptionally well, yielding an almost 5-fold performance increase over baselines in a leave-one-subject-out evaluation (LOSO). * Emotion Detection: When applied to subjective real-life emotion detection under the same LOSO protocol, these general representations failed to surpass naive baselines. * Key Finding: Using an Across-Time validation strategy, the authors established that incorporating an individual's personal data during fine-tuning is the main driver of predictive performance—vastly outweighing population-level pretraining. Ultimately, personalization is shown to be a fundamental requirement for real-world affective inference.