论文标题
对使用生理数据进行情感预测的机器学习方法的简要调查
A Brief Survey of Machine Learning Methods for Emotion Prediction using Physiological Data
论文作者
论文摘要
情感预测是一个关键的新兴研究领域，侧重于识别和预测人类的情绪状态。在其他数据源中，生理数据可以作为情绪的指标，并具有额外的优势，即不能被个人掩盖/篡改，并且可以轻松收集。本文通过自我报告的生态瞬时评估（EMA）得分作为基础真相，调查了多种机器学习方法，该方法将部署智能手机和生理数据实时预测情绪。比较回归，长期记忆（LSTM）网络，卷积神经网络（CNN），强化在线学习（ROL）和深信念网络（DBN），我们展示了用于实现准确情感预测的机器学习方法的可变性。我们比较了最新方法，并强调了实验性能仍然不是很好。可以通过考虑以下问题来改善绩效：提高可伸缩性和可推广性，同步多模式数据，优化EMA采样，将适应性与序列预测相结合，收集无偏见的数据以及利用复杂的功能工程技术。
Emotion prediction is a key emerging research area that focuses on identifying and forecasting the emotional state of a human from multiple modalities. Among other data sources, physiological data can serve as an indicator for emotions with an added advantage that it cannot be masked/tampered by the individual and can be easily collected. This paper surveys multiple machine learning methods that deploy smartphone and physiological data to predict emotions in real-time, using self-reported ecological momentary assessments (EMA) scores as ground-truth. Comparing regression, long short-term memory (LSTM) networks, convolutional neural networks (CNN), reinforcement online learning (ROL), and deep belief networks (DBN), we showcase the variability of machine learning methods employed to achieve accurate emotion prediction. We compare the state-of-the-art methods and highlight that experimental performance is still not very good. The performance can be improved in future works by considering the following issues: improving scalability and generalizability, synchronizing multimodal data, optimizing EMA sampling, integrating adaptability with sequence prediction, collecting unbiased data, and leveraging sophisticated feature engineering techniques.
