Deep Factors with Gaussian Processes for Forecasting
Danielle C. Maddix Amazon Web Services Palo Alto, CA 94303、Yuyang Wang Amazon Web Services Palo Alto, CA 94303、Alex Smola Amazon Web Services Palo Alto, CA 94303
Abstract
A large collection of time series poses significant challenges for classical and neural forecasting approaches. Classical time series models fail to fit data well and to scale to large problems, but succeed at providing uncertainty estimates. The converse is true for deep neural networks. In this paper, we propose a hybrid model that incorporates the benefits of both approaches. Our new method is data-driven and scalable via a latent, global, deep component. It also handles uncertainty through a local classical Gaussian Process model. Our experiments demonstrate that our method obtains higher accuracy than state-of-the-art methods.
中文速览
面对海量相关时间序列的预测任务，传统统计模型（如ARIMA）虽然能给出可靠的不确定性估计，却难以跨序列共享信息、也难以扩展到大规模场景；深度神经网络虽然善于从大量序列中提取模式，却在不确定性建模上天然存在短板。为了兼得两者之长，作者提出了"带高斯过程的深度因子模型"（Deep Factors with Gaussian Processes），用一个全局LSTM网络提取所有时间序列共有的动态因子（相当于数据驱动的主成分），再用一个局部高斯过程（Gaussian Process）为每条序列建模个体随机波动，从而让不确定性在预测区间内自然传播。在公开的电力和交通数据集上与DeepAR、Prophet等方法对比，该模型在均值误差和90分位损失（P90QL）上均取得更优表现，尤其在不确定性捕捉方面优势明显。这项工作的意义在于它提供了一条将经典贝叶斯概率建模与可扩展深度学习融合的可行路径，对需要同时追求预测精度和置信区间质量的业务场景（如供应链、能源调度）具有实际价值。
原文 arXiv:1812.00098;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1812.00098v1