[论文解读] Amortized Predictability-aware Training Framework for Time Series Forecasting and Classification

Xu Zhang, Peng Wang+2 more|arXiv (Cornell University)|Feb 18, 2026

Time Series Analysis and Forecasting被引用 0

一句话总结

tldr: APTF动态识别并惩罚训练过程中低可预测性时间序列样本，利用分层可预测性感知损失（Hierarchical Predictability-aware Loss）和一个摊销模型来提升TSF和TSC性能。

ABSTRACT

Time series data are prone to noise in various domains, and training samples may contain low-predictability patterns that deviate from the normal data distribution, leading to training instability or convergence to poor local minima. Therefore, mitigating the adverse effects of low-predictability samples is crucial for time series analysis tasks such as time series forecasting (TSF) and time series classification (TSC). While many deep learning models have achieved promising performance, few consider how to identify and penalize low-predictability samples to improve model performance from the training perspective. To fill this gap, we propose a general Amortized Predictability-aware Training Framework (APTF) for both TSF and TSC. APTF introduces two key designs that enable the model to focus on high-predictability samples while still learning appropriately from low-predictability ones: (i) a Hierarchical Predictability-aware Loss (HPL) that dynamically identifies low-predictability samples and progressively expands their loss penalty as training evolves, and (ii) an amortization model that mitigates predictability estimation errors caused by model bias, further enhancing HPL's effectiveness. The code is available at https://github.com/Meteor-Stars/APTF.

研究动机与目标

• Motivate the need to address low-predictability samples in time series data to improve training stability and generalization.
• Introduce a general framework (APTF) that applies to both TSF and TSC tasks.
• Develop a loss formulation (HPL) that dynamically buckets samples by predictability and scales their losses.
• Incorporate an amortization model to reduce predictability estimation bias and enhance learning.
• Demonstrate empirical gains across multiple datasets and baseline models.

提出的方法

• Define low-predictability buckets based on loss values and assign decreasing weights to higher-loss buckets.
• Introduce Hierarchical Predictability-aware Loss (HPL) with bucket groups to stabilize gradients and balance learning from high- and low-predictability samples.
• Implement an amortization model to absorb predictability estimation bias and pass estimates with a one-step delay to improve HPL.
• Incorporate predictability evolution by stage-wise reduction of buckets and averaging losses across bucket groups (Hierarchical strategy).
• Evaluate APTF on 11 short-term TSF datasets, 8 long-term TSF datasets, and 128 TSC datasets from UCR, across multiple baselines.

实验结果

研究问题

• RQ1 Can predictability-aware training improve convergence and generalization for TSF and TSC tasks?
• RQ2 How can low-predictability samples be identified and penalized without discarding potentially useful information?
• RQ3 Does an amortization model reduce bias in predictability estimation and improve training stability?
• RQ4 What are the empirical gains of APTF when combined with diverse baseline models on various TSF and TSC benchmarks?

主要发现

• APTF consistently improves accuracy across eight baseline models in short-term TSF, with average improvements of 2.06% to 9.79% for transformer-based and 1.79% to 5.79% for linear models.
• APTF yields average improvements in long-term TSF of 2.01% to 13.14% across transformer models and 1.44% to 5.75% for linear models.
• In TSC on 128 UCR datasets, APTF increases average accuracy from 80.97% to 81.93% and shows more wins than ties vs. baselines.
• Ablation studies show HPL alone reduces errors on several baselines, and the amortization model further reduces forecasting errors by about 1.4% to 2.1% across baselines.
• Loss landscapes become flatter with APTF, indicating improved generalization and robustness.