xTime: Extreme Event Prediction with Hierarchical Knowledge Distillation and Expert Fusion

2025-10-23

AI Summary

Low prediction accuracy for extreme events (e.g., floods, heatwaves, acute medical episodes) in time series stems from severe class imbalance and insufficient exploitation of precursor intermediate events. To address this, we propose a Hierarchical Knowledge Distillation and Dynamic Expert Fusion framework. Our method constructs multiple rarity-specific expert models, enabling hierarchical knowledge transfer—from frequent to rare events—via cross-rarity knowledge distillation. A dynamic gating mechanism adaptively fuses outputs from experts of varying rarity levels. Technically, the framework integrates time-series modeling, Mixture-of-Experts (MoE), and cross-rarity knowledge distillation. Experiments on multiple real-world datasets demonstrate 3%–78% improvements in extreme-event prediction accuracy over state-of-the-art methods. To our knowledge, this is the first approach to systematically achieve effective joint modeling of precursory signals and hierarchical knowledge for rare-event forecasting.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Data Mining & Knowledge Management: Anomaly/Outlier Detection
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs

Abstract

Extreme events frequently occur in real-world time series and often carry significant practical implications. In domains such as climate and healthcare, these events, such as floods, heatwaves, or acute medical episodes, can lead to serious consequences. Accurate forecasting of such events is therefore of substantial importance. Most existing time series forecasting models are optimized for overall performance within the prediction window, but often struggle to accurately predict extreme events, such as high temperatures or heart rate spikes. The main challenges are data imbalance and the neglect of valuable information contained in intermediate events that precede extreme events. In this paper, we propose xTime, a novel framework for extreme event forecasting in time series. xTime leverages knowledge distillation to transfer information from models trained on lower-rarity events, thereby improving prediction performance on rarer ones. In addition, we introduce a mixture of experts (MoE) mechanism that dynamically selects and fuses outputs from expert models across different rarity levels, which further improves the forecasting performance for extreme events. Experiments on multiple datasets show that xTime achieves consistent improvements, with forecasting accuracy on extreme events improving from 3% to 78%.

Problem

• Predicting extreme events in time series data
• Addressing data imbalance in rare event forecasting
• Capturing precursor information for extreme events

Innovation

• Hierarchical knowledge distillation transfers lower-rarity event information
• Mixture of experts mechanism dynamically fuses multi-level expert outputs
• Framework improves extreme event forecasting accuracy up to 78%