Abstract
The distributions of real-life data streams are usually nonstationary, where one exciting setting is that a stream can be decomposed into several offline intervals with a fixed time horizon but different distributions and an out-of-distribution online interval. We call such data multi-distributional data streams, on which learning an on-the-fly expert for unseen samples with a desirable generalization is demanding yet highly challenging owing to the multi-distributional streaming nature, particularly when initially limited data is available for the online interval. To address these challenges, this work introduces a novel optimization method named coupling online-offline learning (CO$_2$) with theoretical guarantees about the knowledge transfer, the regret, and the generalization error. CO$_2$ extracts knowledge by training an offline expert for each offline interval and update an online expert by an off-the-shelf online optimization method in the online interval. CO$_2$ outputs a hypothesis for each sample by adaptively coupling both the offline experts and the underlying online expert through an expert-tracking strategy to adapt to the dynamic environment. To study the generalization performance of the output hypothesis, we propose a general theory to analyze its excess risk bound related to the loss function properties, the hypothesis class, the data distribution, and the regret.
Introduction
In real-world applications like financial forecasting or sensor monitoring, data streams are rarely stationary. The underlying rules generating the data often shift over time, a phenomenon known as "concept drift." Traditional machine learning approaches typically struggle with this: global learners become confused by conflicting historical distributions, while "fresh start" learners suffer from severe cold-start problems every time the distribution shifts.
This paper addresses a more realistic scenario: multi-distributional data streams. In these systems, data can be partitioned into distinct offline intervals (historical patterns) and an active online interval. The authors propose a framework called Mixture of Online and Offline Experts (MOOE), which leverages the Coupling Online-Offline learning (CO$_2$) optimization method to bridge the gap between historical wisdom and real-time adaptation.
The MOOE Framework
MOOE operates by maintaining a dynamic library of experts that are adaptively weighted based on their current performance. This allows the system to instantly "recognize" and reuse a historical distribution if it reappears, while still evolving to handle entirely new patterns.
As shown in the architecture above, the system consists of three core components:
- The Online Expert: A standard learner (using Online Gradient Descent) that updates with every new data point to capture immediate shifts.
- The Offline Expert Set: A collection of static models trained on previous fixed-length intervals, serving as the system's "memory."
- The Meta-Expert: A manager that uses an exponential weighting scheme to decide how much to trust each expert. If an old expert suddenly starts predicting the current data accurately, the Meta-Expert quickly shifts weight toward it.
Theoretical Foundations and the Dilution Effect
Beyond empirical success, the work provides a rigorous theoretical suite, deriving a regret bound of $O(\sqrt{T})$. This is the "gold standard" for online learning, proving that MOOE is as robust as a single-distribution learner even in changing environments.
A key insight from the theory is the "dilution effect." While it might seem beneficial to keep every historical expert, the authors show that having too many weak experts can actually hurt performance by diluting the meta-expert's advice.
Experimental Evaluation
The authors evaluated MOOE against state-of-the-art methods across eight real-world datasets, including network intrusion (Kitsune) and wearable stress sensing (WESAD).
Key results include:
- Superior Accuracy: MOOE outperformed all competitors, achieving 93.4% accuracy on the Kitsune dataset compared to 87.3% for the next best method.
- Recurring Drift Mastery: On datasets where user interests cycle over time (Email and Spam filtering), MOOE showed an 11.8% improvement in accuracy. This confirms the framework's ability to successfully "recycle" old experts when historical patterns return.
- Robustness: The weighting mechanism proved excellent at filtering out "distracted" experts even as Gaussian noise was increased in the stream.
Conclusion
MOOE offers a compelling blueprint for learning from non-stationary data streams. By coupling the stability of offline historical experts with the agility of online optimization, it provides a system that learns from the present without forgetting the lessons of the past. Future work may focus on automating the selection of interval sizes and expert counts for completely unknown streams.
Summary updated