[Paper Review] Detecting and Adapting to Irregular Distribution Shifts in Bayesian Online Learning

Aodong Li, Alex Boyd+2 more|arXiv (Cornell University)|Dec 15, 2020

Data Stream Mining Techniques4 citations

TL;DR

This paper proposes a model-agnostic Bayesian online learning framework that detects and adapts to irregular distribution shifts by introducing a binary change variable to dynamically temper past posteriors, enabling partial forgetting. Using beam search over change-point hypotheses, the method improves performance on concept and covariate drift tasks in both supervised and unsupervised settings, outperforming state-of-the-art approaches on real-world datasets including sensor data, malware, and word embeddings.

ABSTRACT

We consider the problem of online learning in the presence of distribution shifts that occur at an unknown rate and of unknown intensity. We derive a new Bayesian online inference approach to simultaneously infer these distribution shifts and adapt the model to the detected changes by integrating ideas from change point detection, switching dynamical systems, and Bayesian online learning. Using a binary 'change variable,' we construct an informative prior such that--if a change is detected--the model partially erases the information of past model updates by tempering to facilitate adaptation to the new data distribution. Furthermore, the approach uses beam search to track multiple change-point hypotheses and selects the most probable one in hindsight. Our proposed method is model-agnostic, applicable in both supervised and unsupervised learning settings, suitable for an environment of concept drifts or covariate drifts, and yields improvements over state-of-the-art Bayesian online learning approaches.

Motivation & Objective

• To address the challenge of distribution shifts in online learning where data distributions change irregularly and unpredictably after model deployment.
• To enable models to balance remembering past knowledge and adapting to new data distributions without hard-coded assumptions about shift frequency or intensity.
• To develop a scalable, model-agnostic method that integrates change detection with Bayesian online learning for both supervised and unsupervised learning.
• To improve generalization and robustness in non-stationary environments such as autonomous systems, financial trading, and NLP applications.

Proposed method

• Introduces a binary 'change variable' to indicate whether incoming data is compatible with the current data distribution, enabling dynamic adaptation.
• Uses the change variable to condition the prior in Bayesian online learning, allowing partial erasure of past information via posterior tempering when a shift is detected.
• Employs beam search over multiple change-point hypotheses to track and select the most probable sequence of shifts in hindsight.
• Applies the framework to both supervised and unsupervised settings, including sequential SVHN/CIFAR-10 with image rotation-induced covariate shift and word embedding dynamics.
• Uses variational Bayesian inference with a compact, interpretable latent structure to model evolving semantics over time.
• Incorporates beam diversification by dropping the bottom two hypotheses before ranking to enhance exploration of alternative change-point trajectories.

Experimental results

Research questions

• RQ1 How can a Bayesian online learning system detect irregular distribution shifts without prior knowledge of their frequency or intensity?
• RQ2 What mechanism enables effective adaptation to new data distributions while preserving useful memory of past data?
• RQ3 Can a unified framework integrate change detection with Bayesian online learning to improve performance across diverse real-world datasets?
• RQ4 How does the proposed method compare to state-of-the-art approaches in handling concept drift and covariate shift?
• RQ5 To what extent can the method model interpretable semantic shifts in word embeddings over time?

Key findings

• The method significantly outperforms state-of-the-art Bayesian online learning approaches on datasets with concept drift, including basketball player trajectories, malware characteristics, sensor data, and electricity prices.
• On sequential SVHN and CIFAR-10 with covariate drift from image rotations, the model achieves improved generalization and robustness compared to baselines.
• In unsupervised word embedding analysis, the method captures interpretable semantic shifts, such as 'atom' shifting from 'element' to 'nuclear' in 1945, and 'race' shifting post-1991 after the Cold War ended.
• The beam search mechanism enables tracking of multiple plausible change-point sequences, improving robustness and accuracy in shift detection.
• The model is highly scalable, successfully applied to models with hundreds of thousands of parameters and tens of thousands of feature dimensions.
• The rate-distortion curve shows that the method achieves lower L1 dating error than the 'binning' baseline across all episode lengths, with error decreasing as more semantic changes are allowed per year.