Modern machine learning increasingly operates in environments where data arrives over time and conditions evolve. Rather than learning once from a fixed dataset, systems must continuously update, adapt, and remain reliable under changing distributions. This setting can be broadly described as machine learning over time under changing data.

Different assumptions about how data arrives and how knowledge should be retained lead to distinct learning paradigms. Below, we outline the main ones and how they relate.

Learning Paradigms

The traditional setting (batch learning) assumes access to a fixed dataset. Models are trained offline and deployed under the assumption that the data distribution remains stable. This paradigm enables complex optimisation but does not naturally handle evolving data.

Stream learning assumes data arrives sequentially and must be processed incrementally. Storing all past data is often infeasible, and models must update continuously. A key characteristic is that recent data is often more relevant, especially under changing distributions.

halfspace trees auc animation small.gif

In this example, we showcase data arriving over time in an anomaly detection problem and the detector identifying (and missing others). More examples are available here.

Continual learning assumes a large problem is revealed over time through tasks (or experiences). The goal is to learn incrementally while retaining previously acquired knowledge. Unlike stream learning, past and present knowledge are equally important. The central challenge is balancing:
  • Stability: retaining past knowledge
  • Plasticity: adapting to new information
Online Continual Learning (OCL) retains the objectives of continual learning while imposing the constraints of stream learning. Learning must:
  • occur incrementally (no revisiting data),
  • operate under memory and time constraints,
  • and still retain knowledge across tasks.

Learning resources

    • Tutorial , PAKDD, Taiwan, 2024: “Machine Learning for Streaming Data”