Don’t Drift Away - Advances and Applications of Streaming and Continual Learning

Aug 3, 2026

Upload by Quinn

Non-stationary environments where concept drift occurs demand adaptive learning models capable of continuous updates without losing useful prior knowledge. Two research directions address this need from different angles: Continual Learning manages virtual drift by learning new concepts while preventing catastrophic forgetting, whereas Streaming Machine Learning focuses on real drifts and rapid adaptation to changing data distributions. Streaming Continual Learning unifies both goals, retaining relevant past information while efficiently adapting online. The work analyzes core SCL challenges, including temporal dependencies and updating latent representations for personalization and knowledge editing, and proposes SCL benchmarks to advance shared research.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Data streams and concept drift

Virtual vs. real concept drift

Why offline training fails under drift

Continual Learning vs. Streaming Machine Learning

Online CL and related settings

Continual reinforcement learning connection

Motivation for Streaming Continual Learning

Don’t Drift Away: Advances and Applications of Streaming and Continual Learning Andrea Cossu 1 , Davide Bacciu 1 , Alessio Bernardo2 , Emanuele Della Valle2 , Alexander Gepperth3 , Federico Giannini2 , Barbara Hammer4 , and Giacomo Ziffer2 1-University of Pisa 2-DEIB, Politecnico di Milano 3-University of Applied Sciences Fulda 4-Bielefeld University Abstract. Non-stationary environments subject to concept drift require the design of adaptive models that can continuously learn and update. Two primary research communities have emerged to address this challenge: Continual Learning (CL) and Streaming Machine Learning (SML) . CL manages virtual drifts by learning new concepts without forgetting past knowledge, while SML focuses on real drifts, rapidly adapting to evolving data distributions. However, a unified approach is needed to balance adaptation and knowledge retention. Streaming Continual Learning (SCL) bridges the gap between CL and SML, ensuring models retain useful past information while efficiently adapting to new data. We explore key challenges in SCL, including handling temporal dependencies in data streams and adapting latent representations for personalization and knowledge editing. Additionally, we identify promising SCL benchmarks which can foster and promote a unified research effort between CL and SML. 1 Introduction The world we live in is inundated with data. The rise of Social Media and the Internet of Things in the last decade has driven rapid data growth, generating continuous, unbounded flows called data streams. These streams produce infinite sequences of elements that arrive over time and cannot be accessed simultaneously. A significant challenge is developing learning models that adapt to data’s continuously evolving nature. Data may, in fact, change its distribution over time, causing the problem of concept drift. Concept drift is defined as an unforeseeable change in the data-generating process that induces a change in the statistical properties, which is more significant than random fluctuations or anomalies [1, 2] . A concept is the unobservable process that generates data. Concept drifts can be categorized into two main types: virtual and real. Given a machine learning problem, a virtual concept drift happens when the input or the target distribution changes without affecting the rules that map the inputs to the desired targets. For example, in a classification problem that maps the inputs X to the labels y, a concept drift changes the probability P (X|y) or P (y) without affecting P(y|X) . On the other hand, a real concept drift changes the mapping rules between inputs and targets. In a classification problem, this change impacts the probability P (y|X) . For instance, in credit scoring, where the goal is to predict loan default risk based on factors like income and credit history, virtual concept drift occurs when the distribution of these factors shifts over time without altering their relationship with default risk, keeping the decision boundary unchanged. In contrast, real concept drift arises when economic changes, market trends, or new financial policies reshape this relationship, meaning that the same income and credit history may now indicate a different level of risk. This requires the model to adapt its decision boundary. Due to concept drift, it is unfeasible to accumulate the data in a repository and train the model offline. First of all, the repository would contain inconsistent data due to concept drift during data acquisition, and, in addition, subsequent concept drift would render to model obsolete, thus requiring massive re-training using a new repository. The model should, instead, continuously learn and update using the data stream. Two communities emerged to reach this end: Continual Learning (CL) [3] and Streaming Machine Learning (SML) [4] . The main distinctions between the two lie in their objectives and the types of drifts they manage. The most prominent

FAQs

What problem does Streaming Continual Learning target?

It targets non-stationary environments where concept drift requires models to learn continuously while retaining useful past knowledge and adapting efficiently to new data.

How do virtual and real concept drift differ?

Virtual drift changes input/target distributions without altering the mapping implied by P(y|X), while real drift changes the underlying mapping so the decision boundary must adapt.

Why can’t models simply train offline on accumulated data under concept drift?

Accumulated repositories mix inconsistent data collected under different drift conditions, and later drift makes the earlier data and resulting model obsolete, requiring expensive retraining.

Aug 3, 2026

Upload by Quinn