Source-linked AI summary
Learning under Concept Drift: A Review
Jie Lu, Anjin Liu, Fan Dong, Feng Gu, Joao Gama, Guangquan Zhang
TL;DR
Concept drift makes patterns learned from past streaming data unreliable as target distributions change, motivating methods that detect, understand, and adapt to drift. This survey synthesizes recent research, proposes a three-component framework, reviews benchmark datasets, and identifies research directions. It finds that detection methods generally identify when drift occurs, while relatively few determine how severe it is or where it occurs.
Problem
Changing streaming-data distributions can make patterns learned from past data irrelevant, while prior surveys left some concept-drift subproblems and recent developments insufficiently covered.
Method
The survey reviews concept-drift research and organizes learning under drift into detection, understanding, and adaptation, while examining synthetic and real-world datasets.
Results
Detection methods generally identify when drift occurs, but few can determine how severe it is or where it occurs.
Takeaways & Limitations
Future research should provide drift severity and region information for adaptation, address costly or delayed labels, and establish better real-world stream evaluation frameworks.
Takeaways & Limitations
Real-world datasets often lack precise ground-truth drift start and end times and may contain mixed drift types, complicating evaluation.
Abstract
from arXiv · showhide
Concept drift describes unforeseeable changes in the underlying distribution of streaming data over time. Concept drift research involves the development of methodologies and techniques for drift detection, understanding and adaptation. Data analysis has revealed that machine learning in a concept drift environment will result in poor learning results if the drift is not addressed. To help researchers identify which research topics are significant and how to apply related techniques in data analysis tasks, it is necessary that a high quality, instructive review of current research developments and trends in the concept drift field is conducted. In addition, due to the rapid development of concept drift in recent years, the methodologies of learning under concept drift have become noticeably systematic, unveiling a framework which has not been mentioned in literature. This paper reviews over 130 high quality publications in concept drift related research areas, analyzes up-to-date developments in methodologies and techniques, and establishes a framework of learning under concept drift including three main components: concept drift detection, concept drift understanding, and concept drift adaptation. This paper lists and discusses 10 popular synthetic datasets and 14 publicly available benchmark datasets used for evaluating the performance of learning algorithms aiming at handling concept drift. Also, concept drift related research directions are covered and discussed. By providing state-of-the-art knowledge, this survey will directly support researchers in their understanding of research developments in the field of learning under concept drift.
1 INTRODUCTION
Concept drift arises when changing environments make patterns learned from past streaming data unreliable for prediction and decisions. This survey organizes the field around detection, understanding, and adaptation, reviews recent research and datasets, and identifies emerging directions.
- Changing products, markets, and customer behaviors alter streaming-data distributions, making past patterns unreliable for current predictions and decisions.
- Concept drift research increasingly targets detection in noisy data, explainable understanding of drift, and adaptation of knowledge to changing conditions.
- The survey structures learning under concept drift into detection, understanding of when, how, and where drift occurs, and adaptation after drift.
- Earlier surveys left some subproblems uncovered, while newer work introduced multiple-hypothesis-test detection and motivated an updated review of research trends.
- The review selects references from four publication databases using keyword screening, research quality or recency filtering, and grouping into detection, understanding, and adaptation.
- 137 research articles, 10 synthetic datasets, and 14 real-world datasets are listed for discussion, with dataset characteristics, availability, drift suitability, and applications examined.
2 PROBLEM DESCRIPTION
The survey defines concept drift as a time-varying change in the joint distribution of features and labels, which may arise from changes in features, conditional labels, or both. It also distinguishes drift types by how concepts transform over time.
- 2.1 Concept drift definition and the sources: Concept drift occurs when the statistical properties of a target domain change over time, formally when P_t(X, y) differs from P_t+1(X, y).
- 2.1 Concept drift definition and the sources: Because P_t(X, y) decomposes into P_t(X) × P_t(y|X), the survey identifies feature-distribution drift, conditional-label drift, and their mixture as three sources.
- 2.1 Concept drift definition and the sources: Feature-distribution drift leaves the decision boundary unchanged, whereas conditional-label drift changes the boundary and can reduce learning accuracy.
- 2.1 Concept drift definition and the sources: In real-world applications, feature-distribution and decision-boundary drift often occur together, forming the mixed third source.
- 2.2 The types of concept drift: The survey distinguishes four commonly defined drift types and describes adaptation for Types 1–3 as minimizing accuracy loss and accelerating recovery during transformation.
- 2.2 The types of concept drift: Type 4 drift emphasizes retrieving matching historical concepts because the new concept may suddenly, incrementally, or gradually reoccur.
- 2.2 The types of concept drift: Intermediate concepts describe transformations between starting and ending concepts, including mixtures in incremental drift and endpoint concepts in gradual drift.
3 CONCEPT DRIFT DETECTION
Concept drift detection is organized as a four-stage process that retrieves and optionally models streaming data, measures distributional dissimilarity, and tests its statistical significance. The reviewed algorithms span error-rate, data-distribution, and multiple-hypothesis approaches, with trade-offs involving accuracy, computational cost, window specification, and label requirements.
- 3.1 A general framework for drift detection: The general framework retrieves data chunks, optionally models them, calculates test statistics, and applies a hypothesis test.Data retrieval organizes instances into meaningful chunks; data modeling can reduce dimensionality or sample size.
- 3.1 A general framework for drift detection: Accurate and robust dissimilarity measurement remains the most challenging open question in concept drift detection.The measurement quantifies drift severity and forms the test statistic used by the hypothesis test.
- 3.1 A general framework for drift detection: Without hypothesis testing, a drift statistic cannot establish statistical significance or a drift confidence interval.The hypothesis test evaluates whether the observed change is statistically meaningful.
- 3.2.1 Error rate-based drift detection: Error-rate methods monitor classifier errors and trigger model upgrades when error changes become statistically significant.DDM uses warning and drift levels: it begins training a new learner at warning level and replaces the old learner at drift level.
- 3.2.2 Data Distribution-based Drift Detection: Data-distribution methods compare historical and new-data distributions, providing drift timing and location information but usually requiring higher computational cost.These methods commonly require predefined historical and new-data windows.
- 3.2.3 Multiple Hypothesis Test Drift Detection: Hierarchical methods use a detection layer followed by a validation layer, while HHT-AG uses fewer true labels under high verification latency.The validation layer provides a second hypothesis-test-based validation of detected drift.
4 CONCEPT DRIFT UNDERSTANDING
Concept drift understanding augments detection by characterizing when, how severely, and where drift occurs, supplying information for adaptation. The review finds that all detectors identify occurrence time, most distribution-based methods quantify severity, and few locate drift regions.
- Overview: Drift understanding retrieves when, how, and where information and passes it from detection to adaptation.“When” covers occurrence and duration, “how” covers severity, and “where” covers drift regions.
- When: Drift alarms can lag the actual onset because detectors require a minimum number of new observations, while warning levels support model updating.Warning thresholds are more relaxed than drift thresholds, and accumulated warning-to-drift data can train an updated learner.
- How: Severity quantifies the discrepancy between successive distributions, with larger non-negative values indicating larger concept changes.It is represented as ∆ = δ(Pt(X, y), Pt+1(X, y)).
- How: Severity can guide adaptation: low drift may favor incremental learning, whereas high drift may favor discarding the old learner and retraining.The review notes that severity information is not yet widely used in adaptation.
- Where: Drift regions identify feature-space conflicts and help preserve stable-region models, select updated instances, remove obsolete conflicts, and distinguish noise from novel data.Several adaptation approaches use localized regions for model reuse or training-set construction.
- Overview: All drift detection algorithms identify when drift occurs, most distribution-based methods measure severity, and only a few locate drift regions.The review summarizes these capabilities in Table 2.
5 DRIFT ADAPTATION
Drift adaptation updates learning models after detected change through simple retraining, ensemble retraining, or model adjustment. The reviewed strategies trade off freshness, reuse of recurring concepts, and efficiency for localized changes.
- Overview: Drift adaptation comprises simple retraining, ensemble retraining, and model adjusting for different drift types.These methods update existing learning models according to detected drift.
- Simple retraining: Simple retraining replaces an obsolete model with a new model trained on latest data after an explicit detector signals drift.Window strategies commonly preserve recent data for retraining or distribution-change testing.
- Simple retraining: Window size creates a trade-off: small windows reflect recent distributions better, while large windows provide more training data.ADWIN examines possible window cuts rather than requiring a fixed user-specified size.
- Ensemble retraining: Ensembles preserve reusable base classifiers for recurring drift and combine their outputs through voting rules.Adaptive ensembles may add, reweight, activate, or deactivate classifiers as performance changes.
- Model adjusting: Model adjustment partially updates learners when distributions change, which can be more efficient than retraining when drift is localized.Decision trees are common because they can adapt separately to feature-space subregions.
- Model adjusting: Recent studies question whether the Hoeffding bound is appropriate for VFDT node splitting because its information-gain variables are not independent.Alternative impurity measures and relative-frequency calculations are proposed for drift-aware tree adaptation.
6 EVALUATION, DATASETS AND BENCHMARKS
The review evaluates concept-drift learners through validation procedures, specialized metrics, significance tests, and synthetic or real-world datasets. Synthetic data provide controlled drift scenarios, whereas real-world data offer realism but uncertain drift ground truth and mixed drift types.
- Evaluation systems: Evaluation systems are reviewed across validation methodology, evaluation metrics, and statistical significance, with computation equations and usage guidance.The review expands earlier work by organizing evaluation across these three aspects.
- Validation methodology: The three concept-drift validation procedures are holdout, prequential, and controlled permutation.Holdout requires time-matched concepts, prequential tests before training each instance, and controlled permutation preserves local temporal distributions.
- Validation methodology: Prequential evaluation avoids requiring known drift times and makes maximum use of available streaming data.Its error can use landmark-window, sliding-window, or forgetting-mechanism estimates.
- Evaluation metrics: Recommended metrics extend traditional accuracy measures with RAM-hours, Kappa statistics, prequential AUC, and normalized area-under-curve measures.Kappa accounts for class imbalance, while additional measures address temporal dependence and streaming drift.
- Statistical significance: Non-parametric tests include sign, Wilcoxon signed-rank, and Nemenyi tests for comparing learners across instances or multiple datasets.The Nemenyi test compares average algorithm ranks over datasets.
- Datasets: Synthetic datasets support controlled evaluation of specified drift scenarios, while real-world datasets provide realistic benchmarks with unknown precise drift boundaries and potentially mixed drift types.Real-world data therefore make evaluation of some methods difficult.
7 THE CONCEPT DRIFT PROBLEM IN OTHER RESEARCH AREAS
Concept drift handling extends beyond stream mining into imbalanced learning, big-data mining, active and semi-supervised learning, and decision support. These areas adapt models or sampling to changing distributions alongside their domain-specific constraints.
- Cross-domain applications: Concept drift handling is reviewed as an indirect-use problem across research areas beyond standalone drift research.The survey examines developments that benefit from addressing drift.
- Class imbalance: Imbalanced stream learning combines drift adaptation with resampling, penalty constraints, time-decayed metrics, or drift-triggered retraining.Examples include Learn++.CDS, Learn++.NIE, ESOS-ELM, OOB, and UOB.
- Big-data mining: Big-data drift methods address distribution uncertainty together with scalability when streams are too large or unpredictable for one server.Reviewed approaches include online MapReduce drift detection and parallel micro-cluster nearest-neighbor methods.
- Active learning: Active learning under drift manages limited labeling through sampling strategies, budget control, bias prevention, and memory loss factors.The reviewed frameworks combine instance selection with drift adaptation mechanisms.
- Semi-supervised learning: Semi-supervised drift learning leverages unlabeled data through Gaussian mixtures, clustering, adaptive ensembles, and asynchronous handling of labeled and unlabeled streams.Some reviewed methods target gradual or recurring drift and detect changes at cluster boundaries.
- Decision support: Adaptive decision rules handle drift by dynamically adding and removing rules according to error rates monitored by drift detectors.Adaptive VFDR extends Very Fast Decision Rules for streaming data.
8 CONCLUSIONS: FINDINGS AND FUTURE DIREC-
The survey identifies dominant methods, unresolved gaps, and emerging trends in concept drift research, then proposes four directions for future work and practical application.
- Findings: Error-rate and data-distribution methods remain dominant for drift detection, while multiple-hypothesis testing has emerged recently.
- Findings: Detection methods generally identify when drift occurs, but few explain how or where it occurs.
- Findings: Adaptive models and ensemble techniques are increasingly important, whereas retraining with explicit drift detection has slowed.
- Findings and future directions: Most methods assume labels become available after prediction, leaving unsupervised and semi-supervised drift handling comparatively underexplored.
- Future directions: Future detection research should estimate drift timing, severity, and regions, because these signals could support improved adaptation.
- Future directions: Future work should establish real-world stream-selection frameworks and integrate drift-handling techniques with machine-learning methodologies for data-driven applications.