Source-linked AI summary
Learning under Concept Drift: an Overview
Indrė Žliobaitė
TL;DR
Concept drift creates uncertainty because the future data source may differ from the past, complicating incremental prediction and model adaptation. This report formalizes the setting, surveys adaptive training-set and learner-design strategies, and organizes concept-drift methods and applications. It identifies change detectors and ensembles as the two most popular techniques, with ensembles more flexible across change types but potentially slower for sudden drift.
Problem
Concept drift produces uncertainty about the future data source, while training-set formation and learner design must adapt to changing data in incremental prediction.
Method
The report formalizes incremental concept-drift learning, examines adaptive training-set formation, and presents a taxonomy linking drift assumptions to learner design and evaluation.
Results
Change detectors and ensembles are the two most popular techniques; ensembles are more flexible across change types but can react more slowly to sudden drift.
Takeaways & Limitations
Concept-drift learner design depends on assumptions about change type and the future data source, while the framework places the problem within time, adaptivity, and knowledge-transfer dimensions.
Takeaways & Limitations
The framework assumes the future target source is uncertain and commonly assumes the near-future source matches the recent past.
Abstract
from arXiv · showhide
Concept drift refers to a non stationary learning problem over time. The training and the application data often mismatch in real life problems. In this report we present a context of concept drift problem 1. We focus on the issues relevant to adaptive training set formation. We present the framework and terminology, and formulate a global picture of concept drift learners design. We start with formalizing the framework for the concept drifting data in Section 1. In Section 2 we discuss the adaptivity mechanisms of the concept drift learners. In Section 3 we overview the principle mechanisms of concept drift learners. In this chapter we give a general picture of the available algorithms and categorize them based on their properties. Section 5 discusses the related research fields and Section 5 groups and presents major concept drift applications. This report is intended to give a bird's view of concept drift research field, provide a context of the research and position it within broad spectrum of research fields and applications.
1 Framework and Terminology
The framework models incremental learning as sequentially arriving instances whose generating source may change over time. It formalizes concepts, drift causes, prediction steps, and the uncertainty that motivates adaptive training.
- Incremental learning process: At each time step, historical labeled instances are used to build a learner that predicts the label of the next target instance.After the target label becomes available, that labeled instance joins historical data for the next step.
- Concept and drift: A concept is stable when all instances come from one source; concept drift occurs when sources differ across time points.Random noise is excluded when the data-generating source remains unchanged.
- Adaptivity assumption: The target source is uncertain, so the learner may use all historical data or select an adaptive training set.The report focuses on training-set formation strategies represented as a black box in the incremental learning loop.
- Formal representation: A concept or data source consists of class priors and class-conditional probability density functions for all classes.The source is denoted S, while a time-specific source may vary as the data-generating process changes.
- Classification basis: Bayesian classification assigns an instance to the class with the maximal posterior probability, with p(X) constant across classes.The posterior is expressed through class priors, class-conditional densities, and the evidence of X.
- Causes of a concept drift: Drift can involve changing class priors, class-conditional distributions, or posterior class-membership distributions.Changes in p(X|c) may leave class membership unaffected, and the report does not distinguish real from virtual drift in practice.
2 How Do Concept Drift Learners Work?
Concept drift learner design separates assumptions about future data, change patterns, adaptivity mechanisms, and model selection. The framework links these choices to training-set formation and evaluation strategies.
- Design framework: Four design sub-problems are future assumption, change type, learner adaptivity, and model selection.Model selection chooses a learner parametrization at each time step, such as ensemble weights or window size.
- Future assumption: Future assumptions include maintaining the current source, estimating it from the unlabeled target, or predicting the change.Assuming St+1 = St is the most common option, though it is rarely stated explicitly.
- Change types: The main structural drift types are sudden, gradual, incremental, and reoccurring context, based on source configurations over time.The categories are qualitative and not exhaustive; predictability is treated as part of future-assumption design rather than drift type.
- Learner adaptivity: Four adaptivity areas are adaptive base learners, adaptive parametrization, adaptive training-set formation, and ensemble fusion rules.Training-set formation includes selection, manipulation, and feature-set manipulation, and is the report’s stated scope and focus.
- Learner adaptivity: Training-set selection strategies divide into windowing, which uses consecutive instances, and instance selection.Windowing is generally preferred for sudden drift, while instance selection is preferred for gradual drift and reoccurring contexts.
- Model selection: Model evaluation can use theoretical generalization-error analysis or cross-validation, and the choice depends on the assumed future source.The design process connects adaptivity strategies to change-type assumptions and model selection to future-source assumptions.
3 Taxonomy of Available Concept Drift Learners
Concept drift learners are organized by when adaptation is activated and by how models adapt. The overview emphasizes evolving methods, especially ensembles, whose adaptation depends on model evaluation, fusion, and training-set formation.
- Taxonomy: Learners are divided into trigger-based and evolving groups according to when adaptivity is activated.Trigger-based methods use signals that influence model construction; evolving methods generally avoid explicit change detection and select or maintain models using performance estimates.
- Taxonomy: A second grouping distinguishes adaptation through training-set formation from adaptation through base-learner design or parametrization.
- Evolving learners: Classifier ensembles combine or select outputs from several models using fusion rules.Fusion weights indicate expected base-learner competence and may be based on historical performance, selective cross-validation, or learner-specific estimates.
- Evolving learners: Ensemble construction commonly seeks diverse classifiers representing distinct concepts through temporal blocks, differently sized windows, or different learner types.Blocks favor sudden and recurring contexts, while different learner types can use the same training data to diversify the ensemble.
- Scope: Existing techniques generally build individual models only from data already observed, while using unseen noise or unlabeled testing data remains future work.
- Other mechanisms: Other evolving mechanisms adapt by weighting instances, modifying feature spaces, or managing model-specific parameters and designs.Instance weighting can emphasize misclassified examples, while other methods add temporal or performance features, adjust windows, tune regression parameters, or transfer support vectors.
3.2 Learners with triggers
Trigger-based learners use signals to change models or sampling. The overview covers change detectors, heuristic and accuracy-based window sizing, and instance selection, while noting that ensembles are more flexible but may react more slowly to sudden drift.
- Trigger mechanisms: Trigger-based methods determine when models or sampling should change.
- Change detection: Change detection monitors raw data, learner parameters, or learner outputs and usually cuts the training window at a detected change point.The change point and the resulting training window need not coincide.
- Window sizing: Heuristics and base-learner-specific methods determine training-window sizes using error monitoring, lookup-table actions, or historical accuracy.
- Instance selection: Instance-selection triggers inspect unlabeled incoming instances and select training sets using prototypes or historical training instances.
- Discussion: Change detectors and ensembles are the two most popular techniques.Change detectors suit expected sudden drift, whereas ensembles are more flexible across change types but can react more slowly to sudden drift.
4 Related Research Areas
Concept drift connects data mining and machine learning with neighboring fields organized around time, knowledge transfer, and adaptivity. These connections include sequential updating, distribution mismatch, related-instance reuse, and adaptive systems.
- Interdisciplinary context: The related research areas are organized into time, knowledge transfer, and adaptivity categories.The overview positions these areas in Figure 7.
- Time: Time-oriented fields address sequential data, model updating, temporal databases, dynamic causal relations, or nonstationarity.Examples include incremental learning, data stream mining, spatio-temporal and temporal data mining, dynamic Bayesian networks, and ARIMA models.
- Knowledge transfer: Knowledge transfer addresses differences between training and testing distributions, including discrepancies caused by changes in the data-generating process.Related areas include case-based reasoning, instance selection, transfer learning, multitask learning, multiple-source learning, covariate shift, and active learning.
- Adaptivity: Model adaptivity incorporates adaptation into learning, either for changing environments or as incoming examples improve model accuracy.Related approaches include artificial immune systems, adaptive resonance theory, dynamic optimization, and ubiquitous knowledge discovery.
- Conclusion: Concept drift spans all three dimensions—time, adaptivity, and knowledge transfer—and is not limited to data mining or machine learning.
5 Applications
The overview groups concept-drift applications into monitoring and control, personal assistance, decision making, and artificial intelligence. It evaluates them across learning speed, output speed, accuracy, and other application-relevant dimensions.
- Application scope: The application survey covers real-life supervised and unsupervised problems, their drift sources, and their associated learning tasks.
- Application types: Four application types are identified: monitoring and control, personal assistance, decision making, and artificial intelligence.Monitoring and control often uses unsupervised learning for abnormal-behavior detection, including adversarial activity in web, network, telecommunications, and financial data.
- Application dimensions: The application framework defines dimensions including learning and output speed and classification or prediction accuracy.
- Application dimensions: Speed requirements vary by task: fraud detection demands rapid decisions on large data loads, whereas credit evaluation may allow decisions over several days.The examples also differ in the expected relevance of adversarial activity and the required precision.
- Application comparison: The authors summarize the four application types globally in Table 2 and discuss the rationale for their tabulated choices.
5.1 Monitoring and Control
Monitoring and control applications process high-volume streams in real time, covering adversary protection and management monitoring. Their concept drift arises from evolving attacks, changing legitimate behavior, technology, traffic, environmental conditions, and human operation.
- Monitoring and control tasks process large data volumes in real time for adversary protection or management purposes.
- Computer security: Intrusion detection faces concept drift because attackers develop new strategies, technology changes, and normal behavior evolves.
- Telecommunications and finance: Fraud and insider-trading detection involve imprecise labels, severe class imbalance, and adversaries who may change transaction strategies.
- Transportation: Transportation systems must adapt to seasonal and permanent traffic changes, including accidents, driver factors, and irregular seasonality.
- Positioning, climate, and industry: Remote sensing, climate monitoring, and industrial monitoring encounter drift from changing environments, recurring contexts, or human operating habits.
- Industrial and service monitoring: Telecommunication fault detection must handle anomalies whose causes may reflect network usage changes or unrelated events.
5.2 Personal Assistance and Information
Personal assistance and information applications organize or personalize information while adapting to changing users, adversaries, topics, and data distributions. Examples include filtering, profiling, recommendation, spam detection, document organization, and economic forecasting.
- Personal assistance: Personal assistance applications personalize information, with changing user interests as a primary source of concept drift.
- Personal assistance: Text mining addresses drift in news classification, document categorization, digital libraries, and networked media organization.
- Spam filtering: Spam filtering faces rapidly changing adversarial content, seasonal and popular topics, and variation in spam volume.
- Customer profiling: Customer profiling must account for changing individual interests, while users with previously similar interests may diverge over time.
- Recommendation: Recommender systems combine sparse data with temporal variation, and competition winners used temporal information as a key to recommendation.
- Information applications: Information applications distinguish virtual drift in data distribution from real drift in class assignment, with document topics and vocabulary changing over time.
- Document organization: Time-aware topic models analyze changing topic emergence, peaks, declines, and vocabulary across long document collections.
- Economics and management: Macroeconomic forecasts, business-cycle prediction, and software project planning are affected by drifting factors and nonstationary data.
5.3 Decision Making
Decision-making and diagnostic applications often have limited, non-real-time data and high accuracy requirements. Drift is linked to hidden economic contexts, biological adaptation, demographic and treatment changes, and adversarial behavior.
- Decision-making and diagnostic applications typically use limited sequential or timestamped data, tolerate expensive computation, and incur high mistake costs.
- Finance: Bankruptcy prediction and credit scoring can require different models under different economic conditions, although nonstationarity has rarely been acknowledged.
- Biomedical applications: Biomedical drift can result from microorganism mutation and antibiotic resistance, causing treatment effectiveness to diminish over time.
- Biomedical applications: Clinical systems face changes in demographics, disease progression, drugs, and training-testing distributions caused by non-uniform sampling.
- Biometrics: Biometric authentication can drift with physiological changes such as beard growth, but adaptation requires caution because users may behave adversarially.
5.4 AI and Robotics
AI and robotics applications treat concept drift as adaptation to changing environments and user needs. Examples span navigation, robotics, smart homes, household appliances, and virtual reality.
- In AI applications, changing environments create dynamic learning problems in which learners need to adapt.
- Ubiquitous systems: Ubiquitous Knowledge Discovery covers distributed mobile systems operating in complex, dynamic, and unstable environments.
- Navigation: A 2005 DARPA navigation winner used online learning with an adaptive Mixture of Gaussians for road image classification.
- Navigation: The navigation system combined gradual adaptation by adjusting Gaussian parameters with rapid adaptation by replacing Gaussians.
- Robotics and households: Robotics, smart homes, and intelligent appliances require adaptation to changing environments and user needs.
- Virtual reality: Virtual-reality applications face drift from player cheating and differences in user strategies and skills.
- The report summarizes the discussed concept-drift applications in Table 3.
6 Terminology
Concept drift terminology is unsettled because the field is relatively new and spans broad domains. The report therefore provides alternative terminology across research fields.
- Concept drift terminology is not yet fixed because the field is relatively new and spans broad application domains.
7 Concluding Remarks
The report surveys concept-drift techniques and relevant learning tasks, while emphasizing that generalization depends on assumptions about change and that applications pose different challenges. It concludes that tailoring models to specific problems is a promising direction.
- The report provides an overview of concept-drift-responsive techniques and real learning tasks where concept drift is relevant.
- Generalization requires assumptions about the nature of change, which depends on the data and problem.
- The report identifies limited focus on applications and raises lack of real data as a possible concern.
- Concept-drift challenges differ across application types, including questions about change speed, model complexity, adaptability, and security.
- The authors consider focusing on specific models for specific problems a prospective direction.