Learning in the Presence of Concept Drift and Hidden Contexts

G. WidmerM. Kubát

article1996Machine-mediated learning1,657 citations

Presents the FLORA framework of incremental learning algorithms that dynamically adjust sample windows and reuse past concept descriptions to handle recurring hidden contexts and concept drift in continuous data streams.

Listen

Real-world automated systems often operate in dynamic environments where underlying conditions change unpredictably. In domains such as network load balancing, financial modeling, or weather forecasting, the meaning of observed data can shift over time due to hidden context changes. This phenomenon, known as concept drift, presents a critical challenge for online automated systems: models that rely on historical data degrade in performance when contexts shift, while models that adapt too aggressively risk confusing random noise with meaningful operational changes.

The main objective of the article is to evaluate a family of incremental learning algorithms—named the FLORA family—designed to detect context shifts dynamically, discard outdated data through controlled forgetting, reuse previously learned concepts when old contexts recur, and maintain classification accuracy in the presence of data noise.

To evaluate these capabilities, the researchers conducted controlled experimental simulations across several synthetic benchmark domains. The evaluation assessed three system variants: a baseline model that dynamically adjusts the size of its active data window based on recent predictive accuracy and rule complexity (FLORA2); an extended model that stores and reinstalls previous concept descriptions when shifts are detected (FLORA3); and a noise-tolerant variant that replaces strict logical consistency with statistical confidence intervals around each predictive rule (FLORA4). The experiments tested these algorithms under varying conditions, including classification noise levels up to 40%, differing transition speeds between contexts, varying degrees of concept change, and the presence of irrelevant attributes.

The experiments produced several key findings. First, dynamic window adjustment allows systems to rapidly recover predictive accuracy after a concept change without succumbing to the performance slowdowns caused by overtraining on long-standing stable concepts. Second, when operational contexts recur cyclically, retrieving and adapting stored concept models significantly speeds up readjustment compared to learning from scratch, providing superior accuracy across repeated cycles. Third, incorporating statistical confidence intervals allows the system to distinguish effectively between random noise and genuine concept drift; under high noise (up to 40%), the robust variant maintained stable data windows and achieved steady, expected accuracy levels, whereas strictly consistent models destabilized and repeatedly collapsed their data windows. Fourth, the rate of recovery after a shift depends heavily on the syntactic simplicity of the new concept rather than the numerical extent of the drift alone, because simpler concepts are faster to confirm with heuristic measures.

These findings demonstrate that automated decision-making systems can maintain high reliability and low error rates in shifting environments by combining two complementary mechanisms: performance-weighted rule selection and time-based forgetting. In practical operational contexts, this balance minimizes the risk of costly misclassifications, shortens system recovery timelines after environmental changes, and avoids the computational overhead of retraining models entirely from scratch when past conditions return.

Organizations developing or deploying continuous real-time classification systems should consider implementing hybrid models that pair adaptive data windows with statistical rule validation. For environments with recurring operational phases (such as seasonal variations or regular shift patterns), implementing an explicit context repository provides tangible performance advantages. However, because heuristic window-adjustment parameters are sensitive to the complexity of the underlying rules, technical teams should run preliminary tuning trials or implement parameter-optimization techniques (such as cross-validation or parallel parameter search) prior to full deployment.

The conclusions of the article carry high confidence for symbolic, attribute-value classification problems, but readers should observe key limitations. The evaluation relied on controlled, artificial domains with boolean and discrete attributes, without testing continuous numeric variables, complex relational logic, or unconstrained drift rates where changes occur continuously with every single observation. Further validation in live industrial settings is recommended before deploying the framework to safety-critical environments.

Austrian Research Institute for Artificial Intelligence University of Ottawa University of Vienna concept drifts concept drift hidden contexts incremental learning

Widmer et al (1996).pdf

No sufficiently relevant recommendations were found.

Paper: Learning from Time-Changing Data with Adaptive Windowing, Albert Bifet et al. (2007). Extends the concept of dynamic sliding windows introduced in the FLORA framework by providing formal mathematical error bounds and theoretical guarantees with the ADWIN algorithm.
Paper: Learning with Drift Detection, João Gama et al. (2004). Generalizes statistical drift detection and window-resetting mechanisms across arbitrary classifier families using confidence thresholds on online error rates.
Paper: Mining time-changing data streams, Geoff Hulten et al. (2001). Applies sliding-window concept drift adaptation to high-speed data stream mining using incremental, bounded-memory Hoeffding decision trees.
Paper: MOA: Massive Online Analysis, A. Bifet et al. (2010). Provides a comprehensive software platform and standardized benchmarking environment for streaming algorithms operating under evolving distributions and concept drift.
Paper: Learning under Concept Drift: A Review, Jie Lu et al. (2019). Surveys and categorizes decades of subsequent research on concept drift detection, understanding, and adaptation architectures that evolved from early systems like FLORA.
Paper: A Framework for Clustering Evolving Data Streams, Charu C. Aggarwal et al. (2003). Extends the principle of adapting to evolving data streams from supervised rule learning to unsupervised clustering with multi-horizon statistical maintenance.

Cover for Learning in the Presence of Concept Drift and Hidden Contexts

Abstract

On-line learning in domains where the target concept depends on some hidden context poses serious problems. A changing context can induce changes in the target concepts, producing what is known as concept drift. We describe a family of learning algorithms that flexibly react to concept drift and can take advantage of situations where contexts reappear. The general approach underlying all these algorithms consists of (1) keeping only a window of currently trusted examples and hypotheses; (2) storing concept descriptions and re-using them when a previous context re-appears; and (3) controlling both of these functions by a heuristic that constantly monitors the system's behavior. The paper reports on experiments that test the systems' performance under various conditions such as different levels of noise and different extent and rate of concept drift.

Topics

concept drifts

Keywords

concept drift hidden contexts incremental learning instance-based learning window-based forgetting drift tracking computational learning theory classification noise

1. Introduction
2. Learning and Forgetting: The General FLORA Framework
3. Theoretical results concerning learning under drift
4. Flexible Windowing: FLORA2
4.1. Description of FLORA2
4.2. A simple experiment: The STAGGER concepts
5. Dealing with Recurring Contexts: FLORA3
5.1. Description of FLORA3
5.2. An Experiment with recurring contexts
6. Drift vs. Noise: FLORA4
6.1. Description of FLORA4
6.2. Two preliminary experiments
6.2.1. Basic drift tracking
6.2.2. FLORA4 vs. IB3
7. Systematic Experiments
7.1. Varying the amount of noise
7.2. Varying the speed of drift
7.3. Varying the extent of drift
7.4. Adding irrelevant attributes
8. Related Work
9. Conclusion
Acknowledgements
Notes
References

Knowls

Knowl 1 — FLORA Three-Set Hypothesis Representation
model/method
In the FLORA incremental concept learning framework, concept hypotheses are represented using three disjoint sets of conjunctive description items (attribute-value pairs without negation):
- $ADES$ (Accepted DEScriptors): Conjunctive description items matching positive instances in the current sliding window and no negative instances. $ADES$ forms a disjunctive normal form (DNF) formula representing the current positive concept hypothesis. Each item $ADes_i$ maintains an integer counter $AP_i$ indicating how many positive instances in the current window it matches:
$ADES = \{ADes_1/AP_1, ADes_2/AP_2, \dots\}$
- $NDES$ (Negative DEScriptors): Conjunctive description items matching negative instances in the current window and no positive instances. $NDES$ represents the negative concept hypothesis and prevents over-generalization of $ADES$ . Each item $NDes_i$ maintains a counter $NN_i$ tracking matching negative instances:
$NDES = \{NDes_1/NN_i, \dots\}$
- $PDES$ (Potential DEScriptors): A reservoir of overly general description items matching both positive and negative instances in the current window. Each item $PDes_i$ maintains two counters, $PP_i$ (positive matches) and $PN_i$ (negative matches):
$PDES = \{PDes_1/PP_1/PN_1, \dots\}$

Incoming instances update the match counters. A description item is pruned from the system whenever its match counters drop to zero as older instances are dropped from the window.
Knowl 2 — Basic FLORA Incremental Learning and Forgetting Algorithm
algorithm
The basic FLORA algorithm updates description sets and match counters when a new instance is added to the sliding window (learn_from) and when an expired instance is dropped (forget). Hypotheses are generalized using the dropping condition rule.
Function learn_from(I, is_positive):
    if is_positive is true:
        MATCH = false
        for each ADes_i in ADES:
            if ADes_i matches I:
                AP_i = AP_i + 1
                MATCH = true
        if not MATCH:
            G = minimal generalization of some ADes_i covering I without subsuming any descriptor in PDES or NDES
            if G exists:
                replace ADes_i with G in ADES
            else:
                add I to ADES with AP = 1
        for each PDes_i in PDES:
            if PDes_i matches I:
                PP_i = PP_i + 1
        for each NDes_i in NDES:
            if NDes_i matches I:
                delete NDes_i from NDES
                add NDes_i to PDES with PP = 1 and PN = NN_i
    else:
        // Dual operations swapping ADES and NDES for negative incoming instance I
Function forget(I, is_positive):
    if is_positive is true:
        for each ADes_i in ADES:
            if ADes_i matches I:
                AP_i = AP_i - 1
                if AP_i == 0:
                    delete ADes_i from ADES
        for each PDes_i in PDES:
            if PDes_i matches I:
                PP_i = PP_i - 1
                if PP_i == 0:
                    delete PDes_i from PDES
                    add PDes_i to NDES with NN = PN_i
    else:
        // Dual operations swapping ADES and NDES for negative expired instance I
Knowl 3 — FLORA2 Window Adjustment Heuristic
algorithm
FLORA2 dynamically adjusts the sliding window size $|W|$ after each training instance based on monitored predictive accuracy and the syntactic coverage of the hypothesis.

Let $N$ denote the number of positive instances in the current window covered by $ADES$ , and let $S$ be the size of $ADES$ measured as the total number of literals across all items in $ADES$ . Let $Acc$ be the classification accuracy monitored over a sliding window of recent predictions. The user defines three parameters:
- $lc$ : low coverage threshold (e.g., $lc = 1.2$ ),
- $hc$ : high coverage threshold (e.g., $hc = 4.0$ ),
- $p$ : acceptable predictive accuracy threshold (e.g., $p = 70\%$ ).
The heuristic determines $L$ , the number of oldest instances to delete from the window after adding the newest instance:
Function how_many_to_forget(W, N, S, Acc, lc, hc, p):
    Input: Current window W, covered instances N, literal count S, predictive accuracy Acc, thresholds lc, hc, p
    Output: Integer L indicating how many instances to delete from W
    if (N / S < lc) or ((Acc < p) and decreasing(Acc)):
        L = floor(0.2 * |W|) // Suspect concept drift: shrink window by 20%
    else if (N / S > 2 * hc) and (Acc > p):
        L = 2 // Extremely stable: decrease window size by 1
    else if (N / S > hc) and (Acc > p):
        L = 1 // Stable enough: maintain constant window size
    else:
        L = 0 // Need more information: grow window size by 1
    return L
Knowl 4 — FLORA3 Context Storage and Reuse Mechanism
algorithm
FLORA3 extends dynamic windowing by storing stable concept descriptions and recalling them when recurring hidden contexts are suspected.
Function choose_context(ADES, NDES, PDES, current_window, is_stable, drift_suspected, stored_concepts):
    Input: Current description sets, current instance window, stability flag is_stable, drift flag drift_suspected, set of stored concepts
    Output: Updated active concept description
    if is_stable is true:
        if ADES is not already present in stored_concepts:
            add (ADES, NDES, PDES) to stored_concepts
    else if drift_suspected is true and stored_concepts is not empty:
        // 1. Candidate evaluation
        candidates = subset of stored_concepts consistent with current instance window
        C_best = candidate in candidates maximizing ratio of positive to negative matches on current_window
        // 2. Regeneralization
        reset all match counters in C_best to 0
        for each instance I in current_window:
            reprocess I through basic FLORA learning algorithm using C_best
        remove all descriptors in C_best whose counters remain 0
        // 3. Conciseness comparison and replacement
        if total literal count of C_best.ADES < total literal count of current ADES:
            replace active concept (ADES, NDES, PDES) with C_best
Knowl 5 — FLORA4 Confidence Interval-Based Descriptor Maintenance
model/method
FLORA4 replaces the strict consistency requirement of FLORA2/FLORA3 with statistical confidence intervals around classification accuracy to handle noisy data streams.

Let $\mu \in (0, 1)$ be a user-defined confidence level (e.g., $\mu = 0.80$ ). For each description item $X$ , let $[\alpha_l, \alpha_u]$ be the lower and upper endpoints of the two-sided statistical confidence interval (at confidence level $\mu$ ) around $X$ 's classification accuracy over instances in the current window. Let $[\gamma_l, \gamma_u]$ be the confidence interval (at confidence level $\mu$ ) around the observed relative frequency of positive instances in the training stream so far.

FLORA4 maintains description sets using the following classification criteria:
- $ADES$ (Accepted Predictor): Item $X$ is placed or kept in $ADES$ if its lower accuracy endpoint exceeds the upper base rate endpoint: $\alpha_l > \gamma_u$ .
- $PDES$ (Mediocre Predictor): Item $X$ is moved to or kept in $PDES$ if its accuracy interval overlaps with the base rate interval: $\alpha_u \ge \gamma_l$ and $\alpha_l \le \gamma_u$ . Items in $PDES$ are not used for classification.
- Rejected: Item $X$ is deleted entirely if its upper accuracy endpoint falls below the lower base rate endpoint: $\alpha_u < \gamma_l$ .
- $NDES$ : Item $X$ is retained in $NDES$ as long as it is an acceptable predictor of negative instances (evaluated over negative window instances with $\alpha_l > \gamma_u$ ); unacceptable descriptors in $NDES$ are discarded directly without migration.
Knowl 6 — Piecewise Linear Transition Model for Gradual Drift Speed
model/method
Gradual concept drift between two target concepts $A$ and $B$ is modeled using a piecewise linear dominance function $\alpha(x) \in [0, 1]$ , which defines the probability that an instance at index $x$ is generated according to concept $A$ :
$1 - \frac{x - X_1}{\Delta x} & \text{if } X_1 < x < X_1 + \Delta x \ 0 & \text{if } x \ge X_1 + \Delta x \end{cases}$$ where $X_1$ is the instance index at which drift begins, and $\Delta x$ is the transition duration (in number of instances) over which $A$ is replaced by $B$. At each step $x$, the true class of the instance is assigned by concept $A$ with probability $\alpha(x)$ and by concept $B$ with probability $1 - \alpha(x)$. The transition duration $\Delta x$ defines drift speed: smaller $\Delta x$ corresponds to faster drift (e.g., $\Delta x = 50$), while larger $\Delta x$ represents slower gradual drift (e.g., $\Delta x = 200$).$
Knowl 7 — Classification Noise Robustness of FLORA4 compared to FLORA2, FLORA3, and IB3
empirical result
In evaluations on the STAGGER concepts with classification noise levels of 10%, 20%, and 40% (where class labels are flipped randomly with probability $\eta/2$ for noise level $\eta$ ):
- FLORA2 and FLORA3 degrade severely in accuracy and exhibit erratic window oscillations. Because they require strict consistency, isolated mislabeled instances cause valid descriptors in $ADES$ to be demoted to $PDES$ , triggering spurious window shrinking and growing.
- FLORA4 maintains stable window dynamics and predictive accuracy tracking the theoretical optimum under noise (reaching approximately 80% accuracy under 20% noise and 60% accuracy under 40% noise on a balanced two-class problem).
- Compared to the instance-based learner IB3, FLORA4 converges in substantially fewer examples and recovers faster after concept drift because FLORA4's explicit window-based forgetting removes outdated instances, whereas IB3's statistical exemplar retention causes prolonged inertia during concept transitions.
Knowl 8 — Concept Reuse Acceleration in FLORA3 under Recurring Hidden Contexts
empirical result
In experiments using a cyclic sequence of three STAGGER concepts ( $1 \to 2 \to 3 \to 1 \to 2 \to 3 \to 1 \to 2 \to 3$ ) switching every 40 training instances:
- During the initial encounter with each concept (instances 1–120), FLORA3 and FLORA2 exhibit identical convergence profiles because no stored context descriptions are available.
- During subsequent recurrent cycles (instances 121–360), FLORA3 readjusts significantly faster than FLORA2 in four out of six recurrence periods, reaching higher predictive accuracy with fewer instances.
- Storing stable concept descriptions and using them as biases for regeneralization avoids learning recurrent concepts from scratch.
Knowl 9 — Inverse Relationship Between Drift Extent and Recovery Time
empirical result
In an experiment testing concept transitions $A \to B_i$ on a 6-boolean-attribute domain with varying symmetric difference error extents $\text{ext}(A, B_i) = \text{Pr}_{x \sim U}(A(x) \neq B_i(x)) \in \{0.125, 0.25, 0.375, 0.5\}$ :
- FLORA learners readjust faster when the drift extent is largest ( $ext{ext} = 0.5$ ) and slowest when the drift extent is smallest ( $ext{ext} = 0.125$ ).
- This occurs because large concept discrepancies produce rapid accuracy drops and severe coverage decreases ( $N/S < lc$ ), immediately triggering the Window Adjustment Heuristic to shrink the window and purge outdated instances. In contrast, small drift extents produce subtle, sporadic classification errors that do not reliably trigger decisive window reduction.
- Furthermore, concepts with larger drift extents often admit more concise symbolic representations, which FLORA learns more rapidly.
Knowl 10 — Parameter Sensitivity and Expressiveness Limitations in FLORA
limitation
The FLORA algorithm family has two primary structural limitations:
1. Syntactic Parameter Dependency: The Window Adjustment Heuristic relies on the ratio of covered instances to the total number of literals ( $N/S$ ). As a result, the thresholds $lc$ and $hc$ are sensitive to the syntactic complexity and sparsity of the target concept, requiring empirical parameter tuning across different domains.
2. Propositional Expressiveness: The representation language is restricted to conjunctive attribute-value logic without negation, lacking native mechanisms for continuous numerical attributes or first-order relational representations.

Coverage note — Omitted the secondary 10-attribute experiment with irrelevant features as its results only confirmed standard symbolic generalization behavior without introducing distinct mechanisms.

References

1.Aha, D., Kibler, D., & Albert, M.K. (1991). Instance-Based Learning Algorithms. Machine Learning, 6(1), 37–66.
2.Angluin, D. (1988). Queries and Concept Learning. Machine Learning, 2(4), 319–342.
3.Blumer, A., Ehrenfeucht, A., Haussler, D., & Warmuth, M. (1989). Learnability and the Vapnik-Chervonenkis Dimension. Journal of the ACM, 36(4), 929–965.
4.Cost, S., & Salzberg, S. (1993). A Weighted Nearest Neighbor Algorithm for Learning with Symbolic Features. Machine Learning, 10(1), 57–78.
5.Grossberg, S. (1987). Competitive Learning: From Interactive Activation to Adaptive Resonance. Cognitive Science, 11, 23–63.
6.Hecht-Nielsen, R. (1990). Neurocomputing. Reading, MA: Addison-Wesley.
7.Helmbold, D.P., Littlestone, N., & Long, P.M. (1992). Apple-Tasting and Nearly One-Sided Learning. Proceedings of the 33rd Annual Symposium on the Foundations of Computer Science (pp. 493–502). IEEE Computer Science Press.
8.Helmbold, D.P., & Long, P.M. (1991). Tracking Drifting Concepts Using Random Examples. Proceedings of the Fourth Annual Workshop on Computational Learning Theory (pp. 13–23). San Mateo, CA: Morgan Kaufmann.
9.Helmbold, D.P., & Long, P.M. (1994). Tracking Drifting Concepts by Minimizing Disagreements. Machine Learning, 14(1), 27–45.
10.Kaelbling, L.P. (1994). Associative Reinforcement Learning Functions in k-DNF. Machine Learning, 15(3), 279–298.
11.Kilander, F., & Jansson, C.G. (1993). COBBIT — A Control Procedure for COBWEB in the Presence of Concept Drift. Proceedings of the Sixth European Conference on Machine Learning (pp. 244–261). Berlin: Springer Verlag.
12.Krizakova, I., & Kubat, M. (1992). FAVORIT: Concept Formation with Ageing of Knowledge. Pattern Recognition Letters, 13, 19–25.
13.Kubat, M. (1989). Floating Approximation in Time-Varying Knowledge Bases. Pattern Recognition Letters, 10, 223–227.
14.Kubat, M. (1991). Conceptual Inductive Learning: The Case of Unreliable Teachers. Artificial Intelligence, 52, 169–182.
15.Kubat, M. (1992). A Machine Learning Based Approach to Load Balancing in Computer Networks. Cybernetics and Systems, 23, 389–400.
16.Kubat, M. (1993). Flexible Concept Learning in Real Time Systems. Journal of Intelligent and Robotic Systems, 8, 155–171.
17.Kuh, A., Petsche, T., & Rivest, R.L. (1991). Learning Time-varying Concepts. Advances in Neural Information Processing Systems, 3 (pp. 183–189). San Mateo, CA: Morgan Kaufmann.
18.Kuh, A., Petsche, T., & Rivest, R.L. (1992). Incrementally Learning Time-varying Half-planes. Advances in Neural Information Processing Systems, 4 (pp. 920–927). San Mateo, CA: Morgan Kaufmann.
19.Langley, P., Gennari, J.H., & Iba W. (1987). Hill-Climbing Theories of Learning. Proceedings of the Fourth International Workshop on Machine Learning (pp. 312–323). Los Altos, CA: Morgan Kaufmann.
20.Langley, P., & Iba, W. (1993). Average-case Analysis of a Nearest-Neighbor Algorithm. Proceedings of the Thirteenth International Conference on Artificial Intelligence (pp. 889–894). San Mateo, CA: Morgan Kaufmann.
21.Lebowitz, M. (1987). Experiments with Incremental Concept Formation: UNIMEM. Machine Learning, 2(2), 103–138.
22.Maass, W. (1991). On-line Learning with an Oblivious Environment and the Power of Randomization. Proceedings of the Fourth Annual Workshop on Computational Learning Theory (pp. 167–175). San Mateo, CA: Morgan Kaufmann.
23.Markovitch, S., & Scott, P.D. (1988). The Role of Forgetting in Learning. Proceedings of the Fifth International Conference on Machine Learning (pp. 450–465). San Mateo, CA: Morgan Kaufmann.
24.Michalski, R.S. (1983). A Theory and Methodology of Inductive Learning. In R.S. Michalski, J.G. Carbonell, and T.M. Mitchell (eds.), Machine Learning: An Artificial Intelligence Approach, (Vol. 1). San Mateo, CA: Morgan Kaufmann.
25.Minton, S. (1988). Quantitative Results Concerning the Utility of Explanation-Based Learning. Proceedings of the Seventh National Conference on Artificial Intelligence (pp. 564–569). San Mateo, CA: Morgan Kaufmann.
26.Mitchell, T.M., Keller, R.M., & Kedar-Cabelli, S.T. (1986). Explanation-Based Generalization: A Unifying View. Machine Learning, 1(1), 47–80.
27.Moore, A.W. (1992). Fast, Robust Adaptive Control by Learning only Forward Models. Advances in Neural Information Processing Systems, 4 (pp. 571–578). San Mateo, CA: Morgan Kaufmann.
28.Pawlak, Z. (1982). Rough Sets. International Journal of Computer and Information Sciences, 11, 341–356.
29.Salganicoff, M. (1993a). Density-Adaptive Learning and Forgetting. Proceedings of the 10th International Conference on Machine Learning (pp. 276–283). San Mateo, CA: Morgan Kaufmann.
30.Salganicoff, M. (1993b). Exploiting Forgetting Algorithms for Memory-Based Learning (Technical Report MS-CIS-93-80). Philadelphia, PA: University of Pennsylvania, Department of Computer and Information Science.
31.Salzberg, S. (1991). A Nearest Hyperrectangle Learning Method. Machine Learning, 6(3), 251–276.
32.Schlimmer, J.C., & Granger, R.H. (1986). Incremental Learning from Noisy Data. Machine Learning, 1(3), 317–354.
33.Tambe, M., & Newell, A. (1988). Some Chunks are Expensive. Proceedings of the Fifth International Conference on Machine Learning (pp. 451–458). San Mateo, CA: Morgan Kaufmann.
34.Turney, P.D. (1993). Robust Classification with Context-Sensitive Features. Proceedings of the 6th Intl. Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems (pp. 268–276). Edinburgh, Scotland.
35.Valiant, L. (1984). A Theory of the Learnable. Communications of the ACM, 27(11), 1134–1142.
36.Widmer, G. (1994). Combining Robustness and Flexibility in Learning Drifting Concepts. Proceedings of the 11th European Conference on Artificial Intelligence (pp. 468–472). Chichester: Wiley & Sons.
37.Widmer, G., & Kubat, M. (1992). Learning Flexible Concepts from Streams of Examples: FLORA2. Proceedings of the 10th European Conference on Artificial Intelligence (pp. 463–467). Chichester: Wiley & Sons.
38.Widmer, G., & Kubat, M. (1993). Effective Learning in Dynamic Environments by Explicit Context Tracking. Proceedings of the Sixth European Conference on Machine Learning (pp. 227–243), Berlin: Springer Verlag.

Citation

MLA

Widmer, G., and M. Kubat. “Learning in the Presence of Concept Drift and Hidden Contexts”. Machine Learning, vol. 23, no. 1, 1996, pp. 69–101, https://doi.org/10.1023/A:1018046501280.

APA

Widmer, G., & Kubat, M. (1996). Learning in the Presence of Concept Drift and Hidden Contexts. Machine Learning, 23(1), 69–101. https://doi.org/10.1023/A:1018046501280

Chicago

Widmer, G., and M. Kubat. 1996. “Learning in the Presence of Concept Drift and Hidden Contexts”. Machine Learning 23 (1): 69–101. https://doi.org/10.1023/A:1018046501280.

Harvard

Widmer, G. and Kubat, M. (1996) “Learning in the Presence of Concept Drift and Hidden Contexts”, Machine Learning, 23(1), pp. 69–101. Available at: https://doi.org/10.1023/A:1018046501280.

Vancouver

1. Widmer G, Kubat M (1996) Learning in the Presence of Concept Drift and Hidden Contexts. Machine Learning 23:69–101

BibTeX

@article{Widmer_1996, title={Learning in the Presence of Concept Drift and Hidden Contexts}, volume={23}, ISSN={1573-0565}, url={http://dx.doi.org/10.1023/A:1018046501280}, DOI={10.1023/a:1018046501280}, number={1}, journal={Machine Learning}, publisher={Springer Science and Business Media LLC}, author={Widmer, Gerhard and Kubat, Miroslav}, year={1996}, month=Apr, pages={69–101} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Learning in the Presence of Concept Drift and Hidden Contexts

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — FLORA Three-Set Hypothesis Representation

Knowl 2 — Basic FLORA Incremental Learning and Forgetting Algorithm

Knowl 3 — FLORA2 Window Adjustment Heuristic

Knowl 4 — FLORA3 Context Storage and Reuse Mechanism

Knowl 5 — FLORA4 Confidence Interval-Based Descriptor Maintenance

Knowl 6 — Piecewise Linear Transition Model for Gradual Drift Speed

Knowl 7 — Classification Noise Robustness of FLORA4 compared to FLORA2, FLORA3, and IB3

Knowl 8 — Concept Reuse Acceleration in FLORA3 under Recurring Hidden Contexts

Knowl 9 — Inverse Relationship Between Drift Extent and Recovery Time

Knowl 10 — Parameter Sensitivity and Expressiveness Limitations in FLORA

References

Citation

BibTeX

Access the Paper