Paper deep dive
Learner-based Concept Drift Detection: Analysis and Evaluation
Md Moman Ul Haque Khan, Samira Sadaoui
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 6/21/2026, 5:15:52 AM
Summary
This paper provides a comprehensive theoretical analysis and empirical evaluation of learner-based concept drift detection algorithms. It categorizes concept drift into types (Real, Virtual, Mixed) and transition speeds (Sudden, Gradual, Incremental, Recurrent). The study focuses on learner-based detection methods, classifying them into Statistical Process Control (SPC), Window-based, and Ensemble-based approaches. The authors evaluate these methods using both synthetic and real-world streaming datasets to understand their performance in various non-stationary environments.
Entities (9)
Relation Signals (5)
EDDM â isatypeof â SPC-based Detection
confidence 100% ¡ 4.1 EDDM... is an improvement of the DDM... SPC-based Detection
ADWIN â isatypeof â Window-based Detection
confidence 100% ¡ 5.1 ADWIN [12]... Window-based Detection
AWE â isatypeof â Ensemble-based Detection
confidence 100% ¡ 6.1 AWE [56]... Ensemble-based Detection
Real Drift â isatypeof â Concept Drift
confidence 100% ¡ The types of drift are categorized based on the aspects of the probability distribution that is changing.
Sudden Drift â isatypeof â Concept Drift
confidence 100% ¡ Transitions represent the speed of change from the old concept to the new one.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Machine learning algorithms deployed for evolving streaming environments must handle the non-stationary data distributions, commonly referred to as concept drift. The presence of concept drift poses a major challenge for many real-world applications because it can severely degrade their predictive performance, hindering their ability to support robust decision-making. Consequently, the timely and efficient detection of drift events is critical for sustaining high accuracy over time. This study examines theoretically the concept drift characteristics and numerous drift detection algorithms across several categories. Furthermore, we evaluate their performance on both synthetic and real-world datasets exhibiting diverse streaming scenarios and drift characteristics, such as abrupt and gradual changes. This study aims to enhance understanding of the complex notion of concept drift characteristics and behavior of drift detectors, along with their applicability to diverse contexts.
Tags
Links
- Source: https://arxiv.org/abs/2606.20216v1
- Canonical: https://arxiv.org/abs/2606.20216v1
Trouble viewing inline? Open PDF directly â
Full Text
68,803 characters extracted from source content.
Expand or collapse full text
Learner-based Concept Drift Detection: Analysis and Evaluation Md Moman Ul Haque Khan and Samira Sadaoui Department of Computer Science, University of Regina, Canada MdMomanUlHaque.Khan@uregina.ca, Samira.Sadaoui@uregina.ca Contents 1 Introduction3 2 Characteristics of Concept Drift4 2.1 Formal Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 2.2 Types of Drifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 2.3 Transitions of Drifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 3 Concept Drift Detection6 3.1 Active vs. Passive Detection . . . . . . . . . . . . . . . . . . . . . . . . . .6 3.2 Learner-based vs. Distribution-based Detection . . . . . . . . . . . . . . .7 3.3 Learner-based Detectors . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 3.4 A Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9 4 SPC-based Detection9 4.1 EDDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 4.2 FHDDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 4.3 RDDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 4.4 EWMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 4.5 FTDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 5 Window-Based Detection12 5.1 ADWIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 5.2 KSWIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 5.3 MDDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 5.4 FPDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 5.5 WSTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1 arXiv:2606.20216v1 [cs.LG] 18 Jun 2026 5.6 D3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 6 Ensemble-based Detection15 6.1 AWE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 6.2 AUE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 6.3 ARF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 6.4 DWM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 7 Experiments Setup18 7.1 Implementation Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . 18 7.2 Streaming Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 8 Comparison of SPC-based Detectors19 9 Comparison of Window-based Detectors20 10 Comparison of Ensemble-based Detectors21 11 A Summary22 12 Conclusions23 Abstract: Machine learning algorithms deployed for evolving streaming environments must handle the non-stationary data distributions, commonly referred to as concept drift. The presence of concept drift poses a major challenge for many real-world ap- plications because it can severely degrade their predictive performance, hindering their ability to support robust decision-making. Consequently, the timely and efficient detec- tion of drift events is critical for sustaining high accuracy over time. This study examines theoretically the concept drift characteristics and numerous drift detection algorithms across several categories. Furthermore, we evaluate their performance on both synthetic and real-world datasets exhibiting diverse streaming scenarios and drift characteristics, such as abrupt and gradual changes. This study aims to enhance understanding of the complex notion of concept drift characteristics and behavior of drift detectors, along with their applicability to diverse contexts. Keywords:Concept drift types, Transition speeds, Sudden and gradual drifts, Concept drift detection, Learner-based detection, SPC methods, Windowing methods, Ensemble methods, Synthetic stream datasets, Implementation frameworks. 2 1 Introduction Nowadays, most applications evolve over time, such as fraud detection, network in- trusion, health monitoring, financial markets, predictive maintenance and environment monitoring. For example, in e-payment systems, fraudsters often devise new strategies to manipulate the systemâs vulnerabilities, and legitimate buyers adopt new payment methods (e.g., using new mobile wallets) [36, 10]. Regarding the environment monitor- ing systems, sensors can result in shifts due to new pollution causes [60, 38]. Health monitoring devices may also encounter drift because a patientâs lifestyle, age and health status (health is improving or deteriorating) change over time [34]. Industrial equipment also experiences drift as it ages and degrades, and monitoring models must detect fail- ures early [55, 40]. These non-stationary applications process an influx of data produced non-stop, often quickly, and where data can shift unpredictably. These changes lead to a divergence between training and operational data-distributions over time [2, 27]. Such changes in the class distribution and/or feature distribution, known as concept drift [24], often degrade the performance of predictive models [24]. Indeed, ML models trained on old data patterns become invalid and obsolete, leading to poor performance and false alarms. Concept drift occurs when the statistical properties of the incoming data change or the relationships between the independent variables (features) and target variable (class label) change. The evolving data-generating environments present memory and compu- tational challenges, making it critical to develop adaptive algorithms that can effectively address them [27]: 1) Data arrives at high speed, necessitating algorithms with fast data and drift analyzing capabilities, and 2) Data are continuous, making it impractical to store all the data, which requires algorithms to process each sample individually, or select the most relevant data or summarize data on the fly. For each specific application, the model developers must anticipate possible changes in future data and choose the most effective strategies to adapt ML models to real-time changes in data properties, such as discarding outdated knowledge and retraining the models on the adjusted feature space. Our study provides a comprehensive survey of concept drift within the supervised learning setting, with a particular focus on data stream classification. To support a clearer understanding of the complex nature of concept drift, we first examine its key characteristics through several concrete examples, including different drift types, such as real and virtual drift, as well as different rates of change, such as abrupt, gradual, incremental and recurring drift. The focus is placed on learner-based drift detection methods, which are the most widely studied and applied approaches. These methods typically monitor the behavior or performance of a learning model over time in order to identify significant changes that may indicate the presence of drift. We outline a general algorithmic framework for learner-based detection and examine its main categories: Sta- tistical Control Process (SPC), Window-based and Ensemble-based. For each category, we review a selection of representative drift detectors (with a tally of 15 methods). In addition to the survey part, we conduct a thorough empirical evaluation of the reviewed methods on several synthetic (under gradual and sudden drifts) and real-world 3 data streams. We choose artificial datasets because the drift locations are known a apri- ori, which allows for a more precise assessment of each methodâs ability to detect both sudden and gradual drifts. We also consider real-world datasets to evaluate the practical behavior of the methods under realistic conditions, where drifts may be noisy or difficult to identify. By combining a robust review with an extensive experimental evaluation and comparison, our study aims to provide researchers with a clearer and more practical understanding of learner-based detection methods. The remainder of the paper is structured as follows. Section 2 presents the formal definitions of concept drift and discusses its main characteristics using real-world exam- ples. Section 3 examines different classifications of drift detection methods, including active vs. passive, online vs. block, learner-based vs. distribution-based, and SPC- based vs. window-based vs. ensemble-based. Sections 4, 5 and 6 present the theoretical foundations of SPC-, window- and ensemble-based detectors, respectively. Section 7 describes the existing implementation frameworks for these various methods and also introduces the stream datasets used in the experimental evaluation. Sections 8, 9, 10 and 11 evaluate and compare the performance of SPC-, window- and ensemble-based methods. Section 12 concludes our work. 2 Characteristics of Concept Drift 2.1 Formal Definitions Concept drift occurs when learning from dynamic data streams, defined as continuous and limitless sequence of samples, [(X 0 ,y 0 ),...., (X i ,y i )], where X is a multi-dimensional feature vector with a corresponding label y [54]. The Bayesâ decision theorem allows to compute the probability that X is an instance of class y [17]: P (y|X) = P (X|y)P (y) P (X) Here, P (X|y) denotes the likelihood of the input features given the target label, P (y) is the prior probability distribution of the output variable, and P (X) is the uncondi- tional probability distribution of the input features. In predictive modeling, a concept refers to the specific joint probability of input features and target class, P (X,y), which encompasses the prior class probability P (y) and class-conditional probability P (X|y) as follows [15, 27]: P (X,y) = P (y)P (X|y) Concept drift can be defined as a change in the joint probability distribution between two distinct time steps t and t + w. Here, t represents a specific point or interval in time, and w denotes the window over which the change in data distribution has been identified [15, 27]. So, concept drift occurs when P t (X,y)̸= P t+w (X,y) [15, 27]. 4 2.2 Types of Drifts The types of drift are categorized based on the aspects of the probability distribution that is changing. A. Real Drift: Known as actual or true drift, this type occurs when there a change in the posterior probability P (y|X), which can affect the decision boundary of ML models. This drift can be a result of changes in the class distribution P (y) and likelihood distribution P (X|y). It indicates a fundamental change in the concept the model is trying to predict, which over time renders the model less effective [15, 27]: P t (y|X) ̸= P t+w (y|X). An example: We assume we have a model that predicts the likelihood of an illness based on the symptoms observed (X). In this case, real drift refers to changes in the conditional probability of the illnesses given the symptoms, i.e. P (Disease|X). For instance, the changes are due to an update in medical knowledge or emergence of new illness strains. B. Virtual Drift: Known as covariate drift, this type refers to changes in the feature- value distribution P (X), without altering the relationship between the input features and the label i.e., P (y|X). With this type, the model might encounter samples that are underrepresented or differ from those in the training dataset, leading to accuracy issues [15, 27]: P t (X)̸= P t+w (X). An example: In predicting weather conditions, suppose the model was trained with data from the autumn season. If the model starts receiving data from winter (without any change in P (Weather|X)), this would be an instance of virtual drift. C. Mixed Drift: Both real and virtual drifts can happen simultaneously in real-world scenarios, involving changes in both the prior probability of classes and probability of features [61, 59]: P t (y|X)̸= P t+w (y|X) and P t (X)̸= P t+w (X). An example: In credit card fraud detection, we usually encounter real and virtual shifts, such as the consumerâs spending habits change during the holiday season (real drift) and fraudsters change their tactic strategies (virtual drift). 2.3 Transitions of Drifts Transitions represent the speed of change from the old concept to the new one. They can be quantified by the number of samples over which the drift occurs until the new concept is established. These transitions are essential for understanding how to adapt the decision models to maintain accuracy as data evolves. There are four transition types [17, 35], and for each type, concrete examples are provided based on studies such as [15, 36, 35]. ⢠A. Sudden Drift: Occurs when the data distribution changes abruptly from the old concept to a new one at a precise timestamp, which degrades instantly the decision models. Abrupt drifts that can drastically impact the learned models. 5 Examples: 1) A sudden change in consumer behavior due to unexpected events, like the COVID-19 pandemic and new market regulations, 2) a quick equipment failure or 3) unforeseen weather conditions. ⢠B. Gradual Drift: Happens when the data patterns change progressively from the previous concept to the new one. This type shows a longer transition phase that involves a mixture of the old and new concepts. The new concept becomes more predominant over time. Examples: 1) User preferences for a particular service or product can change slowly due to evolving trends, 2) the quality of machinery can degrade slowly over time, and 3) a patientâs health may gradually change due to aging or the progression of a medical condition over the years. ⢠C. Incremental Drift: Occurs when the current concept replaces the past con- cept slowly, by involving intermediate concepts (may not be statistically signifi- cant). Some researchers consider incremental drift as a variant of gradual drift, however, what sets incremental drift apart from gradual drift is the absence of a distinct boundary that separates the old and new concepts. Examples: 1) The slow change in climate patterns (like in temperatures and rainfalls) progressively affect agricultural products, and 2) fraudsters in e-payment systems gradually adopt new tactics by shifting from simple techniques (like stolen credit cards) to modern ones (like virtual cards and bots). ⢠D. Recurrent Drift: Happens when a concept that was encountered earlier reappears after some time has passed. While this shares similarities with gradual drift in that alteration of two concepts, the key distinction is that the old concept reappears after some time interval in the recurring drift. Examples: 1) Seasonal changes in spending behavior, such as increased purchases during the holiday seasons, and 2) with the winter season comes health changes due to people becoming less active and having higher heart rates. 3 Concept Drift Detection 3.1 Active vs. Passive Detection Drift detection methods can be mainly categorized into two folds [23, 53]: active (with an explicit drift detection mechanism) vs. passive (continuous model adaptation). Active or informed detectors monitor data streams for drifts and activate some adaptation mecha- nisms when drifts have occurred, reducing FP and FN rates and saving memory and CPU resources [28, 4]. Recent examples of active detectors include: 1) probabilistic methods that specifically identifies real drift [42] and [43], 2) a sum-product N-based method that detects both real and virtual drifts [41], and 3) a cross-entropy-based method that identifies real drifts within noisy data streams [44] . On the other hand, passive meth- ods continuously update their models with incoming samples without any drift detection 6 because they consider that drifts may occur constantly or periodically. These methods can be handy in detecting gradual/incremental shifts, but they are time-consuming. An example is the incremental feature learning approach introduced in [49] that continu- ously adjusts a single N model to new data chunks. The model is composed of several interconnected sub-NNs, with a new sub-N optimally created for each incoming chunk. To prevent unbounded N growth, only the sub-NNs most relevant to the new data distribution are retained and re-combined to build the optimal model. 3.2 Learner-based vs. Distribution-based Detection Drift detectors can also be broadly classified into three main categories: learner-based (supervised), distribution-based (unsupervised) and hybrid. Each category has specific advantages and disadvantages [35, 17, 57, 61], as explained below. ⢠Learner-based: These methods detect drifts by monitoring the performance of the underlying classifiers, such as the error rates. They are further split into three groups: Statistical Process Control (SPC), Windowing Techniques and Ensemble Learning. The pros of these methods are: 1) Directly linked to the predictive performance of the base models, which makes them intuitive and easy to interpret, and 2) Effective in scenarios where the drifts directly impact the model accuracy. Their cons are: 1) May not detect drifts that do not immediately affect model performance, such as changes in features, 2) Require data to be labeled, and 3) Can be sensitive to noise and random fluctuations in data. ⢠Distribution-based: These methods monitor changes in the data distribution itself over different timestamps to check whether the current and historical data windows come from the same distribution or not. They often use statistical tests (like Kullback-Leibler divergence and Kolmogorov-Smirnov Test) to determine if the data distribution has changed significantly. Their pros are: 1) Detect drifts that do not immediately affect model performance, 2) Often more robust to noise in data, and 3) Do not require data to be labeled. On the other hand, their cons are: 1) May require a large amount of data to detect drifts accurately, 2) More prone to false alarms, and 3) Can be computationally intensive due to continuous statistical testing. ⢠Hybrid: Hybrid detectors combine elements of learner and distribution-based methods to leverage the strengths of both methods and provide more robust drift detection mechanisms. Their pros are: 1) Potentially more robust and versatile in different scenarios, and 2) Can dynamically switch between performance and distribution-based detection based on the context. Their cons are : 1) Can be more complex to implement and tune, and 2) May incur higher computational costs due to the dual nature of detection. 7 3.3 Learner-based Detectors This category explicitly manages drift by incorporating mechanisms to forget old data and retrain or adjust the underlying models to the newly detected concept. To distin- guish drift from noise, the new concept must remain stable for some time [15]. The detectors are usually designed to be independent of specific ML algorithms, as they can adopt various types of learners. The general form of the learner-based detection meth- ods is summarized in Algorithm 1 using the prequential training scheme (test and then train). The algorithm is online, so that the drift is identified in real time. The selected classifier is first pre-trained on an initial robust labeled dateset. The classifier then re- ceives sequentially and continuously a stream of samples and predicts their labels. Next, it uses a statistical test method to detect a drift by comparing a drift metric to a prede- fined statistical threshold. For example, when the error rate (the discrepancy between predicted and true labels) increases significantly, this determines a drift. Other examples would be comparing past accuracy or prediction probability (confidence score) with the accuracy or probability on the new data. Some detectors use a warning threshold to signal an impending drift and monitor the samples more closely. Once a significant change is detected, the classifier retrains on recent data, typically stored in a sliding window or buffer. When the buffer reaches the limit, it forgets old data. Other methods select the most representative samples from the buffer or use all the samples but the new data have higher weights so that the detection model focuses less on old data [2]. Algorithm 1 General Algorithm for Learner-based Drift Detection 1: Inputs: classier, intialDataset, dataStream (continous), statThreshold, RDSize (sliding window size) 2: Pre-train classifier with intialDataset 3: recentData â empty /*window of recent data*/ 4: for each newSample in dataStream do 5:Add newSample to recentData 6: if (length of recentData > RDSize) then 7:Remove oldest sample from recentData 8: end if 9:Predict sampleâs label with classifier 10:Calculate driftMetric on prediction 11: if (driftMetric > statThreshold) /*significant difference*/ then 12:Display âDrift Detectedâ 13:Train classifier on recentData with ground truth 14:recentData â empty 15: end if 16: end for 8 3.4 A Summary Table 3.1 compares numerous detectors (those selected for the experiments) based on their strategies: (1) Active (with an explicit drift detection mechanism) vs. Passive (con- tinuous model adaptation), (2) Online (processes each sample once and then discarded it) vs. Block-based (processes a data chunk at once), and (3) types of learner-based methods: SPC, Window techniques and Ensemble methods. MethodYearPassive vs ActiveOnline vs BlockCategory FTDD [14]2018ActiveOnlineSPC RDDM [7]2017ActiveOnlineSPC FHDDM [46]2016ActiveOnlineSPC EWMA [48]2012ActiveOnlineSPC EDDM [5]2006ActiveOnlineSPC KSWIN [47]2020ActiveOnlineWindow FPDD [14]2018ActiveOnlineWindow WSTD [6]2018ActiveOnlineWindow MDDM [45]2018ActiveOnlineWindow ADWIN [12]2007ActiveOnlineWindow D3 [20]2015ActiveOnlineWindow ARF [19]2017PassiveOnlineEnsemble AUE [13]2011PassiveBlock-basedEnsemble DWM [30]2007ActiveOnlineEnsemble AWE [56]2003ActiveBlock-basedEnsemble Table 3.1: A Summary of Learner-based Drift Detection Algorithms 4 SPC-based Detection SPC detectors treat a modelâs performance as a controlled process [15, 39]. They employ statistical tests to detect deviations from expected behavior or exceedance of predefined control limits. By monitoring the evolution of a modelâs performance, SPCs assess the quality of the learning process [15, 39]. When performance deteriorates or crosses a statistically significant threshold, it indicates concept drift. In such cases, SPCs may trigger classifier retraining or apply alternative strategies to address the drift. This approach enables real-time monitoring and facilitates prompt responses to changes in the underlying data. There are different SPC procedures, such as those defined in DDM, EDDM, RDDM, FHDDM, KAPPA, EWMA, STEPD, FTDD and ACDDM. We describe the most popular methods in the following sections. 9 4.1 EDDM Early Drift Detection Method (EDDM) [5] is an improvement of the DDM [18], designed to better detect gradual (moderate) changes. Instead of considering the number of errors as in DDM, EDDM focuses on the distance between two successive errors. At timestamp t, EDDM calculates [5]: ⢠p t , which is the average distance between two consecutive errors at time t â˘ Ď t , which is the standard deviation of the distances at time t ⢠p max and Ď max , which track the maximum observed performance and take the values of p t and Ď t when (p t + 2Ď t ) reaches its historical maximum value The study [5] defines the warning and drift levels using Îą as the threshold for the warning zone and β as the threshold for the drift zone: ⢠Warning level: (p t + 2Ď t )/(p max + 2Ď max ) < Îą, indicating a possible change ⢠Drift level: (p t + 2Ď t )/(p max + 2Ď max ) < β, signaling a confirmed drift The paper [5] recommends setting Îą and β to 0.95 and 0.90 respectively. When the warning level is detected, the samples are stored in preparation for drift localization. When the drift level is reached, the classifier along with p max and Ď max are reset. In this case, the model can use the data collected in the warning zone together with recent post- drift data to train or update the model. When changes occur gradually at a moderate pace, EDDM is a better option than DDM. However, DDM is a better option when dealing with sudden changes. The limitation of EDDM can be pointed out as it is noise sensitive, which makes it more prone to false positives [39]. 4.2 FHDDM Fast Hoeffding Drift Detection Method (FHDDM) [46] uses the Hoeffding inequality [26] to detect changes in data streams. It continuously monitors the prediction errors and uses the Hoeffding bound to determine whether the error rate has increased significantly or not. FHDDM computes the mean error in this window and compares it with the minimum error observed so far, and uses the following Hoeffding inequality: p(| Ě X â Îź|⼠ξ)⤠2e â2nÎľ 2 where Ě X is the average of the error rates of the current window, Îź is the minimum error rate assuming no drift, n is the size of the sliding window, and Îľ is the acceptable deviation from the minimum error rate. The method flags a drift when the deviation | Ě Xâ Îź| is large enough, i.e. greater than or equal to the calculated threshold Îľ, derived from the Hoeffding bound. 10 4.3 RDDM Reactive Drift Detection Method (RDDM) enhances DDM by dynamically adjusting the detection threshold based on recent error rates [3, 9]. RDDM utilizes the following strategy, based on the calculated error rate and standard deviation [9]: ⢠Warning level: p t +Ď t ⼠p min + (2âĎ min ), for a possible drift; p min is the historical minimum error rate and Ď min is the standard deviation of p min ⢠Drift level: p t + Ď t ⼠p min + (3â Ď min ), for a detected drift RDDM continuously recalculates warning and drift statistics, discarding outdated samples and focusing on recent ones. It identifies concepts that remain active for an extended period or with a long warning window. This adaptability allows to effectively identify both sudden and gradual drifts while minimizing false alarms [6, 53]. Compared to DDM, RDDM achieves higher accuracy in most scenarios, particularly for gradual drifts, by detecting more drifts and identifying them earlier. However, its increased complexity and parameter sensitivity may require careful calibration for optimal perfor- mance in specific applications [15]. 4.4 EWMA The exponentially weighted moving average (EWMA) was proposed in [48]. EWMA for drift detection is a statistical process that identifies drift by monitoring the error rate of a streaming classifier [8, 14].The standard deviation of the EWMA estimator z t is [35]: Ď 2 z t = r p 0 (1â p 0 ) Îť 2â Îť (1â (1â Îť) 2t ) Here, z t = (1â Îť)z tâ1 + Îťx t , p 0 is the baseline error rate under no-drift conditions, and Îť is a smoothing parameter. According to EWMA, drift is detected if [15]: z t > Îź 0 + LĎ z t . Here, L is the control limit, which states the distance z t must deviate from Îź 0 before a change is flagged. Set- ting the appropriate smoothing factor (Îť) is crucial for achieving optimal performance. Îť = 0.2 has been suggested by [35]. A high Îť value increases sensitivity to recent changes but may make the method susceptible to noise. A low Îť value can lead to delayed drift detection[48]. The main advantage of EWMA is its ability to detect gradual and abrupt drifts [3]. It is also computationally efficient, with an overhead of O(1) [48]. 4.5 FTDD The Fisher Test Drift Detector (FTDD) is one of three drift detection algorithms pro- posed by [14], based on an efficient implementation of Fisherâs Exact Test [16]. FTDD is based on the Sequential Testing with Estimation of Proportion Difference (STEPD) algorithm [39], but addresses its limitations when dealing with imbalanced data [51] Unlike STEPD, which uses a test of equal proportions, FTDD employs Fisherâs Exact 11 test in all comparisons to detect concept drift [6, 8, 15]. The p value calculation is a key feature of FTDD, contributing to its effectiveness in accurately detecting drifts [14]: p value = |w r +w p | |w r | |w p | Ă |c r +c p | |c r | |c p | p value = pvalueĂ constF Ă 2 Where w r is the number of errors in the recent window, w p is the proportionally calculated number of errors in past window, c r is the number of correct predictions in the recent window, c p is the proportionally calculated number of correct predictions in past window. constF is a constant factor derived from the factorial calculations for the window size (w). It helps to optimize the p value computation by pre-calculating and storing a common factor. Here, the probability of observing the specific distribution of errors and correct predictions in the two windows (recent and past) is calculated given the assumption that no drift has occurred [14]. The warning is triggered when the p value falls below the threshold Îą w = 0.005 and a drift is confirmed when the p value is typically less than Îą d = 0.003. 5 Window-Based Detection Windowing detectors handle drifts by monitoring the model performance or other sta- tistical measures over two data windows (of a fixed or dynamic size) and checking for significant discrepancies between them [15, 51]. They treat the classifierâs prediction results as the data stream. They assess the chosen metric on each window and flag drift when the difference exceeds a predefined threshold or a threshold derived from a statisti- cal test [15, 51]. The dynamic windowing technique adjusts the size based on the driftâs length, retaining samples until the drift occurs. It allows for flexibility without the need to set a predefined window size, which is a difficult task. The decision for window size is crucial because large windows can increase accuracy but may miss the detection of rapid drifts, and small windows are more effective in detecting sudden drifts but might not capture gradual changes. There are different windowing strategies, as described in the following selected methods: ADWIN, KSWIN, MDDM, FPDD, WSTD and D3. 5.1 ADWIN ADaptive WINdowing (ADWIN) [12] maintains a sliding window w containing the most recent errors and tests, after every new instance, whether the average error in an older window w p differs significantly from that in a newer window w r . ADWIN detects drift by comparing the mean error rates (Îź w p for w p , Îź w r for w r ) of the two sub-windows. A drift is signaled when the difference in error rates exceeds a threshold derived from the Hoeffding bound [51, 15]: |Îź w p â Îź w r | > Îľ, where Îľ is the threshold calculated based on the Hoeffding bound [12]: 12 Îľ = r 1 2m ln( 4|W| δ ) Where m is the harmonic mean of the lengths of sub-windows w r and w p , |w| is the length of the entire window w, δ is a predefined confidence parameter, and ln() is the natural logarithm. ADWIN has been shown to be effective in detecting both abrupt and gradual concept drifts [1]. 5.2 KSWIN KolmogorovâSmirnov Windowing (KSWIN) [47] maintains two sliding windows of error observations, an older window, w p , and a recent window, w r [44, 40]. After each new prediction, the sup-norm KS (KolmogorovâSmirnov) distance, dist(w r ,w c ) is computed for sample x, where sup-norm distance is the maximum point-wise gap between two functions, making it highly sensitive to even a single large deviation and thus effective for detecting distribution changes [47]: dist(w p ,w r ) = sup|f w p (x)â f w r (x)| Here, f (x) = 1 n P n i=1 I (ââ,x) (x i ) and I (ââ,x) (x i ) is an indicator function. A drift is signaled by KSWIN when the calculated distance exceeds a threshold defined by the formula below [44, 47]: dist(w p ,w r ) > r â ln(Îą) n Where Îą is the significance level representing the probability of incorrectly rejecting the null hypothesis. Common values for Îą are 0.05, 0.01, or even smaller depending on the desired level of confidence, and n is the size of the two sub-windows [47]. 5.3 MDDM McDiarmid Drift Detection Method (MDDM) [45] uses the McDiarmidâs inequality intro- duced in [37]. MDDM operates under the assumption that in a streaming environment, recent samples are more relevant than older ones. To reflect this, MDDM uses a sliding window with a weighting scheme that prioritizes the most recent samples [45]. MDDM slides a window of size n over the prediction results [22]. This window stores binary values, where 1 represents an incorrect prediction and 0 represents a correct prediction [15]. Each element in the window is associated with a weight, where w i < w i+1 . While inputs are processed, the current weighted average of the elements of the window is cal- culated, i.e. Îź t , as well as the maximum weighted mean observed so far, i.e. Îź m . When Îź m < Îź t then Îź m = Îź t [51, 45]. Hence, with MDDM, concept drif is declared when [45]: (Îź m â Îź t )⼠ξ 13 where Îľ is the threshold calculated with the McDiarmidâs inequality: Îľ = r P n i=1 v 2 i 2 â ln( 1 δ ) Here, δ is the confidence level, and v i is the weight of the i-th element, calculated as: v i = w i P n i=1 w i [45]. 5.4 FPDD The Fisher Proportions Drift Detector (FPDD)[14] is designed to enhance the perfor- mance of the Statistical Test of Equal Proportions(STEPD), particularly when dealing with small sample sizes [15]. It leverages Fisherâs Exact Test to provide a more robust statistical analysis in these scenarios [8]. This leads to more accurate drift detection when drifts occur rapidly. Similar to STEPD, FPDD compares two windows, the recent window (w r ) and past window (w p ), to analyze the data stream. The core distinction of FPDD lies in its conditional application of Fisherâs Exact Test. Suppose the number of errors or correct predictions in either of the two windows is smaller than five. In this case, FPDD uses Fisherâs Exact Test to assess the statistical significance of the difference in error rates between the two windows. In cases where the sample sizes are sufficient(5 or more), FPDD reverts to the standard test of equal proportions used by STEPD[51]. Based on the chosen statistical test (Fisherâs Exact Test or the test of equal proportions), FPDD calculates the p value (as computed in section 4.5), and uses it as follows [14]: ⢠Drift detected: if pvalue⤠ι d ⢠Warning signal: if Îą d < pvalue⤠ι w FPDD uses two similar threshold parameters like STEPD, the significance levels for the detection of drifts ( Îą d = 0.003 ) and warning ( Îą w = 0.05 ). 5.5 WSTD Wilcoxon Rank Sum Test Drift Detector (WSTD) [7], which closely resembles STEPD, detects drifts based on an efficient implementation of the Wilcoxon rank sum statistical test [58], whereas STEPD uses the test of equal proportions [6, 23]. WSTD requires setting a significance level (Îą) and applies the normal distribution to evaluate the null hypothesis. Given two sample set, n 1 and n 2 , they are combined in the ascending order as below [7]: ⢠Test statistic is z = (RâÎź R ) Ď R ⢠Population mean is Îź R = n 1 Ă (n 1 +n 2 +1) 2 ⢠Standard deviation is Ď R = q n 1 Ă n 2 Ă (n 1 +n 2 +1) 12 14 where R is the smallest sum of the ranks of both sample sets, n 1 is the size of the smallest sample set, and n 2 is the size of the largest sample set. The z value is used to reject the null hypothesis, and the pvalue (or obtained probability) is also required to find the z value [3, 7]. Similar to STEPD, WSTD monitors the base learnerâs predictions using two windows: a recent window and an older one. It relies on a statistical test to issue warnings or confirm drifts. WSTD incorporates three key parameters with default values: the recent window size (w = 30), the significance level for drift detection (Îą d = 0.003), and the significance level for warnings (Îą w = 0.05) [7]. 5.6 D3 Discriminative Drift Detector (D3) [20] detects changes by comparing recent and his- torical samples. D3 operates on the principle that a change in the underlying data distribution P (X) will manifest as a difference between recent and historical data. By training a classifier to distinguish between these sets, D3 can detect drifts when the classifier performs well, which suggests that the data distributions have diverged. D3 employs a sliding window, denoted as W , to store the most recent samples. The size of this window is determined by w(1 + Ď), where [20]: ⢠w is the number of âoldâ samples, which correspond to the historical data â˘ Ď is the fraction of ânewâ samples, relative to the size of the old data. This essentially controls the size of the new dataset within the window A drift is signaled when the classifier AUC score (trained to distinguish between the old and new sample) is greater than or equal to the threshold, θ, which is typically set between 0.5 and 1.0 [50, 20]. The performance of D3 can vary significantly based on the specific classifier and dataset being used [25]. Additionally, the methodâs capability is constrained to identifying only linear drift patterns within the feature space [21]. 6 Ensemble-based Detection Ensemble-based detectors are a robust approach to handling drifts because combining the outputs of several learners is more effective to changing data patterns. Many such detectors adopt the Weighted Majority Algorithm (WMA). The ensemble remains up to date with new concepts as follows [15, 1]: (1) a new classifier trained on the current data is added to the ensemble, (2) the learners are then weighted according to their performance with the current data, and (3) the least-weighted learners, deemed less effective, are removed from the ensemble. In addition to adding/removing learners, the ensemble may use incremental learning to adapt the learners. The ensemble approach is efficient in identifying gradual drifts while maintaining high levels of accuracy [15, 1], but comes with high computational and memory costs. There are numerous ensemble- based detectors, which can be categorized as follows: (1) methods that implement their own detection mechanisms, such as AWE, AUE, ACE and DWM, and (2) methods that 15 utilize an existing detector, such as ADWIN, which is used in algorithms like ARF. We describe these methods in the following sections. 6.1 AWE Accuracy Weighted Ensemble (AWE) [56] selects the most efficient classifiers using the mean squared error (MSE), which is calculated based on the probabilities of class as- signments. AWE identifies the top n classifiers based on their performance on the most recent data chunk [56]. It employs MSE for assigning weights: classifiers that exhibit an error rate meeting or exceeding a certain threshold are excluded. This method ensures that only the classifiers most aligned with the current data patterns are retained, thereby enhancing the ensembleâs overall accuracy. The weight (w i ) assigned to a classifier c i is calculated using the following formulas [15, 51, 30]: w i = MSE r â MSE i Here MSE r is the reference MSE calculated based on the class distribution of the current data chunk S n [56]: MSE r = X c p(c)(1â p(c)) 2 where p(c) is the probability of class c, which is estimated from the distribution of classes in S n . MSE i represents the error for the i th classifier, calculated over S n as follows [56]: MSE i = 1 S n X (x,c)ÎľS n (1â p i c (x)) 2 where p i c (x) is the probability that classifier i assigns sample x to class c. 6.2 AUE Accuracy Updated Ensemble (AUE) [13], an advancement over AWE, updates only the top k weighted classifiers that meet a certain accuracy threshold on the most recent data block, rather than updating all classifiers. Additionally, AUE utilizes a more straight- forward weighting function compared to AWE [51]. AUE incorporates online classifiers, typically Very Fast Decision Trees (VFDT) or Hoeffding Trees, by updating them in- dividually in addition to weight modifications. This adaptability ensures that in the absence of drift across chunks, the classifiers enhance their performance as though they were trained on a singular, extensive dataset [51]. Consequently, this flexibility permits a reduction in chunk size without compromising the ensembleâs accuracy. The weighting function employed in AUE is simplified as below[13]: w i = 1 MSE i + Îľ 16 where MSE i is computed in the same manner as in AWE, and Îľ is a small constant added to facilitate weighting calculations even when MSE i equals zero [13]. The en- semble size in AUE is typically fixed to a predetermined number, k, of classifiers [51]. Comparative experimental studies have demonstrated that AUE outperforms AWE in various datasets, except for one instance where both achieved comparable accuracy [13, 51]. 6.3 ARF The Adaptive Random Forest (ARF) algorithm [19] handles evolving data streams by incorporating several key strategies. ARF extends the traditional random forest algo- rithm to adapt to changes in the data distribution by replacing under-performing trees with new ones [29, 35]. ARF uses Hoeffding Trees (HT) as its base learners, which are a type of very fast decision tree. These trees are incrementally updated with each new sample, making them suitable for streaming data [19, 13]. ARF employs a resampling method based on online bagging [53]. Instead of growing each tree sequentially on dif- ferent subsets of data, as in traditional random forests, ARF simulates sampling with reposition using a Poisson distribution, specifically Poisson (Îť = 6) [19]. ARF handles concept drift through a combination of online bagging for resampling, drift detectors for each tree, background training of new trees, and weighted majority voting. These mech- anisms allow the model to maintain high accuracy when dealing with non-stationary data streams[23]. 6.4 DWM The Dynamic Weighted Majority (DWM)[30] extends the weighted majority algorithm [32] by dynamically adjusting the weights of its base classifiers and adding/removing classifiers based on their performance [31]. DWM is a passive adaptation method that does not use an explicit drift detection mechanism, but rather continuously adapts its base learner [27]. This differentiates it from methods that explicitly monitor error rates or other statistics to detect drift. With DWM, weights are adjusted after each new classified sample. This online ap- proach allows DWM to adapt to gradual changes more easily [15, 23]: ⢠If a base classifier predicts correctly, its weight is multiplicatively increased by a factor β > 1 i.e., w i (t + 1) = βâ w i (t) [51] ⢠If a base classifier predicts incorrectly, its weight is decreased multiplicatively by a factor Îą < 1 i.e., w i (t + 1) = Îąâ w i (t) [51] where w i (t) is the weight of classifier i at time t, and w i (t + 1) is the updated weight at time t + 1. Typical values for the parameters can be β = 1.1 and Îą = 0.9. 17 7 Experiments Setup We utilize both artificial and real-world datasets from the Harvard Dataverse [33], under abrupt and gradual scenarios, to assess the performance of drift detectors of different categories. These datasets present distinct data characteristics, providing an extensive testbed. When the drift is identified, the base learners, Naive Bayes and Hoeffding Tree, are retrained with the most recent samples to ensure that predictions remain accurate. We assess and compare these detectorâs performance using the AUC metric. AUC is useful for assessing performance degradation caused by drifts. For each detectorâs hyperparameters, we use their default values mentioned in earlier sections. In all the experiments, we set the window size to 50 and the ensemble size to 15. In the empirical tables, the best results are shown in red and the worst ones in blue. 7.1 Implementation Frameworks Table 7.1 provides the implementation frameworks of the 15 selected drift detectors. These frameworks support a broad collection of ML algorithms, such as classification, regression, clustering and drift detection. Here, NA indicates that there was no publicly available implementations found online. Therefore, we implemented the Python code with the assistance from Github Copilot. Table 7.1: Drift Detectors and their Implementation Frameworks URL FTDD [14]Git (Fisherâs Exact Test RDDM [7]Tornado FHDDM [46]River EWMA [48]Tornado EDDM [5]River KSWIN [47]River FPDD [14]Git (Fisherâs Exact Test) WSTD [6]NA MDDM [45]Tornado ADWIN [12]River ARF [19]River D3 [20]Git AUE [13]NA DWM [30]Scikit-multiflow AWE [56]Scikit-multiflow 18 7.2 Streaming Datasets Our empirical study utilizes synthetic and real datasets, summarized in Table 7.2, to evaluate the performance of various drift detectors. The six synthetic datasets (balanced binary class label and without noise) have been produced using three main generators, Random Tree Stream (RT), SINE and MIXED, by incorporating abrupt and gradual drifts [33]: ⢠RT-abrupt and RT-gradual, with the arrangement of 8873985678962563 ⢠Sine-abrupt and Sine-gradual, with the arrangement of 0123 ⢠Mixed-abrupt and Mixed-gradual, with the arrangement of 0101 Each dataset possess four distinct concepts and three drift locations [33]. The sudden datasets experience abrupt shifts at locations of 10,000, 20,000 and 30,000. The gradual datasets undergo gradual transitions at locations of 9,500, 20,000 and 30,500, with each transition having a drift width of 1,000 samples. On the other hand, we employ two real datasets (balanced binary class label), as presented in Table 7.2. The Electricity dataset ELEC2, collected from the Australian New South Wales electricity market, predicts whether the prices (set every five minutes) of electricity will increase or decrease [11]. The Intrusion Detection Evaluation dataset CIC-IDS2017 contains network traffic collected for five days, from Monday, July 3, 2017, at 9am to Friday, July 7, 2017, at 5pm [52]. DatasetSize#Features-Modality#Drifts RT a 40,0002-numerical3 Abrupt RT g 40,0002-numerical3 Gradual Sine a 40,0002-numerical3 Abrupt Sine g 40,0002-numerical3 Gradual Mixed a 40,0004-numerical3 Abrupt Mixed g 40,0004-numerical3 Gradual ELE45,3128-mixed (6 numerical, 2 categorical)Unknown CIC28,30319-mixed (16 numerical, 3 categorical)Unknown Table 7.2: Configurations of Synthetic and Real-world Datasets 8 Comparison of SPC-based Detectors For SPC detectors (cf. Table 8.1), on abrupt-drift datasets, NB shows modest per- formance, with FTDD achieves the highest AUC average, while the other methods all yield a lower average. With HT, performance is considerably higher across all detec- tors, confirming the advantage of the more expressive base learner. FTDD performs best, followed by EWMA, while RDDM shows the lowest performance. On gradual-drift 19 datasets, NB performance remains uniform across all SPC methods, where HT yields moderately higher values and EWMA and EDDM tied at the top. On real-world data streams, NB reveals a notable divergence among detectors: FTDD records the lowest average, whereas the remaining methods achieve a substantially higher average. For HT, the results follow a similar pattern: FTTD yields the lowest average AUC, whereas the other methods achieve comparable performance. When considering all dataset types together, overall HT surpasses NB. The results indicate that while FTDD holds an advantage on abrupt-drift data, EWMA and EDDM deliver the most reliable performance across all dataset types. NBHT FTDDRDDMFHDDMEWMAEDDMFTDDRDDMFHDDMEWMAEDDM RT a 0.570.620.630.630.630.600.700.740.750.76 Sine a 0.610.520.510.510.510.780.500.560.560.56 Mixed a 0.550.520.520.520.520.690.550.560.570.54 Aver a 0.580.550.550.550.550.690.580.620.630.62 RT g 0.630.630.630.630.630.760.770.770.780.78 Sine g 0.500.500.510.510.510.490.500.490.520.52 Mixed g 0.510.510.510.510.510.500.470.490.490.49 Aver g 0.550.550.550.550.550.580.580.580.600.60 ELE0.700.750.750.750.750.710.750.750.750.75 CIC0.380.800.800.800.800.710.780.780.780.77 Aver r 0.540.780.780.780.780.710.770.770.770.76 Table 8.1: Comparisons of SPC-based Methods Combined With Two Base Learners 9 Comparison of Window-based Detectors For window-based detectors (cf. Table 9.1), on abrupt-drift datasets, NB shows uniform and modest performance, with all six methods yielding an identical average AUC, offering no differentiation among detectors. With HT, performance is consistently higher and slight differences emerge: KSWIN, WSTD and D3 each reach the same top average value, while FPDD, MDDM and ADWIN obtain a marginally lower average. On gradual-drift datasets, NB again remains stable and uniform across all methods. HT yields moderately higher values, where WSTD, MDDM, ADWIN and D3 all reach the same top average, while KSWIN and FPDD each fall slightly behind. On real-world data streams, NB reveals a notable divergence among detectors: KSWIN records the lowest average, driven by a sharp drop on the CIC dataset, whereas the remaining methods achieve a substantially higher average. For HT, the results follow a similar pattern: KSWIN yields the lowest average, while the other methods achieve comparable or slightly higher performance. 20 When considering all dataset types together, HT surpasses NB across all window- based methods, though the margin remains modest. Overall, window-based methods yield similar aggregate performance levels to SPC detectors, and no single window- based detector emerges as clearly dominant within either base learner when all dataset types are combined. The results indicate that while KSWIN, WSTD and D3 share an advantage on abrupt-drift data, WSTD and D3 deliver the most reliable performance across all dataset types. NBHT KSWINFPDDWSTDMDDMADWIND3KSWINFPDDWSTDMDDMADWIND3 RT a 0.630.630.640.630.630.630.770.770.780.780.780.78 Sine a 0.510.500.500.500.500.500.540.530.540.510.520.53 Mixed a 0.520.520.520.510.510.510.520.500.500.510.500.51 Aver a 0.550.550.550.550.550.550.610.600.610.600.600.61 RT g 0.640.630.630.630.630.630.780.790.790.790.800.80 Sine g 0.510.510.510.510.510.510.510.500.510.500.520.50 Mixed g 0.510.510.510.510.510.510.490.490.500.500.490.49 Aver g 0.550.550.550.550.550.550.590.590.600.600.600.60 ELE0.700.750.750.750.750.750.710.750.750.750.750.75 CIC0.380.800.800.800.800.800.710.780.780.780.770.77 Aver r 0.540.780.780.780.780.780.710.770.770.770.760.76 Table 9.1: Comparison of Window-based Methods Combined With Two Base Learners 10 Comparison of Ensemble-based Detectors The ensemble detectors (cf. Table 10.1) demonstrate considerably higher performance across most dataset categories compared to SPC and window-based methods. On abrupt- drift datasets, NB combined with ARF achieves the strongest average, followed by AWE, AUE, and DWM in descending order. With HT, ARF further improves and continues to lead, with AWE, AUE, and DWM preserving the same ranking across both learners. On gradual-drift datasets, NB with ARF again obtains the highest average, followed by AWE, AUE, and DWM. With HT, ARF continues to lead, followed by AWE, AUE, and DWM, maintaining a consistent ordering across both base learners. On real-world data streams, the ranking shifts considerably. With NB, AUE achieves the highest average, followed by AWE, ARF, and DWM. With HT, AUE becomes the clear top performer, followed by AWE, DWM, and ARF, indicating that AUE holds a particular advantage on real-world streams while ARFâs dominance does not transfer from synthetic data. When considering all dataset types together, HT surpasses NB across all ensemble methods. ARF+HT achieves the highest overall average, closely followed by AUE+HT 21 and AWE+HT, with DWM+HT reaching the lowest among HT-based methods. For NB, AUE leads, followed by AWE, ARF, and DWM. The results indicate that while ARF consistently dominates on both abrupt and gradual synthetic drift scenarios, AUE proves most effective on real-world data, underscoring the importance of dataset characteristics when selecting an ensemble-based drift detection method. NBHT ARFAUEDWMAWEARFAUEDWMAWE RT a 0.830.730.650.760.880.740.690.77 Sine a 0.950.780.590.900.980.830.800.91 Mixed a 0.970.680.560.870.970.780.680.89 Aver a 0.920.730.600.840.940.780.720.86 RT g 0.760.730.650.780.880.740.760.80 Sine g 0.930.810.530.880.970.840.530.90 Mixed g 0.960.700.530.840.960.790.540.87 Aver g 0.880.750.570.830.940.720.610.86 ELE0.810.880.700.740.820.890.710.75 CIC0.360.700.370.640.630.870.770.75 Aver r 0.590.790.540.690.730.880.740.75 Table 10.1: Comparisons of Ensemble-based Methods Combined With Two Base Learn- ers 11 A Summary As presented in Table 11, HT generally outperforms NB across most dataset categories and detector families. However, a notable exception arises on real-world data streams, where NB achieves equal or slightly superior performance to HT within both SPC and window-based methods. This suggests that the advantage of a more expressive base learner is not universal and depends on the nature of the data. Ensemble methods consistently outperform both SPC and window-based detectors across all dataset types, and their strongest results are achieved with HT, reinforcing that the choice of base learner significantly impacts predictive accuracy under concept drift. Focusing on single-detector families, EWMA+HT and EDDM+HT offer the best over- all SPC performance (â 0.69) and the best gradual-drift accuracy (â 0.60). For real- world streams, NB paired with RDDM, FHDDM, EWMA, or EDDM achieves the high- est SPC result of approximately 0.78. Within the window-based family, KSWIN+HT, WSTD+HT, and D3+HT are jointly strongest for abrupt drifts (â 0.61), while WSTD+HT, MDDM+HT, ADWIN+HT and D3+HT perform best on gradual drifts (â 0.60). On real-world streams, all window-based methods paired with NB uniformly achieve the top result of approximately 0.78. Overall, all window-based methods with HT yield an identical aggregate performance of 0.68, with no single detector emerging as dom- 22 Cat.Drift type Learner Best method SPCAbruptHTFTDD SPCGradualHTEWMA/EDDM SPCReal-worldNBRDDM/FHDDM/EWMA/EDDM OverallHTEWMA/EDDM WindowAbruptHTKSWIN/WSTD/D3 WindowGradualHTWSTD/MDDM/ADWIN/D3 WindowReal-worldNBFPDD/WSTD/MDDM/ADWIN/D3 OverallHTWSTD/D3 Ensemble AbruptHTARF Ensemble GradualHTARF Ensemble Real-worldHTAUE OverallHTARF Table 11.1: Best-performing drift detectors by category, drift type and base learner inant. Among ensemble-based methods, ARF+HT achieves the strongest results on both abrupt and gradual drift scenarios (â 0.94) and ranks best in overall aggregate performance (â 0.83), while AUE+HT proves most effective on real-world data streams (â 0.88). Overall, although SPC and window-based detectors remain competitive, par- ticularly on real-world data with NB, the most reliable and consistent improvements across all drift scenarios come from ensemble-based adaptation mechanisms, especially ARF+HT. 12 Conclusions Concept drifts are prevalent in todayâs streaming applications where data distributions and relationships change over time. They have significant consequences in many critical domains, such as healthcare, banking, finance, cybersecurity, email spam and phishing, IoT, sensor networks and recommendation systems. In these applications, drifts can degrade the performance of decision models and increase the risk of incorrect classifi- cations. Therefore, accurately identifying such drifts is crucial to prevent classification models from deteriorating. The aim of our study is to simplify the description of a broad range of learner-based concept drift detection algorithms in order to enhance the comprehension of these com- plex methods. These algorithms are classified into three main groups, Statistical Process Control (SPC) methods, windowing techniques and ensemble-based approaches. These methods can operate in either active or passive modes and can be implemented in online or block-based settings. We have also conducted an extensive evaluation and comparison 23 of these detectors using both synthetic and real-world datasets to assess their perfor- mance. References [1] Osama A.Mehdi et al. âRoadmap of Concept Drift Adaptation in Data Stream Mining, Years Laterâ. In: IEEE Access P (Jan. 2024), p. 1â1. doi: 10.1109/ ACCESS.2024.3358817. [2] Joao Paulo Papa Adriana Sayuri Iwashita. âAn Overview on Concept Drift Learn- ingâ. In: IEEE Access (2019). [3] Supriya Agrahari and Anil Kumar Singh. âConcept Drift Detection in Data Stream Mining : A literature reviewâ. In: Journal of King Saud University - Computer and Information Sciences 34.10, Part B (2022), p. 9523â9540. issn: 1319-1578. doi: https://doi.org/10.1016/j.jksuci.2021.11.006. url: https://w. sciencedirect.com/science/article/pii/S1319157821003062. [4] Gabriel Aguiar, Bartosz Krawczyk, and Alberto Cano. âA survey on learning from imbalanced data streams: taxonomy, challenges, empirical study, and reproducible experimental frameworkâ. In: Machine Learning 113 (June 2023), p. 1â79. doi: 10.1007/s10994-023-06353-6. [5] M. Baena-GarcÄąa et al. âEarly drift detection methodâ. In: International Workshop on Knowledge Discovery from Data Streams (2006). [6] R. S. M. Barros and S. G. T. C. Santos. âA large-scale comparison of concept drift detectorsâ. In: Information Sciences 451-452 (2018), p. 348â370. issn: 0020- 0255. doi: https://doi.org/10.1016/j.ins.2018.04.014. url: https: //w.sciencedirect.com/science/article/pii/S0020025518302743. [7] Roberto Barros, Juan Gonzalez Hidalgo, and Danilo Cabral. âWilcoxon Rank Sum Test Drift Detectorâ. In: Neurocomputing 275 (Jan. 2018), p. 1954â1963. doi: 10.1016/j.neucom.2017.10.051. [8] Roberto Barros and Silas Santos. âAn Overview and Comprehensive Comparison of Ensembles for Concept Driftâ. In: Information Fusion 52 (Dec. 2019), p. 213â 244. doi: 10.1016/j.inffus.2019.03.006. [9] Roberto Barros et al. âRDDM: Reactive drift detection methodâ. In: Expert Sys- tems with Applications 90 (Dec. 2017), p. 344â355. doi: 10.1016/j.eswa.2017. 08.023. [10] BarÄą ̧s Bayram, Bilge K Ěoro Ěglu, and Mehmet G Ěonen. âImproving Fraud Detection and Concept Drift Adaptation in Credit Card Transactions Using Incremental Gradient Boosting Treesâ. In: 2020 19th IEEE International Conference on Ma- chine Learning and Applications (ICMLA). 2020, p. 545â550. doi: 10.1109/ ICMLA51294.2020.00091. 24 [11] Alessio Bernardo. Artificial and Real dataset with different concept drift type and imbalance ratio. Version V1. 2020. doi: 10.7910/DVN/RKY6RD. url: https:// doi.org/10.7910/DVN/RKY6RD. [12] A. Bifet and R. Gavalda. âLearning from Time-Changing Data with Adaptive Windowingâ. In: SDM (2007). [13] D. Brzezinski and J. Stefanowski. âAccuracy Updated Ensemble for Data Streams with Concept Driftâ. In: vol. 6679. May 2011, p. 155â163. isbn: 978-3-642-21221- 5. doi: 10.1007/978-3-642-21222-2_19. [14] D. R. L. Cabral and R. S. M. Barros. âConcept drift detection based on Fisherâs Exact testâ. In: Information Sciences 442-443 (2018), p. 220â234. issn: 0020- 0255. doi: https://doi.org/10.1016/j.ins.2018.02.054. url: https: //w.sciencedirect.com/science/article/pii/S0020025518301403. [15] Andreas Kassler Firas Bayram Bestoun S. Ahmed. âFrom concept drift to model degradation: An overview on performance-aware drift detectorsâ. In: Knowledge- Based Systems (2022). [16] R. A. Fisher. âOn the Interpretation of X2 from Contingency Tables, and the Calculation of Pâ. In: Journal of the Royal Statistical Society 85.1 (1922), p. 87â 94. issn: 09528385. url: http://w.jstor.org/stable/2340521 (visited on 11/12/2024). [17] J. Gama et al. âA Survey on Concept Drift Adaptationâ. In: ACM Computing Surveys (CSUR) (2014). [18] J. Gama et al. âLearning with Drift Detectionâ. In: Intelligent Data Analysis (2004). [19] Heitor M. Gomes et al. âCorrection to: Adaptive random forests for evolving data stream classificationâ. In: Mach. Learn. 108.10 (Oct. 2019), p. 1877â1878. issn: 0885-6125. doi: 10.1007/s10994-019-05793-3. url: https://doi.org/10. 1007/s10994-019-05793-3. [20] Ě Omer G Ěoz Ěua ̧cik et al. âUnsupervised Concept Drift Detection with a Discrimina- tive Classifierâ. In: Proceedings of the 28th ACM International Conference on In- formation and Knowledge Management (2019). url: https://api.semanticscholar. org/CorpusID:207757814. [21] Ě Omer G Ěoz Ěua ̧cÄąk and Fazli Can. âConcept learning using one-class classifiers for implicit drift detection in evolving data streamsâ. In: Artif. Intell. Rev. 54.5 (June 2021), p. 3725â3747. issn: 0269-2821. doi: 10.1007/s10462-020-09939-x. url: https://doi.org/10.1007/s10462-020-09939-x. [22] Ege Gulcan and Fazli Can. âUnsupervised concept drift detection for multi-label data streamsâ. In: Artificial Intelligence Review 56 (July 2022). doi: 10.1007/ s10462-022-10232-2. 25 [23] Meng Han et al. âA survey of active and passive concept drift handling methodsâ. In: Computational Intelligence 38.4 (2022), p. 1492â1535. doi: https://doi. org/10.1111/coin.12520. eprint: https://onlinelibrary.wiley.com/doi/ pdf/10.1111/coin.12520. url: https://onlinelibrary.wiley.com/doi/abs/ 10.1111/coin.12520. [24] Tegjyot S. Sethi Hanqing Hu Mehmed Kantardzic. âNo Free Lunch Theorem for concept drift detection in streaming data classification: A reviewâ. In: Wiley In- terdisciplinary Reviews: Data Mining and Knowledge Discovery (2019). [25] Fabian Hinder, Valerie Vaquet, and Barbara Hammer. One or Two Things We know about Concept Drift â A Survey on Monitoring Evolving Environments. 2023. arXiv: 2310.15826 [cs.LG]. url: https://arxiv.org/abs/2310.15826. [26] Wassily Hoeffding. âProbability Inequalities for Sums of Bounded Random Vari- ablesâ. In: Journal of the American Statistical Association 58.301 (1963), p. 13â 30. issn: 01621459, 1537274X. url: http://w.jstor.org/stable/2282952 (visited on 10/22/2024). [27] M. Hammami I. Khamassi M. Sayed-Mouchaweh and K. Gh Ěedira. âDiscussion and review on evolving data streams and concept drift adaptingâ. In: Evolving Systems (2018). [28] Meenal Jain, Gagandeep Kaur, and Vikas Saxena. âA K-Means clustering and SVM based hybrid concept drift detection technique for network anomaly detec- tionâ. In: Expert Systems with Applications 193 (2022), p. 116510. issn: 0957- 4174. doi: https://doi.org/10.1016/j.eswa.2022.116510. url: https: //w.sciencedirect.com/science/article/pii/S0957417422000112. [29] Botao Jiao et al. âDynamic Ensemble Selection for Imbalanced Data Streams With Concept Driftâ. In: IEEE Transactions on Neural Networks and Learning Systems 35.1 (2024), p. 1278â1291. doi: 10.1109/TNNLS.2022.3183120. [30] J. Zico Kolter and Marcus A. Maloof. âDynamic Weighted Majority: An Ensemble Method for Drifting Conceptsâ. In: J. Mach. Learn. Res. 8 (Dec. 2007), p. 2755â 2790. issn: 1532-4435. [31] Bartosz Krawczyk et al. âEnsemble learning for data stream analysis: A surveyâ. In: Information Fusion 37 (2017), p. 132â156. issn: 1566-2535. doi: https:// doi.org/10.1016/j.inffus.2017.02.004. url: https://w.sciencedirect. com/science/article/pii/S1566253516302329. [32] N. Littlestone and M.K. Warmuth. âThe Weighted Majority Algorithmâ. In: Infor- mation and Computation 108.2 (1994), p. 212â261. issn: 0890-5401. doi: https: //doi.org/10.1006/inco.1994.1009. url: https://w.sciencedirect.com/ science/article/pii/S0890540184710091. [33] Jes Ěus L Ěopez Lobo. Synthetic datasets for concept drift detection purposes. url: https://doi.org/10.7910/DVN/5OWRGB. Harvard Dataverse. 26 [34] Jie Lu et al. âData-driven decision support under concept drift in streamed big dataâ. In: Complex and Intelligent Systems 6 (2020). doi: 10.1007/s40747-019- 00124-4. [35] Jie Lu et al. âLearning under Concept Drift: A Reviewâ. In: IEEE Transactions on Knowledge and Data Engineering 31.12 (2019), p. 2346â2363. doi: 10.1109/ TKDE.2018.2876857. [36] Donia Malekian and Mahmoud Reza Hashemi. âAn adaptive profile based fraud detection framework for handling concept driftâ. In: 2013 10th International ISC Conference on Information Security and Cryptology (ISCISC). 2013, p. 1â6. doi: 10.1109/ISCISC.2013.6767338. [37] Colin McDiarmid. âSurveys in Combinatorics, 1989: On the method of bounded differencesâ. In: 1989. url: https://api.semanticscholar.org/CorpusID: 116663483. [38] Hassan Mehmood et al. âConcept Drift Adaptation Techniques in Distributed Environment for Real-World Data Streamsâ. In: Smart Cities (2021). url: https: //api.semanticscholar.org/CorpusID:233623878. [39] K. Nishida and K. Yamauchi. âDetecting Concept Drift Using Statistical Testingâ. In: Discovery Science, 10th International Conference, DS 2007 (2007). [40] Yuri Thomas P. Nunes and Luiz Affonso Guedes. âConcept Drift Detection Based on Typicality and Eccentricityâ. In: IEEE Access 12 (2024), p. 13795â13808. doi: 10.1109/ACCESS.2024.3355959. [41] Sirvan Parasteh and Samira Sadaoui. âA Novel Probabilistic Approach for Detect- ing Concept Drift in Streaming Dataâ. In: Deep Learning Theory and Applications. Ed. by Donatello Conte et al. Cham: Springer Nature Switzerland, 2023, p. 173â 188. isbn: 978-3-031-39059-3. [42] Sirvan Parasteh and Samira Sadaoui. âA Probabilistic Approach for Detecting Real Concept Drift.â In: Proc. of 16th International Conference on Agents and Artificial Intelligence-Volume 2, ICAART (2). 2024, p. 301â311. [43] Sirvan Parasteh and Samira Sadaoui. âA Robust Probabilistic Framework for Iden- tifying and Evaluating Concept Drift in Abrupt and Gradual Scenariosâ. In: Agents and Artificial Intelligence. Ed. by Ana Paula Rocha, Luc Steels, and Jaap van den Herik. Springer Nature, 2025, p. 353â367. [44] Sirvan Parasteh, Samira Sadaoui, and Mohammad Sadegh Khosravani. âDetection of Real Concept Drift Under Noisy Data Streamâ. In: 2023 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE. 2023, p. 1316â1321. [45] A. Pesaranghader, H. Viktor, and E. Paquet. McDiarmid Drift Detection Methods for Evolving Data Streams. 2018. arXiv: 1710.02030 [stat.ML]. 27 [46] Ali Pesaranghader and Herna L. Viktor. âFast Hoeffding Drift Detection Method for Evolving Data Streamsâ. In: Machine Learning and Knowledge Discovery in Databases. Cham: Springer International Publishing, 2016, p. 96â111. isbn: 978- 3-319-46227-1. [47] C. Raab, M. Heusinger, and F. Schleif. âReactive Soft Prototype Computing for Concept Drift Streamsâ. In: Neurocomputing (2020). [48] Gordon J. Ross et al. âExponentially weighted moving average charts for detecting concept driftâ. In: Pattern Recognition Letters 33.2 (2012), p. 191â198. issn: 0167- 8655. doi: https://doi.org/10.1016/j.patrec.2011.08.019. [49] Armin Sadreddin and Samira Sadaoui. âChunk-based incremental feature learn- ing for credit-card fraud data streamâ. In: Journal of Experimental & Theoretical Artificial Intelligence (2022), p. 1â19. doi: 10.1080/0952813X.2022.2153277. url: https://doi.org/10.1080/0952813X.2022.2153277. [50] Bruno Henrique Schwengber et al. âLearning From Network Data Changes for Unsupervised Botnet Detectionâ. In: IEEE Transactions on Network and Service Management 19.1 (2022), p. 601â613. doi: 10.1109/TNSM.2021.3109076. [51] Eyad Elyan Scott Wares John Isaacs. âData stream mining: methods and chal- lenges for handling concept driftâ. In: SN Applied Sciences (2019). [52] Iman Sharafaldin, Arash Habibi Lashkari, and Ali A. Ghorbani. âToward Gener- ating a New Intrusion Detection Dataset and Intrusion Traffic Characterizationâ. In: International Conference on Information Systems Security and Privacy. 2018. url: https://api.semanticscholar.org/CorpusID:4707749. [53] Andr Ěes L. Su Ěarez-Cetrulo, David Quintana, and Alejandro Cervantes. âA survey on machine learning for recurring concept drifting data streamsâ. In: Expert Systems with Applications 213 (2023), p. 118934. issn: 0957-4174. doi: https://doi. org/10.1016/j.eswa.2022.118934. url: https://w.sciencedirect.com/ science/article/pii/S0957417422019522. [54] N.V. Chawla T.R. Hoens R. Polikar. âLearning from streaming data with concept drift and imbalance: an overviewâ. In: Progress in Artificial Intelligence (2012). [55] Maurras Ulbricht Togbe et al. âAnomalies Detection Using Isolation in Concept- Drifting Data Streamsâ. In: Computers 10.1 (2021). issn: 2073-431X. doi: 10. 3390/computers10010013. url: https://w.mdpi.com/2073-431X/10/1/13. [56] H. Wang et al. âMining Concept-Drifting Data Streams Using Ensemble Classi- fiersâ. In: (July 2003). doi: 10.1145/956750.956778. [57] Geoffrey I. Webb et al. âUnderstanding Concept Driftâ. In: ArXiv (2017). [58] Frank Wilcoxon. âIndividual Comparisons by Ranking Methodsâ. In: Biometrics Bulletin 1.6 (1945), p. 80â83. issn: 00994987. url: http://w.jstor.org/ stable/3001968 (visited on 10/28/2024). 28 [59] Q. Xiang et al. âConcept Drift Adaptation Methods under the Deep Learning Framework: A Literature Reviewâ. In: Applied Sciences 13 (May 2023), p. 6515. doi: 10.3390/app13116515. [60] Lingyu Zhang, Jiabao Zhao, and Wei Li. âOnline and Unsupervised Anomaly De- tection for Streaming Data Using an Array of Sliding Windows and PDDsâ. In: IEEE Transactions on Cybernetics 51.4 (2021), p. 2284â2289. doi: 10.1109/ TCYB.2019.2935066. [61] Indre Ë Zliobaite. âLearning under Concept Drift: an Overviewâ. In: ArXiv (2010). 29