Their creation, called Dynamic Continuous Progressive Neural Networks, or DYNcPNN, addresses a scenario the authors term Evolving Streaming Time Series. In such settings, data arrives as an unbounded stream, past values influence future ones, the underlying statistical relationships shift without warning, and the model must somehow absorb new knowledge without erasing what it has already learned. Each of these demands has traditionally been the province of a separate research field. Streaming Machine Learning emphasizes rapid adaptation to drifts but happily abandons old knowledge in the process. Continual Learning protects deep networks from catastrophic forgetting but assumes drifts merely introduce new data distributions rather than contradictory ones. Time Series Analysis, particularly through recurrent architectures like Long Short-Term Memory networks, handles temporal dependence but is rarely deployed in true streaming contexts. The Milanese team’s Streaming Continual Learning paradigm, introduced in previous work and now embodied in DYNcPNN, unifies all three.

The problem DYNcPNN confronts is subtler than it first appears. Concept drift comes in flavors that demand different responses. The authors draw a novel distinction between what they call contradictory drifts and input drifts. A contradictory drift changes the decision boundary itself: the same input that once meant “normal conditions” may now signify a storm, perhaps because an agency has revised its alert thresholds. An input drift, by contrast, introduces unfamiliar data—new slang in customer reviews, novel sensor readings—without invalidating the old classification rules. The difference matters enormously. Contradictory drifts force a model to relearn, while input drifts reward a model that can weave new knowledge into existing structures without disturbing them. Real streams present both, often at once, along with temporal dependence, where the probability of an event at time t genuinely depends on what happened at t minus some lag. Standard streaming classifiers, which typically assume data points are independent, ignore this temporal structure entirely and pay for it in accuracy.

DYNcPNN builds on the team’s earlier architecture, cPNN, which itself adapted the Progressive Neural Networks strategy from Continual Learning to the streaming domain. The original Progressive Neural Networks grow by adding an entirely new “column” of neural network layers whenever a new task appears, while freezing the weights of previous columns so their knowledge remains intact. Transfer connections let new columns draw on old columns’ representations, enabling selective reuse of past expertise. The cPNN variant applied this logic on top of a Continuous LSTM—a recurrent base learner that buffers the incoming stream in fixed-size mini-batches and builds sequences with a hopping window—thereby capturing temporal dependence while learning continuously. But cPNN suffered from two crippling limitations: it required knowing the exact timestamps of concept drifts in advance, and it expanded its architecture every single time a drift was detected, whether or not such expansion was warranted.

The new work eliminates both limitations through two intertwined mechanisms. The first is an automatic concept drift detector based on ADWIN—ADaptive WINdowing—a well-established streaming technique that maintains a sliding window of recent observations, here a binary correctness signal of the model’s predictions. ADWIN continuously evaluates all ways of splitting the window into a recent half and an older half; when the difference between the means of the two sub-windows exceeds a statistically motivated threshold, a drift is signaled and the oldest observations are discarded. By wiring this detector into the architecture, DYNcPNN no longer needs to be told when the world has changed—it notices on its own, without manual intervention.

The second and arguably more consequential innovation is the dynamic decision mechanism that governs when the network actually expands. When a drift is detected, expanding the architecture is not always the right response. Adding a column costs memory, and if the new concept is mild or closely related to the current one, simple continued training on the existing network may suffice. DYNcPNN therefore evaluates whether the drift is severe enough to justify growth. If the drift is judged contradictory and substantial, a new column is added and the old ones are frozen, preserving their knowledge forever. If, however, the model decides to adapt without expanding—continuing to train the existing network—it risks overwriting previously learned knowledge. To prevent exactly this form of catastrophic forgetting, the authors introduce a protective strategy that safeguards the relevant prior knowledge during this no-expansion adaptation phase. The net effect is a model that spends memory only when necessary and consistently outperforms its predecessors in both accuracy and footprint.

The severity of a drift, which drives this decision, is formalized in the paper as the discrepancy between the joint probability distributions before and after the change. Mild drifts require only minor adjustments; severe ones demand substantial relearning. For contradictory drifts, severity can be gauged through the “influence zone”—the proportion of the previously observed input space where the new concept changed the labels. This quantitative framing allows the architecture to make principled rather than arbitrary choices about when to grow.

The experimental campaign is notably thorough. The team benchmarked DYNcPNN against its own predecessor cPNN, against the underlying cLSTM trained continuously, and against the heavyweights of the Streaming Machine Learning world, including Hoeffding Adaptive Trees and Adaptive Random Forests. The benchmarks were crafted to include significant temporal dependence: the authors injected elaborate temporal dependencies into the labels of synthetic streams, and they also tested on real-world data covering weather prediction, air pollution, and power consumption—domains where sensor readings evolve continuously and past measurements genuinely shape the future. The results, reported in the paper, show DYNcPNN consistently outperforming all competitors, adapting more quickly to concept drifts, mitigating catastrophic forgetting effectively, and optimizing memory usage. Equally revealing were the failures of the baselines: the streaming decision-tree and forest models, unable to account for temporal dependence, lagged behind on all temporally structured tasks.

The significance of this work extends beyond a single architecture. It serves as a pointed critique of the assumptions underpinning much of classical machine learning. Offline learning rests on the i.i.d. assumption—that data are independent and identically distributed—which the Empirical Risk Minimization principle requires for meaningful statistical guarantees. Streaming data with temporal dependence and concept drift violates both conditions at once. Correlated samples reduce the effective number of independent observations, biasing risk estimates, while shifting distributions make convergence impossible in principle. The authors’ response is not to patch the old framework but to build one designed for the violation, drawing on an emerging body of SCL research that has recently begun to emphasize taming temporal dependence in streams.

Practical applications are easy to imagine. The paper’s own motivating example is environmental nowcasting: a classifier receiving an unbounded stream of temperature, humidity, and radiation measurements from sensors must predict the current weather condition in real time, before official meteorological confirmation arrives. Seasonal transitions produce input drifts; revised alert thresholds produce contradictory drifts; rain persistence and gradual temperature trends produce temporal dependence; and recurring phenomena—summer storms, winter snow showers—demand that old knowledge remain retrievable rather than destroyed. The same structure applies to network security, financial monitoring, industrial sensor networks, and any Internet of Things deployment where the ground truth arrives late or never.

The authors are candid about limits and future directions, and DYNcPNN is presented as a pioneering embodiment of a young paradigm rather than a finished solution. Yet the trajectory is clear: as data-hungry applications proliferate and the gap between laboratory conditions and operational reality widens, architectures that learn continuously, remember selectively, and adapt autonomously will become less exotic and more essential. With DYNcPNN, the Streaming Continual Learning paradigm has gained a concrete, benchmarked, and open blueprint for what that future might look like—neural networks that grow only when they must, forget only what they should, and never stop reading the stream.

The work, published in Data Mining and Knowledge Discovery, addresses a problem that has become increasingly urgent as machine-learning systems move from laboratories into high-stakes environments. A model can lose predictive power for several different reasons, and those causes are not always easy to distinguish. Concept drift occurs when the relationship between inputs and outcomes changes. In a medical system, for example, the same symptoms may become associated with a different diagnosis because a disease evolves or treatment practices change. Virtual concept drift occurs when the distribution of incoming inputs changes while the underlying relationship between inputs and labels remains stable. A financial model may encounter a new population of customers without the rules governing credit risk changing. Label drift, as defined in the study, involves systematic changes in observed labels, including label flipping or the appearance of previously unseen labels. Concept evolution goes further: it describes the emergence of genuinely new classes that were absent from the original training data.

Crypsis is designed to treat these phenomena not as one undifferentiated problem, but as related yet distinguishable events. Its central component, the Observer, stores statistical metadata for each known class. For every class, the system records lower and upper feature bounds, ranges of population variance, and an index tracking known concepts and their evolution. In the authors’ formulation, if there are (m) classes and (n) features, the Observer is represented as an (m \times (2n+3)) matrix. The feature boundaries help determine whether a new record resembles a known class, while population-variance ranges provide additional information about the normal spread of that class. A record falling outside expected ranges may be an outlier, a drifted example, or evidence of a new class. The Observer therefore acts as a compact statistical memory, allowing the system to compare each incoming instance with the structural profile of previously learned concepts.

When a record arrives during supervised testing, Crypsis evaluates its features through a set of confidence scores. These scores are combined into a confidence matrix that indicates how closely the record matches the system’s stored expectations. High confidence suggests that the record belongs to a familiar, stable concept and can be handled by a base model. Low confidence triggers a more careful investigation. The system must then determine whether the unusual record is merely noise, an outlier, a shifted example from an existing class, or part of a real drift event. This distinction matters because blindly retraining on every unusual observation could make a model less stable rather than more adaptive. Crypsis uses thresholds and statistical comparisons to reduce this risk, while its design aims to avoid excessive reliance on ensembles that may become overconfident when several models make the same incorrect assumption.

The Grid provides the framework’s response to recognized drift. It is organized around mappings between original and observed labels. In a simple four-class example, the diagonal of the Grid contains base models for situations in which the original and observed labels agree. Off-diagonal positions represent specialized models trained to manage particular forms of drift—for instance, cases in which examples originating from one class are observed under another label. Once the system identifies a drift pattern, it can direct new records to the corresponding model rather than forcing every situation through a single classifier. The Heap supports this process by temporarily collecting records suspected of belonging to the same drift or emerging class. When enough related examples accumulate to cross a predefined threshold, the cluster is retrieved and used to train or update a specialized model. The Heap is then cleared of those records, limiting the memory burden of continual adaptation.

The most distinctive part of Crypsis appears before the Observer and Grid are fully trusted. In real data streams, the initial training material may already contain hidden drift, inconsistent labels, or contaminated regions. Training directly on such data could embed the problem into the system’s statistical memory. To address this uncertainty, the authors introduce an “Elitism” approach. The dataset is divided into multiple sections, and machine-learning models are trained and tested on different divisions. The resulting accuracy scores are aggregated while division sizes and configurations are varied. Data segments that meet or exceed a chosen performance threshold are retained as elite divisions, while weaker segments are discarded. In a second stage, elite divisions are concatenated in different combinations. Their joint performance is measured through a quantity called “Reputation,” which reflects how often a division participates in combinations that achieve high accuracy. Divisions with consistently strong reputations are selected for subsequent base-model training, with the goal of creating an Observer from comparatively reliable data.

This selection process is intended to reduce the danger of model overconfidence, but it also introduces a transition problem. If only elite divisions are used at first, the Observer may not cover the full range of legitimate feature values. During an initial transitional testing phase, the system can consequently produce false positives by treating normal records as suspicious. Crypsis addresses this by using early testing to gather information and then performing limited model refinement before actual deployment. In effect, the system first learns which apparently unusual records are normal for the broader stream, then updates its base models and Observer. The approach combines incremental learning—updating a model as new examples arrive—with aspects of lifelong learning, in which knowledge from earlier concepts is retained and used to support future adaptation. The distinction is important: online learning generally focuses on the same task over time, whereas lifelong learning aims to preserve and reuse knowledge across changing concepts.

The researchers evaluated Crypsis using synthetic datasets and the Electricity, or ELEC2, benchmark. The synthetic experiments were designed to create controlled label-drift and concept-evolution scenarios, including multiple classes, changing labels, overlapping feature spaces, and newly emerging categories. Four widely used classifiers—K-nearest neighbors, support vector machines, random forests, and XGBoost—were tested under ordinary conditions, under injected drift or evolution, and with Crypsis-enhanced processing. The authors report consistent improvements across accuracy, precision, recall, and F1 score. They also conducted paired t-tests comparing the enhanced and baseline models, reporting statistically significant gains in both label-drift and concept-evolution experiments. However, the paper notes that the synthetic datasets were balanced and that misclassifications were symmetrically distributed, making the four reported metrics numerically identical in the presented results. This detail is important because equal scores do not automatically imply that a system performs equally well on every class in an imbalanced real-world setting.

The real-world evaluation used ELEC2, which contains 45,312 observations and eight input attributes collected from the Australian New South Wales electricity market between 1996 and 1998. The task is to predict whether the next electricity price will rise or fall relative to a moving average calculated over the previous 24 hours. Because electricity demand, supply, weather, seasonality, and market behavior continually change, the dataset is a classic test bed for concept drift. Crypsis was evaluated under several configurations of its Elitism procedure and compared with published approaches, including incremental one-class ensembles, dynamically weighted ensembles, weighted incremental–decremental support-vector machines, Bhattacharyya-distance drift detection, ElStream, and parameter-estimation procedures for adaptive ensembles. The authors report that Crypsis achieved higher accuracy than the listed methods and maintained strong performance across its own parameter variations. Yet the comparison should be interpreted carefully: the study states that results for earlier methods were taken from their original publications rather than reproduced under identical hardware, code, or experimental settings. The authors identify a standardized computational benchmark as an important direction for future work.

Crypsis ultimately presents model-drift management as a layered process of observation, diagnosis, memory, selection, and adaptation rather than as a single alarm mechanism. Its claimed contribution is the integration of concept-drift detection, label-drift handling, novel-class discovery, statistical class profiling, and selective pre-training into one observer-based architecture that extends beyond binary classification. The framework is also intended to provide interpretability by indicating whether a change resembles an outlier, a known drift pathway, or an emerging class. The researchers envision applications in areas such as finance, healthcare, cybersecurity, autonomous systems, and energy forecasting, where silent performance degradation can carry serious consequences. Future work will focus on multimodal streams combining text, images, and sensor data, as well as on measuring runtime, memory consumption, scalability, and deployment cost. If those challenges can be addressed, Crypsis could help transform drift detection from a reactive maintenance task into a continuous form of machine intelligence—one that notices when the world has changed before an outdated model becomes a hidden liability.

Subject of Research: Detection and mitigation of concept drift, label drift, concept evolution, and model drift in streaming machine-learning systems.

Article Title: “Crypsis: an elitism-driven observer-based approach for detection and mitigation of concept drift, concept evolution, and label drift”

Article References: Yanni, G. S., Rashad, H. S., & Maghraby, F. A. (2026). “Crypsis: an elitism-driven observer-based approach for detection and mitigation of concept drift, concept evolution, and label drift.” Data Mining and Knowledge Discovery, 40, Article 33. Related references include Gama et al. (2014), Lu et al. (2018), Masud et al. (2011), and Nguyen et al. (2016).

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01191-1

Keywords: Concept drift, label drift, model drift, concept evolution, novel class detection, incremental learning, lifelong learning, data streams, drift detection, machine learning adaptation.