[Research Briefing] Addressing Concept Drift via XAI-based Profile Drift Detection (PDD)

Introduction

Modern enterprise AI services operate on continuously evolving real-time data streams rather than static datasets. However, many companies rely solely on basic accuracy monitoring after deployment, which leads to critical delays in response during ‘concept drift’—situations where the fundamental relationships within the data change.

Concept drift occurs when the correlation between input data and the target variable changes, even if the distribution of the input data remains constant. For example, if consumer behavior patterns shift abruptly, model accuracy may decline gradually; however, without understanding the root cause, organizations fall into the inefficiency of repetitive, blind retraining.

At this juncture, the adoption of XAI (Explainable AI) must evolve beyond simple ‘explanation’ into a tool for ‘monitoring.’ By tracking the ‘profile’ of how a model interprets data, system anomalies can be captured before performance degradation becomes visible.

This research presents a strategy to secure both model reliability and maintenance efficiency by integrating XAI-based drift detection into the MLOps pipeline through a specific methodology called PDD.

Core Thesis

The paper seeks to address data drift—specifically ‘Concept Drift,’ where the relationship between explanatory variables and the response variable changes—which leads to degraded predictive model performance as data distributions evolve. It points out that traditional detection methods based on accuracy or marginal distributions are limited in their ability to capture subtle conceptual shifts.

The researchers propose Profile Drift Detection (PDD) based on Partial Dependence Profiles (PDP). PDD utilizes three complementary metrics to quantify changes in PDP: the Partial Dependence Index (PDI), L2 distance, and L2Der (derivative distance). This approach enables continuous model monitoring and adaptive retraining within an MLOps framework.

Six synthetic and real-world datasets (SEA, Hyperplane, NOAA, Ozone, Elec2, Friedman) were used for experimentation. Logistic Regression (LR), Decision Trees (DT), and Random Forests (RF) were employed as models, and PDD was analyzed in comparison to existing statistical detection methods such as DDM, EDDM, KSWIN, Page-Hinkley, and HDDM.

Consequently, PDD demonstrated greater stability by exhibiting a more conservative tendency than aggressive detection methods like KSWIN or EDDM, thereby reducing false positives. In particular, it proved balanced performance by recording competitive accuracy and appropriate drift detection counts in the Elec2 and Friedman datasets.

In-Depth Analysis of Key Failure Factors

First, there is a risk of distorted detection signals due to model overfitting. In experiments with the Hyperplane, NOAA, and Ozone datasets, the Random Forest (RF) model exhibited overfitting, characterized by very high training accuracy but low test accuracy. In these instances, PDD’s detection counts were inconsistent or lacked stability, indicating that when a model learns noise rather than general relationships, the PDP profile itself becomes unstable, reducing the reliability of drift detection.

Second, there is a risk of undetected drift due to the limitations of single-variable-centric detection. PDD fundamentally monitors a single most important variable selected based on Permutation Importance. However, in real-world environments, drift may occur through the complex interaction of multiple variables, or changes starting from low-importance variables may gradually affect the entire model. In such cases, tracking only the single most important variable may miss subtle conceptual changes.

Third, there are risks related to computational cost and scalability constraints. The generation of PDPs and the calculation of derivatives (L2Der) increase in computational cost as data volume grows. Particularly when extending to multi-class classification problems, profiles must be tracked and compared for each class, which can lead to spikes in computational load in real-time streaming environments, causing monitoring latency or degrading overall system throughput.

Source: Photo on Unsplash / Technology Team Archive

Analysis of PDD Quantitative Metrics: PDI, L2, and L2Der

PDD (Profile Drift Detection) goes beyond simple performance monitoring to detect concept drift by quantifying changes in the Partial Dependence Profile (PDP). The core of this methodology is the use of three complementary metrics—PDI, L2, and L2Der—to measure the distance between PDP curves. Each metric analyzes drift from different perspectives (shape, magnitude, and rate of change) to minimize the possibility of false negatives or false positives associated with single-metric systems.

The first metric, L2 distance, measures the overall difference and scale between two profiles. It is calculated by summing the squared differences of values at all points on the profile, thereby evaluating the structural similarity or general fluctuation of the PDP curve. Even if the overall shape of two profiles is similar but they have shifted parallel to the vertical axis, the L2 distance recognizes this as a significant difference and detects it.

The second metric, L2Der, compares the distances between the first derivatives of the profiles. Its role is to analyze the speed and severity of value changes within the profile. Unlike L2, which looks at value differences, L2Der precisely identifies changes in the slope of the curve and local fluctuations. This allows for the capture of dynamic behavioral changes in specific sections as well as the overall shape, providing a deeper understanding of the cause of the drift.

The third metric, PDI (Partial Dependence Index), focuses on behavioral similarity, evaluating whether the profile increases or decreases at the same data point. PDI quantitatively expresses the directional change between two profiles to analyze whether trends are aligned or opposite. It is specialized in identifying how the ‘direction of the relationship’ has changed rather than the absolute distance.

According to Algorithm 1 in the paper, PDD organically combines these three metrics to determine drift. First, PDPs for the most important variable are calculated for the training and test sets, and the resulting PDI, L2, and L2Der values are set as thresholds. Subsequently, when a new data batch arrives, the PDP for that batch is calculated and compared against the existing thresholds. Concept drift is determined to have occurred only when all three metrics—PDI, L2, and L2Der—exceed their respective thresholds.

This multi-metric system operates on complementary principles. For example, if the profile shape is nearly identical but the position has shifted, a false negative may occur if relying solely on PDI; in this case, the L2 metric compensates. Conversely, situations involving a reversal of relationships rather than simple value fluctuations trigger sensitive responses from PDI and L2Der. Consequently, by reviewing distance, rate of change, and directionality, PDD provides interpretable insights into which variable is driving the drift and how the relationship has evolved, unlike black-box detection models.

Detection Performance Comparison Across 6 Datasets

Experiments using six benchmark datasets (SEA, Hyperplane, NOAA, Ozone, Elec2, Friedman) revealed that PDD possesses distinct detection characteristics compared to traditional statistical methods. While methodologies like KSWIN and EDDM detect drift very aggressively, PDD tends to take a relatively conservative approach, reducing unnecessary alarms and increasing system stability.

In the SEA dataset, PDD did not detect a single drift instance for the Logistic Regression (LR) model. This was because the PDPs of the training and test sets overlapped almost perfectly, leaving the PDI value near zero. In contrast, KSWIN and EDDM reacted very sensitively, recording up to 30 drifts depending on batch size. In Decision Tree (DT) and Random Forest (RF) models, PDD’s detection count increased as batch size grew, yet it remained stable and controlled compared to KSWIN.

In the Hyperplane dataset, PDD’s sensitivity varied significantly by model type and batch size. In the LR model, PDD remained conservative, failing to detect drift despite performance degradation. The DT model showed irregular patterns, rapidly detecting 15 drifts in 20 batches before dropping to 1 drift in 30 batches, suggesting that PDD’s sensitivity can be influenced by changes in the data stream or model behavior.

Experiments on NOAA and Ozone datasets observed the impact of model complexity and overfitting on detection performance. Specifically, in the RF model, inconsistencies occurred where PDD’s detection counts spiked or plummeted. In the NOAA dataset’s RF model, detections jumped from 0 in 20 batches to 22 in 30 batches. This suggests that overfitting in the RF model may have distorted the signals during PDP comparison, meaning PDD’s reliability can fluctuate when a model learns noise.

In the Elec2 dataset, PDD demonstrated its most balanced performance. It maintained competitive accuracy across LR, DT, and RF models while managing drift detection counts at appropriate levels. For instance, while other methodologies uniformly recorded 10 drifts in the LR model, PDD responded flexibly, detecting 4–12 drifts depending on batch size. In this dataset, the ‘nswprice’ variable was consistently identified as the most important across all models, allowing for high-reliability detection by tracking its profile changes.

Finally, in the Friedman dataset (a regression task), PDD showed highly efficient detection capabilities while maintaining stable RMSE values. It detected 1–3 drifts in the LR model and 2–3 in the DT model, minimizing unnecessary retraining alarms. This stands in stark contrast to KSWIN, which generated 20–30 drifts across all models. By precisely tracking changes in the predictive mechanism centered on the key variable ‘variable_4,’ PDD successfully captured actual conceptual changes while lowering the risk of false positives.

Overall, PDD adopts a strategy of quantifying actual changes in the relationship between variables and the response variable, rather than reacting to every minute change like KSWIN or EDDM. This characteristic makes it a powerful tool in real-time production environments for detecting internal relationship changes when performance metrics remain silent, without compromising model stability. However, the fact that detection sensitivity varies based on the degree of model overfitting and batch size settings is a factor that must be considered during operation.

Correlation Between Variable Importance Shifts and Drift Detection

The core of the PDD methodology is tracking the relationship between the variables the model uses for prediction and the response variable via the Partial Dependence Profile (PDP). Specifically, rather than monitoring all variables, PDD detects drift focusing on the most important variable selected via Permutation Importance. Variable importance tables in Appendix A.9 show that the key variables the model focuses on are not fixed but change as data batches progress. This transition of key variables directly impacts PDD’s detection results, suggesting that beyond simple data distribution changes, the model’s ‘interpretive framework’ itself is evolving.

The Ozone dataset clearly illustrates the correlation between variable importance shifts and drift detection. In the RF model for this dataset, variable V56 was most important at the 10-batch stage, but the key variable shifted to V61 at the 20- and 30-batch stages. Similarly, the LR model shifted from V25 to V19, and the DT model from V31 to V36. This indicates that as the data stream flows, the priority of information the model utilizes to predict the target variable is readjusted.

These shifts in variable importance introduce volatility into PDD’s detection sensitivity. In the DT model of the Ozone dataset, no drift was detected at 10 batches, and only 1 was detected at 20 batches, but the count spiked to 10 at 30 batches. Analysis indicates that this surge is closely related to the shift of the most influential variable from V31 to V36.

In other words, PDD does not merely compare statistical figures of data; it determines drift based on the PDP curve of the ‘highest-priority variable’ selected by the model. Therefore, if the core variable being monitored changes, thresholds or profile shapes set based on the previous variable are no longer valid, which can generate strong drift signals as new variable-response relationships form. This demonstrates that PDD is highly sensitive to changes in the model’s internal decision-making structure.

Conversely, more stable detection results appear in cases where the key variable remains constant across all batches. In the Elec2 dataset, ‘nswprice’ was consistently identified as the most important variable across all batches for LR, DT, and RF models. In such an environment with a fixed key variable, PDD maintained a conservative detection trend with low volatility, efficiently capturing actual conceptual changes while reducing unnecessary false positives.

In conclusion, PDD’s detection mechanism aligns with the dynamics of model variable importance. Environments where key variables change frequently can be interpreted as signals that the model’s predictive logic itself is fluctuating, which PDD quantifies as changes in PDP shape. This enables a specific root cause analysis—such as ‘the model is now judging based on V36 instead of V31’—going beyond the result-oriented metric of ‘performance has dropped.’

However, relying on a single most important variable introduces risks. The original text notes that when the number of explanatory variables is very large or when the contributions of multiple variables increase simultaneously, tracking only one top variable may result in missing subtle drifts occurring in other variables. To resolve this, strategies such as updating important variables every batch or simultaneously monitoring multiple top-importance variables are required, which inevitably increases computational costs.

PDD-based MLOps Adaptive Retraining Workflow

The essence of PDD (Profile Drift Detection), as specified in Algorithm 1, is managing the model lifecycle through an adaptive loop of ‘Detection $ightarrow$ Data Addition $ightarrow$ Model Update $ightarrow$ Profile Refresh.’ This process aims to preemptively respond to concept drift by tracking changes in the Partial Dependence Profile (PDP)—the model’s way of interpreting data—in real-time, rather than waiting for performance degradation. From an MLOps perspective, this is implemented as a continuous pipeline integrating model monitoring and adaptation strategies.

The first execution step is the establishment of the baseline profile and threshold definition. After training the initial model using training and test data, the most influential variable is selected via Permutation Importance. The PDPs for the training and test sets are calculated for this variable, and PDI, L2, and L2Der metrics are derived and stored as the thresholds for drift determination. This serves as the foundational work to define the degree of volatility the model exhibits in a normal state.

The second step is continuous monitoring of real-time batch data. Whenever a new batch enters the data stream, the model performs predictions and immediately generates a PDP for that batch. The new profile is compared with the baseline profile using the three previously established metrics (PDI, L2, L2Der). If all three metrics exceed their set thresholds, the system determines that ‘concept drift’—a change in the relationship between the variable and the target—has occurred, and triggers an alarm.

The third step is data augmentation and adaptive model updating. Upon drift detection, the system expands the data pool by adding the new batch data to the existing training dataset. The model is then retrained based on the expanded data to reflect the evolved concepts. This process ensures the model does not stagnate but optimizes itself according to the dynamic environment, serving as a key mechanism to prevent abrupt drops in predictive performance.

The final step is the dynamic refresh of the baseline profile. Since the relationship between variables and the target is redefined when the model is updated, the PDP is recalculated based on the new model to update the baseline profile. Through this, the system recognizes the evolved concept as the new ‘normal state’ and sets a new reference point for subsequent drift detection. As this loop repeats, the model becomes an adaptive system that evolves in alignment with the direction of the data stream.

This workflow provides particularly strong advantages in real-time environments. Traditional methods must wait until performance metrics like accuracy drop, but the PDD loop can detect drift even when performance metrics are silent because it captures structural changes in the PDP. Furthermore, because the variable inducing the drift can be visually verified, operators gain interpretable insights into which data characteristics changed, moving beyond simple retraining.

However, in actual implementation, the batch size setting significantly impacts detection sensitivity. According to experimental results, a batch size that is too large may miss drift detection, while one that is too small increases the risk of false positives. Therefore, determining the optimal batch size considering the data inflow rate and model complexity is a critical factor in ensuring the stability of a PDD-based MLOps pipeline.

Checklist for Building XAI-based Concept Drift Monitoring

To prevent such system failures in a practical engineering pipeline, the following core defense mechanisms must be enforced during the operational phase:

• Key Variable Selection: Are you periodically identifying the variables that most heavily influence model predictions via Permutation Importance or similar methods?
• PDP Baseline Establishment: Have the L2, L2Der, and PDI values of the PDP derived from the training and test sets been set as baseline thresholds?
• Multi-Metric Cross-Validation: Is drift determined by considering all three metrics—distance, rate of change, and directionality (PDI)—rather than a single metric?
• Batch Size Optimization: Have you set an appropriate data batch size to balance detection sensitivity and false positive rates?
• Overfitting Monitoring: Are you constantly monitoring the gap between training and test accuracy to evaluate the reliability of PDP-based detection results?
• Adaptive Retraining Trigger: Is there a procedure to analyze the root cause via changed PDP graphs when drift is detected, rather than simply retraining?
• Computational Cost Optimization: For large-scale datasets, have you considered applying techniques such as ‘Compress then Explain’?
• Variable Importance Refresh: Do you recognize that the model’s key variables may change according to the data stream and update important variables periodically?

Author’s Perspectives and Practical Recommendations

The most challenging aspect of building MLOps in the field is establishing a clear criterion for ‘when to retrain the model.’ Most retrain when accuracy drops below a certain level, but this is a reactive prescription. The PDD approach proposed in this paper is a highly intelligent approach in that it measures whether the ‘way the model views the world’ has changed.

It is particularly impressive that it considers not only L2 distance but also the derivative (L2Der) and directionality (PDI). This is because a change in the ‘slope of change’—rather than a simple difference in value—signifies a fundamental shift in business logic. This allows data scientists, when explaining performance drops to executives, to report specifically that ‘the influence of variable A on the result was previously a positive correlation but has now shifted to a negative correlation,’ rather than simply stating ‘accuracy dropped by 5%.’

However, as mentioned in the paper, relying on a single most important variable can be risky in practice. Since multiple variables interact in real services, it seems necessary to expand this into a profile ensemble detection system for Top-K variables.

In conclusion, PDD holds great value by converting the model’s black box into monitoring metrics and can serve as a bridge from performance-centric MLOps to interpretation-centric MLOps.

Conclusion and Enterprise Governance Guide

For internal AI projects to move beyond PoC-stage demos and settle into actual business value, the following governance principles must be established from the planning stage:

1. Interpretability-First Principle: All performance degradation alarms must be accompanied by XAI-based explanations of which variable relationships have changed, not just numerical values.
2. Preemptive Response Principle: Any significant change in variable profiles should be defined as a potential risk and trigger an investigation, even before performance metrics decline.
3. Dynamic Baseline Management Principle: To distinguish between natural evolution of data distribution and concept drift, baseline thresholds must be updated periodically.
4. Overfitting Verification Principle: To ensure the reliability of XAI-based detection tools, the correlation between the model’s generalization performance and detection signals must be constantly verified.
5. Evidence-Based Retraining Principle: Avoid simple periodic retraining; perform model updates only when the business significance of the detected profile change is confirmed.

References and Original Analysis Material