When Concept Drift Is More Than a Distribution Shift

Explaining how non-linear causal relationships reveal hidden process changes.

Most systems can detect that something changed.
Very few can explain why it changed.

Machine learning models rarely fail dramatically.

They degrade quietly. Accuracy drops slowly. Predictions become unstable. Business impact often appears before monitoring dashboards raise alarms.

In many process-driven systems — banking, manufacturing, healthcare — this degradation is caused by concept drift, where the underlying relationships between variables change over time.

But the real challenge is this:

Systems usually detect that drift occurred but they rarely explain what caused it.

This research focuses on closing that gap.

The Limitation of Traditional Drift Detection

Most drift detection techniques rely on:

  • Statistical change detection
  • Window-based distribution comparison
  • Linear Granger causality
  • Prediction error monitoring

These approaches are useful, but they mainly detect symptoms rather than causes.

Traditional Granger causality also assumes linear relationships, however, real-world business processes rarely behave linearly.

Examples include:

  • Workload affecting turnaround time in complex ways
  • Customer attributes influencing outcomes non-linearly
  • Resource allocation interacting dynamically with control-flow behavior

Linear tools often miss these hidden dependencies.

Reframing Drift: A Change in Causal Structure

Instead of asking:

Has the data distribution shifted?

I asked a different question:

Has the causal structure between process variables changed?

Event logs typically contain two categories of information.

Primary Features

Control-flow patterns representing how the process executes.

Secondary Features

Contextual attributes such as age, workload, or resource usage. If contextual variables begin influencing process behavior differently over time, it indicates structural drift.

To capture this, the proposed framework combines:

  • Process mining
  • Change point detection (PELT)
  • Non-linear causal modeling using MLP
  • Statistical validation via Wilcoxon test

Core Idea of the Framework

Two predictive models are trained.

Model 1

Uses only primary process features.

Model 2

Uses both primary and contextual features.

If Model 2 significantly improves prediction accuracy, we infer:

Contextual variables are causally influencing process behavior.

This converts traditional drift detection into:

Causal Drift Explanation

Why Use MLP Instead of Linear Granger?

Traditional Granger causality identifies only linear dependencies.

Real-world systems often contain:

  • Sinusoidal patterns
  • Exponential effects
  • Non-linear lag relationships
  • Multi-variable interactions

Multilayer Perceptrons (MLPs) can naturally model these complex relationships.

The framework therefore:

  • Uses lag-based time windows
  • Employs 2–3 hidden layers with ReLU activation
  • Applies dropout and early stopping
  • Evaluates performance using Mean Squared Error (MSE)
  • Validates causal significance using the Wilcoxon signed-rank test

If contextual variables consistently reduce prediction error — and the improvement is statistically significant, they likely represent true causal influence.

Experimental Validation

The framework was evaluated using three datasets.

Synthetic CPI-generated datasets

Controlled experiments with:

  • Sudden drift
  • Gradual drift
  • Recurring drift

BPI Challenge 2017 dataset

A real-world dataset containing 31,000+ loan application event logs.

Large synthetic non-linear dataset

High-dimensional time series including functions such as:

  • sin
  • log
  • exp
  • sqrt with injected drift.

Key Results

  • Consistent MSE reduction when contextual variables were included
  • Over 85% statistically significant improvements
  • Detection of non-linear dependencies missed by linear models
  • Low false causal detection

These results demonstrate strong robustness, sensitivity, and statistical reliability.

Why This Matters for Industry

In production ML systems, monitoring usually answers:

  • Is model performance degrading?
  • Is the data distribution changing?

But organizations need deeper insight:

  • What caused the degradation?
  • Which contextual factors changed?
  • Is the drift operational, behavioral, or structural?

This research moves toward a more powerful paradigm:

Drift Detection → Drift Explanation → Root Cause Insight

The long-term goal is to build ML systems that:

do not just detect failure — but understand structural change.

Limitations & Future Work

Some challenges remain:

  • Fixed lag windows may miss dynamic temporal dependencies
  • Pairwise modeling limits full multivariate causal discovery
  • Real-world logs lack ground-truth causal labels
  • Hyperparameter tuning remains important

Future directions include:

  • Neural Granger approaches
  • Transformer-based temporal modeling
  • Multivariate structural causal models
  • Self-Adaptive MLOps Pipelines
  • Online adaptive drift monitoring

Conference Presentation

In July 2025, I presented this research paper Explainable Concept Drift in Process Mining Using MLP-Based Non-linear Causality Detection

at the 16th International Conference on Computing, Communication and Networking Technologies (ICCCNT 2025) held at IIT Indore.

Pradeep Pant presenting at ICCCNT 2025 - Certificate of Presentation

Final Thoughts

Concept drift detection alone is no longer sufficient.

Modern ML systems must be:

  • Causally aware
  • Non-linear
  • Statistically validated
  • Interpretable

Drift is not merely a statistical anomaly.

It is a signal that the system’s environment, structure, or context has changed.

Understanding why it changed is where intelligent systems truly begin.

Discussion

This work is part of my ongoing PhD research.
If you have suggestions, feedback, or ideas for collaboration in areas such as concept drift, machine learning, process mining, or adaptive ML systems, I would love to hear your perspective.

Feel free to leave a comment or reach out via email.