Concept Drift Detection and Adaptation Pipeline

Pipeline for detecting sudden, gradual, and recurring concept drift in a data stream, with model pool and incremental adaptation (retrain/fine-tune).

Workflow (high level)

1. Data stream → Preprocessing (prediction errors from y, ŷ)
2. Meta-detector (voting ensemble) → if drift: trigger ECPF or RCD recurring check
3. Recurring drift detector (inputs: prediction errors, drift timestamp, concept memory) → Recurring? Yes → restore model from pool; No → Drift type classifier
4. Drift type classifier → Sudden vs Gradual vs Incremental
5. Model pool + Prediction model → Incremental adaptation on drift
6. Offline evaluation → Correct Detection score, precision/recall/F1, prediction MAE

Setup

1. Clone and enter the project directory:

   git clone https://github.com/Diteaphr/drift_detect.git
   cd drift_detect

2. Create and activate a virtual environment:

   python -m venv .venv
   source .venv/bin/activate   # macOS/Linux

3. Install dependencies:

   pip install -r requirements.txt

Run full pipeline

python main.py [tsv|dwm|statistical|dwme]

Loads data/sudden_drift/recurring_sudden_sea100k_g03.csv, runs the full pipeline, then prints Correct Detection score, precision/recall/F1, and prediction MAE.

Optional: save a timeline plot after evaluation:

python main.py dwm --plot                          # saves to outputs/recurring_sudden_sea100k_g03_dwm_cd.png
python main.py dwm --plot --plot-path my_plot.png  # custom path

The plot shows rolling accuracy, perturbation intervals (drift interval + 1000 samples), and each alert coloured green (TP) or red (FP).

To only record alert timestamps, from the project root:

python scripts/collect_alert_times.py --csv path/to/data.csv --out alerts.json

Optional flags: --meta tsv|dwm|statistical, --config-json, --ecpf.

Run ECPF on a recurring-drift CSV

python run_ecpf_recurring.py \
  --csv data/recurring_drift/recurring_sud_sea100k_g00.csv \
  --signal-mode oracle_60

Prints a Correct Detection score at the end and saves an event log CSV and timeline PNG to outputs/.

Batch across g00–g09:

python run_ecpf_recurring_batch.py \
  --signal-mode oracle_60 \
  --data-dir data/recurring_drift

Each row reports cd(TP=…, FP=…, N=…, score=…%). The final summary prints the mean Correct Detection score across all files.

Correct Detection score

Score formula (Bifet et al.):

score = max(0, (TP - FP) / N) × 100

• N — number of ground-truth drift intervals in the stream
• TP — each interval counts at most once, regardless of how many alerts fall inside it
• FP — each alert outside every interval counts as one false positive
• Perturbation interval — ground-truth drift interval extended by +1000 samples on the right (e.g. [250, 2000] → [250, 3000])

Implementation:

File	Role
src/metrics/correct_detection.py	compute_correct_detection, build_perturbation_intervals — pure calculation
tests/evaluation.py	correct_detection_from_detections — wraps DriftDetection objects
main.py	Calls evaluation, optionally plots results
run_ecpf_recurring.py	Prints score after each single-file run
run_ecpf_recurring_batch.py	Adds cd_tp/fp/n/score_pct columns to summary CSV

Use pipeline on your own stream

from src.config import PipelineConfig
from src.pipeline import ConceptDriftPipeline
import numpy as np

config = PipelineConfig(update_batch_size=50, recurrence_threshold=0.5)
pipeline = ConceptDriftPipeline(config=config)
pipeline.warm_start(X[:100], y[:100])

for i in range(100, len(X)):
    y_pred, detections, drift_occurred = pipeline.step(X[i], y[i], index=i)
    for d in detections:
        print(f"Drift at {d.timestamp}: {d.drift_type.value}")

Use recurring detector only

See tests/example_usage.py: pass prediction errors, drift alert timestamp, and a ConceptMemory instance; get back recurring: bool.

python -m tests.example_usage