How lakshmikanthan07/drift-aware-continual-lstm-for-emerging-fraud-detection works
DriftGuard is a streaming fraud-detection and continual-learning framework. Its principal workflow is to generate card-level transaction sequences, score them with a deep sequence model, detect performance drift, and adapt the learner while retaining prior examples; results are evaluated and visualized. The README also describes preprocessing, static baselines, adaptive business thresholds, replay, A-GEM, and EWC. Some learner integrations are unsampled, so uncertain runtime connections are left dashed rather than inferred as confirmed calls.
Main entry points: continual_learner.py Continual Learner · realtime_generator.py Sequence Generator · dual.py Dual Drift Detector · preprocess_data.py Data Preprocessing
Components 18
Stream and Inputs
- Data Preprocessing
src/preprocess_data.py - Prepared Data
- Sequence Generator
src/realtime_generator.py - Feature Metadata
data/metadata.json
Continual Learning
- Continual Learner
src/continual_learner.py - Sequence Model
src/models/lstm_model.py - Adaptive Threshold
src/continual_learner.py - Random Replay
src/memory/random.py - Balanced Replay
src/memory/balanced.py - Prioritized Replay
src/memory/prioritized.py - Gradient and EWC
src/continual_learner.py
Drift Monitoring
- Dual Drift Detector
src/drift_detection/dual.py - ADWIN Detector
src/drift_detection/adwin.py - Recall Window
src/drift_detection/window.py
Evaluation and Results
- Static Baselines
src/baselines.py - Model Evaluator
src/model_evaluator.py - Interactive Dashboard
src/dashboard.py
Other components
- Transaction Data
Connections 16
- Transaction Data → Data Preprocessing: preprocesses
README.md - Prepared Data → Sequence Generator: loads sequences
README.md - Sequence Generator → Feature Metadata: reads features
src/realtime_generator.py - Sequence Generator → Continual Learner: supplies batches
src/continual_learner.py - Continual Learner → Sequence Model: uses model
src/continual_learner.py - Continual Learner → Adaptive Threshold: tunes threshold
src/continual_learner.py - Continual Learner → Random Replay: uses replay
src/continual_learner.py - Continual Learner → Balanced Replay: uses replay
src/continual_learner.py - Continual Learner → Prioritized Replay: uses replay
src/continual_learner.py - Continual Learner → Gradient and EWC: applies mitigation
README.md - Continual Learner → Dual Drift Detector: checks drift
src/continual_learner.py - Dual Drift Detector → ADWIN Detector: updates
src/drift_detection/dual.py - Dual Drift Detector → Recall Window: updates
src/drift_detection/dual.py - Continual Learner → Static Baselines: compares strategies
README.md - Continual Learner → Model Evaluator: evaluates metrics
src/continual_learner.py - Continual Learner → Interactive Dashboard: visualizes stream
README.md