COMPASS is a framework for online continual next activity prediction using Foundation Models (FMs). It addresses the cold-start problem and catastrophic forgetting that affect existing Predictive Process Monitoring (PPM) approaches by combining parameter-efficient fine-tuning with unsupervised drift detection and gradient-based knowledge consolidation — without requiring explicit task boundaries.
COMPASS/
│
├── Data/ # Raw event logs (.csv)
├── Methods/
│ └── COMPASS/
│ ├── run.py
│ ├── engine.py # Streaming loop, drift detection, training
│ └── keeplora_handler.py # Subspace management and LoRA reinitialization
├── Utils/
│ ├── preprocess.py # Event log parsing and sequence extraction
│ └── metrics.py # Accuracy, F1, and per-event logging
├── runs/ # Output directory for results and CSVs
├── requirements.txt
└── README.md
python -m venv venv
source venv/bin/activate
pip install -r requirements.txtRequires Python ≥ 3.10 and a CUDA-capable GPU. Experiments were run on NVIDIA Tesla V100 (16 GB).
Place raw event logs (.csv format) in the Data/ directory. Preprocessing, timestamp formatting, and prefix sequence extraction are handled automatically by Utils/preprocess.py. Preprocessed objects are cached as .pkl files under Preprocessed/ to speed up repeated runs.
python -m Methods.COMPASS.run \
--dataset Data/BPIC2015_Recurrent.csv \
--model-name arnir0/Tiny-LLM \
--window-size 100 \
--seed 42python -m Methods.COMPASS.run \
--dataset Data/BPIC2015_Recurrent.csv \
--model-name distilbert/distilgpt2 \
--oracle-drift \
--seed 42Hyperparameters were tuned on the first 15% of each event log (validation split). The remaining 85% is used for online evaluation.
| Parameter | Values |
|---|---|
| Window size (W, reffered to as Γ in the paper) | {100, 500, 1000} |
| Learning rate (η) | {2e-3, 2e-4, 2e-5} |
| Variance threshold (τ) | {5e-2, 5e-3, 5e-4} |
| LoRA rank (r) | {8, 64, 256, 512} |
| LoRA alpha (α) | {16, 128, 512, 1024} |
All runs use Γ = 100, energy thresholds ϵᵥᵥ = 0.75, ϵ_f = 0.95, and max subspace rank R = 64.
