Resource-Centric-NAP

Overview

This repository implements a research pipeline for resource-centric next activity prediction using process event logs. The core question is how different sequence encoding techniques affect predictive performance on downstream classifiers.

The project evaluates 7 embedding and encoding approaches using two classifiers:

• Random Forest (run_experiments_rf.py)
• LightGBM (run_experiments_lgbm.py)

It also includes an experiment script to measure encoding complexity and a results analysis utility.

Key Concepts

• Resource-centric prediction: The dataset is split by resource (org:resource), so resources in the test set do not appear in the training set.
• Next activity prediction: For each partial sequence of resource activity, predict the next activity.
• Prefix strategies:
  • prefix: fixed prefix of the first n activities
  • last_k: last k activities before the target point
  • sliding_window: windowed prefixes across the trace

Encoding / Embedding Methods

The repository supports seven methods:

1. Baseline - position-wise label encoding of the prefix
2. OHE - one-hot encoding of positional activity features
3. Bigram - bigram transition count features from activity sequences
4. W2V - Word2Vec embeddings trained on activity prefixes
5. D2V - Doc2Vec embeddings trained on full resource traces
6. BERT - custom BERT-style encoder pre-trained on activity sequences
7. ACF - Activity Context Frequency embeddings with PMI post-processing

Datasets

The repository uses BPI Challenge event logs stored in datasets/:

• BPI_Challenge_2013_Incidents.xes
• BPI_Challenge_2017.xes
• BPI_Challenge_2018.xes
• BPI_Challenge_2019.xes

These logs are loaded and converted to a resource-centric format containing:

• case:concept:name
• concept:name
• org:resource
• time:timestamp

Main Scripts

run_experiments_rf.py

Run Random Forest experiments across configured datasets, strategies, prefix lengths, and embedding methods.

run_experiments_lgbm.py

Run LightGBM experiments across configured datasets, strategies, prefix lengths, and embedding methods.

run_complexity_experiments.py

Measure encoding complexity for each method, including feature count, matrix memory size, and encoding time.

win_counts.py

Analyze existing experiment CSV results and compute ranking statistics such as:

• total wins per method
• wins per dataset/model configuration
• total points using a top-7 scoring system

Project Structure

• modules/
  • data_loader.py - dataset import, cleaning, resource split, and prefix generation
  • rf_trainer.py - Random Forest training and grid search utilities
  • lgbm_trainer.py - LightGBM training and grid search utilities
  • encoders/ - encoding and embedding implementations
    • baseline.py
    • one_hot_encoding.py
    • bigram.py
    • word2vec.py
    • doc2vec.py
    • bert.py
    • acf.py
• ACF_code/ - custom implementation of activity-context frequency / PMI embedding generation
• datasets/ - input XES event logs
• results/ - experiment output CSV files
• plots/ - generated visualizations and plots
• notebooks/ - exploratory analysis notebooks

Requirements

The code uses the following Python packages (at least):

• pandas
• numpy
• pm4py
• scikit-learn
• gensim
• torch
• transformers
• lightgbm
• matplotlib
• seaborn

If you use the LightGBM GPU settings in modules/lgbm_trainer.py, ensure a compatible GPU environment is available or update the trainer configuration to use CPU.

Usage

1. Install dependencies in your Python environment.
2. Place the target XES datasets in datasets/.
3. Run the desired script:

python run_experiments_rf.py
python run_experiments_lgbm.py
python run_complexity_experiments.py
python win_counts.py

4. Inspect generated CSVs in results/ and plots/ for visual summaries.

Notes

• Most experiment scripts are configurable via the constants defined at the top of each file (DATASETS, PREFIX_LENGTHS, STRATEGIES, METHODS, USE_GRID_SEARCH).
• The repository is designed for research experimentation.
• The data_analysis.ipynb and notebooks/RandomForest.ipynb notebooks contain additional exploratory work and visualization examples, the latter might not be completely up-to-date with the actual encoders, as this was mainly used for first experimentation.