cooprecsys 0.0.1rc2

Koperasi Recommender System - Ultra-optimized collaborative filtering using Cython + OpenMP

CoopRecSys

Koperasi Recommender System ML/AI Module A production-grade machine learning and AI module for building intelligent recommendation systems tailored for cooperative (koperasi) product recommendations. This system combines collaborative filtering, learning-to-rank techniques, and explainable AI dashboards.

---

Table of Contents

• Features
• Tech Stack
• Installation
• Quick Start
• Usage Examples
• Architecture
• Dashboard & Explainability
• API Reference
• Configuration
• Testing
• Contributing
• Author
• License

---

Features

Advanced Recommendation Models

• AryColBring: Ultra-optimized collaborative filtering powered by Cython + OpenMP

  • Multi-loss support: Logistic, WARP, BPR, WARP-kOS
  • Sparse matrix acceleration with DuckDB integration
  • Per-thread buffers with no Python GIL in hot paths

• LTR-LightGBM: Learning-to-Rank using LightGBM

  • Group-aware train/test splitting
  • Ranking metrics: NDCG, MAP, AUC
  • MLflow experiment tracking

Explainable AI Dashboard

• Interactive HTML-based dashboard powered by JavaScript
• Feature importance visualization
• Model prediction explanations
• Real-time ranking visualization
• SHAP-style local interpretability

Performance Optimizations

• 69.7% Python for core logic and data processing
• 9.7% Cython for high-performance numerical kernels
• 14.9% CSS for responsive UI styling
• 2.9% JavaScript for interactive dashboards
• Multi-threading support with joblib parallelization
• Sparse matrix operations with SciPy

Production Ready

• CI/CD Pipeline with GitHub Actions
• Comprehensive error handling and logging
• Model persistence with cloudpickle
• DuckDB-backed data ingestion
• TQDM progress bars for user feedback

---

Tech Stack

Component	Technology	Version
Language	Python	3.10+
Performance	Cython	0.29+
ML Framework	LightGBM	Latest
Matrix Ops	NumPy, SciPy	1.21+, 1.7+
Data Processing	Pandas, DuckDB	1.3+, 0.8+
Experiment Tracking	MLflow	1.20+
Parallelization	joblib	1.1+
Visualization	Matplotlib, Seaborn	3.4+, 0.11+
Frontend	HTML5, CSS3, JavaScript	Modern

---

Installation

Prerequisites

# System dependencies (Ubuntu/Debian)
sudo apt-get install build-essential python3-dev gcc g++ make

# macOS
brew install gcc llvm libomp

Setup

1. Clone the repository

   git clone https://github.com/masterofray/cooprecsys.git
   cd cooprecsys
   

2. Create virtual environment

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   

3. Install dependencies

   pip install --upgrade pip
   pip install -r requirements.txt
   

4. Compile Cython extensions (optional, for performance)

   python setup.py build_ext --inplace
   

---

Quick Start

Using the Test Suite

The test suite demonstrates how to use both recommendation models:

LTR LightGBM Example

python -m test.ltrlgbm_test.ltrlgbm_example

Test location: test/ltrlgbm_test/ltrlgbm_example.py

This script:

• Loads sample data from data/sampledata.parquet
• Performs group-aware train/test splitting
• Trains the LightGBM LTR model with hyperparameter tuning
• Generates rankings and MLflow tracking
• Produces HTML reports and artifact visualizations

LTR Inference Test

python -m test.ltrlgbm_test.ltrlgbm_inferencing

Test location: test/ltrlgbm_test/ltrlgbm_inferencing.py

This script demonstrates inference using trained models with fallback strategies.

Collaborative Filtering Tests

pytest test/arycolbring_tests/test_model.py -v
pytest test/arycolbring_tests/test_evaluation.py -v
pytest test/arycolbring_tests/test_data_utils.py -v

Test locations:

• test/arycolbring_tests/test_model.py - Model initialization and fitting
• test/arycolbring_tests/test_evaluation.py - Evaluation metrics (Precision@k, Recall@k, AUC)
• test/arycolbring_tests/test_data_utils.py - Data loading and preprocessing
• test/arycolbring_tests/test_cross_validation.py - Train/test splitting strategies

---

Usage Examples

Example 1: LTR LightGBM Pipeline

from src.configs import LTRConfig
from src.models.ltr_lgbm import lgbm_fit_transform
import pandas as pd

# Load your data
data = pd.read_parquet('data/sampledata.parquet')

# Split data by customer ID (group-aware)
from sklearn.model_selection import GroupShuffleSplit
gss = GroupShuffleSplit(n_splits=1, test_size=0.2)
train_idx, test_idx = next(gss.split(data, groups=data['CustomerID']))
train_df = data.iloc[train_idx]
test_df = data.iloc[test_idx]

# Configure model
config = LTRConfig.from_ini('src/configs/configuration.ini', 
                             features=['ProductName', 'ProductPrice', ...])
config.feature.label = 'CategoryID'
config.feature.query_id = 'CustomerID'

# Train with hyperparameter tuning
trainer = lgbm_fit_transform(
    config=config,
    train=train_df,
    test=test_df,
    run_tuning=True,
    run_name="my_recommendation_model"
)

# Access results
print(f"Best Iteration: {trainer.best_iteration}")
print(f"Runtime: {trainer.runtime_minutes} minutes")

Example 2: Collaborative Filtering with AryColBring

from src.models.arycolbring import AryColBring
import scipy.sparse as sp
import numpy as np

# Create sparse interaction matrix (n_users × n_items)
interactions = sp.coo_matrix(
    (np.ones(1000), (np.random.randint(0, 100, 1000), 
                     np.random.randint(0, 50, 1000))),
    shape=(100, 50)
)

# Initialize model
model = AryColBring(
    no_components=32,
    loss='warp',  # or 'logistic', 'bpr', 'warp-kos'
    learning_rate=0.05,
    random_state=42
)

# Train
model.fit(interactions, epochs=10, num_threads=4)

# Get recommendations
user_id = 0
item_scores = model.predict([user_id] * 50, np.arange(50))
top_items = np.argsort(item_scores)[::-1][:10]

---

Architecture

cooprecsys/
├── src/
│   ├── models/
│   │   ├── arycolbring/          # Collaborative filtering
│   │   │   ├── model.py
│   │   │   ├── cy/               # Cython kernels
│   │   │   ├── evaluation.py
│   │   │   └── cross_validation.py
│   │   ├── ltr_lgbm/             # Learning-to-Rank
│   │   │   ├── model.py
│   │   │   ├── pipeline.py
│   │   │   └── explainer.py
│   │   └── dashboard/            # AI Explainability
│   │       ├── index.html        # Main dashboard
│   │       ├── styles.css        # Responsive styling
│   │       └── app.js            # Interactive logic
│   ├── configs/
│   │   ├── configuration.ini
│   │   └── logger.py
│   └── utils/
│       ├── data_utils.py
│       └── metrics.py
├── test/
│   ├── arycolbring_tests/        # CF unit tests
│   │   ├── test_model.py
│   │   ├── test_evaluation.py
│   │   └── test_data_utils.py
│   └── ltrlgbm_test/             # LTR integration tests
│       ├── ltrlgbm_example.py    # ← Main test script
│       └── ltrlgbm_inferencing.py
├── .github/workflows/
│   └── pipeline.yml              # CI/CD automation
├── img/
│   └── logo_navi.jpg             # Project logo
├── requirements.txt
└── README.md

---

Dashboard & Explainability

CoopRecSys Explainable AI Dashboard is a lightweight web interface implemented with Jinja2 templates, JavaScript, and CSS. It is intended for production monitoring and diagnostic workflows of an LTR LightGBM ranking model, presenting concise, actionable metrics, temporal trends, and dataset context. The dashboard enables data scientists and engineers to rapidly assess model health, detect anomalies or drift, and investigate root causes of performance changes through an integrated explainability‑focused view. The dashboard provides real-time explanations for model predictions:

Features

• Feature Importance: Visualization of which features drive recommendations
• Ranking Explanations: Why specific items are ranked higher
• Comparison View: Side-by-side model performance comparison
• Interactive Filters: Drill down by user, item, or category
• Export Reports: Generate PDF/HTML reports with explanations

Running the Dashboard

# Start the web server
python -m http.server 8000 --directory ./artifacts/reports

# Open browser to http://localhost:8000/20260528_training_report.html

Overviews Page

Figure 1. Overview of the CoopRecSys Explainable AI Dashboard.

Figure 1 displays the primary overview screen of the dashboard, combining a high‑level scorecard, temporal visualizations, navigation tabs, and dataset context to provide an immediate assessment of model status. Key elements and purpose:

• Header — identifies the dashboard and the monitored model for orientation.
• Navigation Tabs — Overview, Rankings, Diagnostics, Config for structured access to analytical modules.
• Scorecard Metrics — compact presentation of prediction statistics (PRED MAX, PRED MEAN, PRED MIN, PRED STD) and ranking performance (NDCG@5, NDCG@10 for train and test) for rapid appraisal.
• Control Panel Visitor Analytics — summary count of recent predictions and an interactive multi‑series line chart to reveal trends, spikes, or drift.
• Sidebar Dataset Statistics — contextual counts such as USERS, PRODUCTS, CATEGORIES, and ROWS to indicate scale and coverage.
• Multi‑series Line Chart — overlays prediction and evaluation metrics to facilitate correlation analysis and anomaly detection.

---

Rankings Page

Figure 2. Rankings CoopRecSys Dashboard.

The Rankings page provides a transparent, interactive view of model inference results and the feature context that produced each ranking. It is intended for analysts and engineers who require a sortable, filterable listing of top predictions together with per‑row explainability signals so that individual decisions can be inspected, validated, and traced back to input features.

Key Components

• Ranking Results Table — Primary component showing the top N predictions produced by the LTR LightGBM model with configurable columns for identifiers, features, prediction score, and explainability metrics.
• Row Explainability Panel — Per‑row detail pane that surfaces feature contributions (e.g., SHAP values), top contributing features, and short textual explanation for the predicted rank.
• Filters and Facets — Controls to restrict the table by date range, user segment, product category, prediction score range, or custom tags.
• Sorting and Pagination — Stable, server‑side or client‑side sorting by score and any feature column, with efficient pagination for large result sets.
• Export and Snapshot — Export current view to CSV and capture a snapshot (timestamped) of the displayed ranking for audit or reporting.
• Contextual Metadata — Small summary area showing dataset scope (rows, users, products), generation timestamp, and model version used for the ranking.

Interactions and Controls

• Global Filters — Date range picker; dropdowns for product category and user segment; numeric sliders for prediction score and years working.
• Column Filters — Per‑column quick filters (text search, numeric range).
• Row Inspection — Clicking a row opens the Row Explainability Panel with:
  • Full feature vector for that row.
  • SHAP waterfall or bar chart showing positive and negative contributions.
  • A short natural language explanation generated from the top contributions.
• Compare Mode — Select two or more rows to view a side‑by‑side comparison of features and contributions.
• Server Mode — For large datasets, enable server‑side pagination and sorting; otherwise use client‑side DataTables for small to medium result sets.
• Audit Trail — Each exported snapshot includes metadata: model version, timestamp, and filter state.

---

Diagnostics Page

Figure 3. diagnostics Graph for CoopRecSys model.

The Diagnostics page provides a consolidated environment for model introspection and validation. It combines global diagnostics (feature importance and distributional checks), prediction‑level diagnostics (relevance score histograms and drift indicators), and per‑sample explainability artifacts (SHAP summaries and downloadable SHAP files). The page is intended for data scientists, ML engineers, and auditors who require both high‑level signals and the ability to drill into individual explanations.

Key Components

• Feature Importance Chart — Horizontal bar chart showing top features by chosen importance metric (GAIN, SPLIT, or permutation importance). Interactive: sort, change metric, and toggle top‑K.
• Histogram of Relevance Predictions — Binned histogram of model relevance scores (test / production) with overlayed reference distribution (train) and summary statistics (mean, median, std, skewness).
• SHAP Samples Panel — A sample browser that lists available SHAP files (timestamped), allows download, and previews selected samples with a SHAP waterfall or bar chart and raw feature vector.
• Drift & Distribution Alerts — Small indicator cards that flag features with significant distributional shift (KS test, PSI) and prediction drift (population mean shift).
• Sample Inspector — On selecting a sample from the SHAP list or from the top predictions, show: raw features, SHAP contributions (positive/negative), cumulative contribution to score, and a short natural‑language explanation.
• Export & Audit — Buttons to export diagnostics snapshot (CSV/JSON) and to attach model version, feature engineering commit, and timestamp for reproducibility.

Operational and Implementation Notes

• Server responsibilities

  • Precompute feature importance (GAIN/SPLIT) and expose as JSON.
  • Provide binned relevance distributions for train/test/production to avoid heavy client computation.
  • Serve SHAP files on demand and paginate sample lists for large files.

• Performance

  • For large SHAP files, fetch only sample metadata for the list and request full sample SHAP vectors when the user inspects a sample.
  • Use server‑side aggregation for histograms and KS/PSI calculations.

• Accessibility & UX

  • Ensure charts have aria-label and textual summaries for screen readers.
  • Allow keyboard navigation for the SHAP sample list and close preview with Esc.
  • Provide clear tooltips explaining each diagnostic metric (e.g., GAIN vs SPLIT).

• Reproducibility & Audit

  • Every diagnostics snapshot must include model version, feature engineering commit hash, data window, and timestamp.
  • Exported diagnostics should embed this metadata.

---

Configs Page

Figure 4. Configs table as parameter LTR LGBM.

The Config page centralizes model configuration and hyperparameter management for the LTR LightGBM ranking model. It provides a controlled interface to view, edit, validate, version, and apply training and inference parameters while preserving auditability and reproducibility. The page is intended for ML engineers and platform operators who must safely tune model behavior in production or prepare reproducible training runs.

Key Components

• Configuration Table — Tabular display of current parameter names and values (e.g., objective, metric, ndcg_eval_at, learning_rate, max_depth, num_leaves, feature_fraction, bagging_fraction, bagging_freq, lambda_l1). Each row shows Parameter, Value, Type, Source (default / experiment / production), and Last modified timestamp.
• Edit Controls — Inline editors for editable parameters with appropriate input types: numeric fields, dropdowns for enumerated options, multi-value arrays for list parameters, and toggles for booleans. Edits are staged until explicitly saved.
• Validation Engine — Client and server validation rules that enforce type constraints, allowed ranges, and inter‑parameter consistency (e.g., num_leaves consistent with max_depth, feature_fraction in (0,1]).
• Preview and Dry Run — A preview panel that shows the effective configuration JSON and a dry‑run button that triggers a lightweight validation job (no training) to check compatibility with current feature schema and training pipeline.

---

Configuration

Environment Variables

# MLflow tracking
export MLFLOW_TRACKING_URI=file:./mlruns
# or for remote server
export MLFLOW_TRACKING_URI=http://localhost:5000

# Logging
export LOG_LEVEL=DEBUG

Configuration File

Edit src/configs/configuration.ini:

[model]
loss = warp
num_components = 32
learning_rate = 0.05
epochs = 10

[data]
test_size = 0.2
random_state = 42

[mlflow]
experiment_name = cooprecsys_prod

---

Testing

Run All Tests

# Integration tests
python -m test.ltrlgbm_test.ltrlgbm_example

# Unit tests
pytest test/arycolbring_tests/ -v --tb=short

# With coverage
pytest test/arycolbring_tests/ --cov=src --cov-report=html

CI/CD Pipeline

All tests run automatically on:

• Push to master or dev branches
• Pull requests
• Manual workflow dispatch

Status Badge:

Test Results

Tests verify: [x] Model initialization and parameter validation [x] Data loading and preprocessing [x] Training and inference [x] Evaluation metrics computation [x] Cross-validation splitting strategies [x] Security checks (Bandit)

---

Performance

Benchmarks (Sample Dataset)

Model	Training Time	Inference Time	Memory
AryColBring (WARP)	2.1s	0.3s	45MB
LTR-LightGBM	5.4s	0.8s	120MB
Ensemble	8.2s	1.2s	180MB

Dataset: 100k interactions, 5k users, 2k items Hardware: 4-core CPU, 8GB RAM

---

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Add tests for new functionality
4. Run tests locally (pytest test/arycolbring_tests/)
5. Commit changes (git commit -m 'Add amazing feature')
6. Push to branch (git push origin feature/amazing-feature)
7. Open a Pull Request

Development Guidelines

• Follow PEP 8 style guide
• Add docstrings to all functions
• Include type hints
• Write unit tests for new code
• Update README for new features

---

Author

Aryanto (masterofray)

• Email: [e-mail]
• GitHub: @masterofray
• Role: Author & Maintainer

---

License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.

---

References

• AryColBring Model Documentation
• Collaborative Filtering - ACM
• Learning-to-Rank Overview
• LightGBM Documentation
• MLflow Documentation
• Cython Documentation

---

Support & Issues

• Report Issues: GitHub Issues
• Documentation: See /docs folder
• Email: [e-mail]

---

Built with ❤️ for better product recommendations in cooperative systems

If you find this useful, please star the repository!