A lightweight GUI Python toolbox for detecting, visualizing, and understanding data shift.

---

✨ Overview

DomainSAT is a simple, powerful, and interactive domain (data) shift analysis toolbox built in Python. It runs on macOS, Linux, and Windows directly through a browser-based interface, with no installation and no coding required. It helps researchers and practitioners easily detect, quantify, and visualize domain shift across datasets.

🔎 Domain (Data) Shift Detection Methods

• Univariate Statistical Tests

  • Kolmogorov–Smirnov (KS)
  • Mann–Whitney U
  • Cramér–von Mises
  • Chi-square (categorical)

• Distance & Divergence Metrics

  • MMD (per-feature)
  • MMD (multivariate)
  • Wasserstein distance (per-feature)
  • Wasserstein distance (multivariate)
  • KL Divergence
  • JS Divergence
  • Mahalanobis distance

• Classifier-Based Drift Detection

  • Domain classifier (AUC-based)
  • C2ST – Logistic Regression
  • C2ST – Random Forest

• Representation-Based Detection

  • Autoencoder

📊 Visualization

• Per-feature distributions (Histogram + KDE (kernel density estimation) Curve)
• UMAP embedding (2D projection)
• PCA projection (2D projection)
• Interactive feature selection

---

🔧 Prerequisites

One of the core design goals of DomainSAT is simplicity:

• You do NOT need to install the package system-wide, build wheels, or configure environments.
• Just copy the project folder to any location and run it directly.

✔️ What you must have

• Python 3.8+ installed
  (Windows, macOS, or Linux all supported)

• A modern web browser
  (Chrome, Safari, Firefox, Edge, ...)

✔️ Minimal Python dependencies

DomainSAT requires only a few lightweight packages:

• streamlit umap-learn
• numpy, pandas, scikit-learn, scipy, matplotlib (typically pre-installed in most environments)

You may install them using pip:

pip install streamlit umap-learn numpy pandas scikit-learn scipy matplotlib

or using conda:

conda install -c conda-forge streamlit numpy pandas scikit-learn scipy umap-learn matplotlib

---

🚀Usage

👉Start

Place the project folder anywhere on your computer. Once dependencies are installed, simply open a terminal (or Command Prompt on Windows), and run (in the path of the project folder):

streamlit run DomainSAT.py

Then the toolbox will launch in your browser. The first startup may take some time (approximately 20 seconds).

---

👉Upload your datasets

Upload two CSV files:

1. Source dataset
2. Target dataset

DomainSAT will automatically detect the shared features and prepare the data for analysis.

Note:
The first row of each CSV must contain the feature names (column headers).
If your data consists only of embeddings without headers, please insert a header row (e.g., f1, f2, f3, ...) before loading.

For convenience, several sample datasets are provided in the folder Exp Data for quick testing and validation.

---

👉Choose a domain shift-detection method

From the list, you can choose one shift detection method. Currently, DomainSAT provides several categories of methods:

• Statistical tests: KS, Mann-Whitney U, Cramér-von Mises, Chi-square
• Distance-based: MMD, Wasserstein, Mahalanobis, KL/JS Divergence
• Classifier-based: Domain classifier (AUC), C2ST (Logistic), C2ST (Random Forest)
• Representation-based: Autoencoder

Adjust method-specific parameters (p-value threshold, distance threshold, AUC threshold, etc.) from the sidebar.

---

👉Run the analysis

Click "Run Shift Detection" button, DomainSAT will compute shift metrics and display:

• Summary table of results
• Whether shift is detected (per feature or global)
• Downloadable CSV report

---

👉Visualize the shift

Feature-level histograms

• You can select any feature to inspect and visualize its distribution as a histogram for both the source and target datasets.
• For each feature, overlay source vs target distributions with KDE curves for easy inspection.

Set-level distribution

DomainSAT provides two projection methods: UMAP and PCA, that reduce high-dimensional data into a 2D space for visualizing the distribution of source and target datasets. Simply click the checkbox, and the corresponding visualization will be generated.

Note: UMAP typically takes longer time to compute, while PCA is much faster.

---

Copyright and License

© 2025 Hao Guan. All rights reserved.

DomainSAT (Domain Shift Analysis Toolbox) is an original work developed by the author and publicly released under the Apache License, Version 2.0.

Any reuse, redistribution, or derivative work must comply with the license terms, including proper attribution, preservation of copyright notices, and inclusion of the original license.

Unauthorized copying of the project name, repository structure, documentation, or core design without proper attribution or license compliance constitutes a violation of the license and may result in formal takedown actions.