DES646 Project

Dataset Bias and Quality Diagnostics Dashboard

Project Description

This project presents a lightweight, interactive Streamlit dashboard designed to help researchers and designers detect, visualize, and mitigate dataset bias and quality issues before training machine learning models.
By combining principles of explainable AI and visual analytics, it enables transparent and interpretable exploration of image datasets even for users without deep technical expertise.

The tool automatically identifies common dataset issues such as duplicates, class imbalance, outliers, and potential bias-conflicting samples using pretrained vision models. It also supports semi-supervised labeling for partially labeled datasets and provides actionable feedback for improving data quality and fairness.

Key functionalities include:

• Data Diagnostics: Detect duplicates, imbalance, outliers, and low-quality samples.
• Bias & Self-Influence Analysis: Identify samples that disproportionately affect model fairness or accuracy.
• Visual Exploration: Explore class distributions, embedding projections (UMAP), and influence-ranked images.
• Label Completion (Optional): Suggest probable labels for unlabeled images using embedding similarity.
• Actionable Feedback: Recommend relabeling or removing problematic samples to improve dataset quality.

---

How to Run

1. Clone the repository and navigate to the project directory

git clone <repo-url>
cd <repo-name>

2. Install dependencies

Make sure you have Python 3.8 or above, then run:

pip install -r requirements.txt

3. Launch the Streamlit dashboard

streamlit run dashboard/app.py

4. Use the interface

• Choose whether to upload a semi-labeled dataset or run diagnostics on an existing dataset.

• Explore duplicates, imbalance, outliers, and influence scores through interactive visualizations.

• Download reports for duplicate and imbalance analysis.