v0.3.0 — Multi-Feature Fairness Auditing & Representation Diagnostics

[Khanz9664]

TrustLens v0.3.0: The Multi-Feature Fairness & Representation Update

We are excited to announce the release of TrustLens v0.3.0! 🚀

This update marks a major milestone in our mission to make model reliability diagnostics more accessible and high-resolution. v0.3.0 introduces a powerful orchestration layer for fairness diagnostics, advanced latent space visualizations, and a new engine for comparative model auditing.

---

🌟 Highlights

⚖️ Multi-Feature Fairness Diagnostics

TrustLens now supports high-resolution fairness auditing across multiple sensitive features simultaneously. The TrustReport.plot_bias() API has been upgraded with an opt-in multi_feature mode that generates independent diagnostics for every demographic group in your dataset (e.g., Age, Gender, Ethnicity) in a single call.

• Per-Feature Figures: Return structured dictionaries of figures for seamless integration into reports.
• Mode Orchestration: Support for "subgroup", "equalized_odds", and "gap" modes across all features.
• Deterministic Logic: Guaranteed stable feature ordering and consistent structure.

🌌 Deep Representation Analysis

Go beyond class-level metrics with our new representation diagnostic tools. We've introduced 2D embedding projections and separability scoring to help you visualize why your model might be struggling.

• 2D Embedding Visualization: Automatic UMAP/t-SNE/PCA projections with class-colored scatter plots.
• Separability Metrics: Silhouette scores and within/between-class distance analysis to quantify latent space quality.
• Automatic Fallbacks: Intelligent fallback logic ensures you always get a visualization, regardless of your environment's installed dependencies.

🤖 Model Comparison & Pattern Detection

Compare models head-to-head with the new trustlens.compare API. Identify "Calibration Drift" and "Confidently Wrong" patterns to surface high-level semantic risks that simple accuracy scores often hide.

---

🛠 What's New?

Added

• Multi-Feature Support: TrustReport.plot_bias(multi_feature=True) for parallel feature analysis.
• Embedding Projections: plot_embedding_2d with automatic dimensionality reduction.
• New Metrics: embedding_separability and enhanced equalized_odds() validation.
• Comparative Audit: trustlens.compare for multi-model benchmarking.
• Explainable Scores: Ranked explanation layer for Trust Score deductions.

Improved

• Deterministic Ordering: Canonical ordering for all bias and fairness visual results.
• Memory Hygiene: Automated figure cleanup in high-volume analysis pipelines.
• Rich Docstrings: Detailed return-shape matrices and usage examples for all primary APIs.

---

🤝 Contributors

A huge thank you to the community for making this release possible!

Special thanks to:

• @komoike-oss28-ui for the multi-feature visualization orchestration and fairness metric enhancements.
• @WeiGuang-2099 for the embedding visualization pipeline and representation metrics.
• @sidharth-vijayan for the bias analysis demonstrations.
• @MustansirNisar for security infrastructure.

---

Get Started

Update to the latest version via pip:

pip install trustlens --upgrade

Check out the updated Demo Notebook to see v0.3.0 in action!

---

This discussion was created from the release v0.3.0 — Multi-Feature Fairness Auditing & Representation Diagnostics.