S12-24/bias-in-bios-profession-classifier

Model description

Multinomial classifier predicting a person's profession (28 classes) from the free-text hard_text biography in the LabHC/bias_in_bios dataset (derived from De-Arteaga et al., Bias in Bios, FAccT 2019, arXiv:1901.09451).

This is an educational / research demo project studying gender bias in automated resume screening. It is NOT intended, tested, or suitable for real hiring or HR decisions.

Intended uses & limitations

Intended use: demonstrating and studying gender bias in text-based occupation classifiers. Out of scope: any real candidate screening, resume filtering, or employment decision. See the 'Bias, Risks and Limitations' section below for the full fairness audit.

Training Procedure

TF-IDF vectorizer + logistic_regression (scikit-learn), tuned with randomized search over clf__C (cv=3, scoring=f1_macro).

Best hyperparameters: {'clf__C': 10.0}.

Hyperparameters

Click to expand

Hyperparameter	Value
memory
steps	[('tfidf', TfidfVectorizer(max_df=0.9, max_features=30000, min_df=2, ngram_range=(1, 2),
	sublinear_tf=True)), ('clf', LogisticRegression(C=10.0, max_iter=2000, random_state=42))]
verbose	False
tfidf	TfidfVectorizer(max_df=0.9, max_features=30000, min_df=2, ngram_range=(1, 2),
	sublinear_tf=True)
clf	LogisticRegression(C=10.0, max_iter=2000, random_state=42)
tfidf__analyzer	word
tfidf__binary	False
tfidf__decode_error	strict
tfidf__dtype	<class 'numpy.float64'>
tfidf__encoding	utf-8
tfidf__input	content
tfidf__lowercase	True
tfidf__max_df	0.9
tfidf__max_features	30000
tfidf__min_df	2
tfidf__ngram_range	(1, 2)
tfidf__norm	l2
tfidf__preprocessor
tfidf__smooth_idf	True
tfidf__stop_words
tfidf__strip_accents
tfidf__sublinear_tf	True
tfidf__token_pattern	(?u)\b\w\w+\b
tfidf__tokenizer
tfidf__use_idf	True
tfidf__vocabulary
clf__C	10.0
clf__class_weight
clf__dual	False
clf__fit_intercept	True
clf__intercept_scaling	1
clf__l1_ratio
clf__max_iter	2000
clf__multi_class	deprecated
clf__n_jobs
clf__penalty	l2
clf__random_state	42
clf__solver	lbfgs
clf__tol	0.0001
clf__verbose	0
clf__warm_start	False

Model Plot

Pipeline(steps=[('tfidf',TfidfVectorizer(max_df=0.9, max_features=30000, min_df=2,ngram_range=(1, 2), sublinear_tf=True)),('clf',LogisticRegression(C=10.0, max_iter=2000, random_state=42))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Evaluation Results

Metrics computed on the held-out test split of LabHC/bias_in_bios.

Metric	Value
accuracy	0.7838
f1_macro	0.6818
n_samples	9999

How to Get Started with the Model

[More Information Needed]

Model Card Authors

This model card is written by following authors:

[More Information Needed]

Model Card Contact

You can contact the model card authors through following channels: [More Information Needed]

Citation

Below you can find information related to citation.

BibTeX:

[More Information Needed]

Bias, Risks, and Limitations

This model was audited for gender bias with fairlearn, using gender as the sensitive feature and the True Positive Rate (TPR) gap — the difference in recall between men and women for a given predicted profession — as the primary fairness metric, following the methodology of De-Arteaga et al., Bias in Bios (FAccT 2019, arXiv:1901.09451).

• Mean absolute TPR gap across all professions: 0.1295
• Largest TPR gap: 0.4056, for profession model
• Professions with the most gender-imbalanced training data: dietitian, rapper, nurse, dj, surgeon, paralegal, software_engineer, composer

See fairness_report.json in this repo for the full per-profession breakdown, and the project README's Bias Dashboard for a visual comparison.

Disclaimer: this is a pedagogical / research project. Do not use this model, or its predictions, to make real hiring, screening, or employment decisions.

Accuracy

Accuracy / recall by gender

fairlearn MetricFrame computed on the test split with gender as the sensitive feature.

Click to expand

accuracy	recall_macro
0.00626503	0.0105893
0.787156	0.594779
0.780891	0.584189
0.992041	0.982196