A machine learning approach to predict ethnicity using personal name and census location in Canada

Area of Science:

• Computational Social Science
• Population Health Informatics
• Machine Learning Applications

Background:

• Canada's ethnic diversity is underrepresented in large databases, hindering population research and interventions.
• Automated ethnicity classification via machine learning shows promise for addressing this data gap.
• The performance of such methods in the Canadian context remains largely unexamined.

Purpose of the Study:

• To develop and evaluate a large-scale machine learning framework for predicting ethnicity in Canada.
• To assess the utility of name and census location features for ethnicity classification.
• To determine the performance of automated ethnicity prediction across diverse Canadian ethnic groups.

Main Methods:

• Utilized the 1901 Canadian census data, encompassing 4,812,958 unique individuals.
• Developed multiclass and binary classification machine learning pipelines using logistic regression, C-support vector, and naïve Bayes.
• Incorporated name features (string, substrings, phonetic, entity patterns) and location features (province, district, subdistrict).

Main Results:

• Achieved 76% F1 score and 91% accuracy for multiclass classification.
• Binary classification F1 scores ranged from 68% to 95% (median 87%) for Chinese, French, Italian, Japanese, Russian, and others.
• Performance was lower for English, Irish, and Scottish (F1 63-67%) due to shared heritage; adding location data improved Aboriginal classification F1 from 50% to 84%.

Conclusions:

• An automated machine learning approach using name and census location features can predict Canadian ethnicity.
• Predictive performance varies across ethnic categories.
• Census location features significantly enhance the prediction accuracy for Aboriginal classifications.