Reproduce “Comparison of Feature Extraction Methods and Predictors for Income Inference” on OpenResearch

About this paper

Martin Fixman, Martin Minnoni, Carlos Sarraute

Researchers evaluated various feature extraction techniques and machine learning models for inferring user income levels from mobile phone communication patterns. The study found that a Bayesian graph-based approach achieved the highest AUC of 0.746 on the inner graph, while Random Forest classifiers with categorical ego-network features performed best on the full graph, with an accuracy of 0.614.

Key insights

• Categorical features, which discriminate communication based on the income level of connected partners, consistently improved prediction performance.
• Extending ego network depth beyond two levels (Ring₂/Cat₂) did not consistently yield additional performance improvements, sometimes leading to slight degradation due to noise.
• The Bayesian method, which directly leverages communication graph structure and homophily, demonstrated superior performance on portions of the graph with known neighbor income.

Summary by alphaXiv.