Census2Vec: Enhancing Socioeconomic Predictive Models with Geo-Embedded Data

Artificial intelligence’s role in distilling insights from data has emerged as a pivotal solution to contemporary challenges within our data-centric society. Census data, while offering rich demographic and socioeconomic insights, is limited by its complex dimensionality posing obstacles to developing universally applicable models. This study introduces an approach to leverage census data through the creation of location embeddings across various domains. Utilizing the Optuna framework, this research tuned autoencoders to optimize the bottleneck layer size, producing compact, low-dimensional embeddings that encapsulate critical relationships. These embeddings are further enriched with Federal Information Processing Standard (FIPS) codes, maintaining geographic identifiers. The methodology’s effectiveness is demonstrated through regression models trained on the American Community Survey, accurately predicting key indicators like median gross rent and per capita income with an 8–10% higher average accuracy compared to traditional PCA-based methods. These findings suggest a novel paradigm for overcoming the limitations of geographically specific huge dimensional data for synthesizing insights. In practical terms, such as in public policy, this approach could enable more precise targeting of socio-economic interventions based on nuanced community profiles. This innovative technique in representation learning shows considerable promise for enhancing machine learning applications across diverse sectors, including marketing, and real estate.