Data Preprocessing for Culturally Competent Machine Learning in Social Robotics

In recent years, the growing integration of intelligent machines into daily life has raised significant concerns about their ethical and social implications. Among these, a critical challenge is the need for social robots to demonstrate cultural competence—namely, the ability to understand, respect, and adapt to diverse cultural contexts. This issue becomes especially prominent in machine learning (ML)-empowered systems, where the under-representation of certain cultural groups in training data can reinforce existing social disparities. Such disparities not only perpetuate stereotypes but also hinder social robots’ ability to interact effectively and equitably across diverse populations. This paper addresses the issue of cultural competence in social robotics, focusing on the impact of cultural under-representation in robotic visual perception. In particular, we considered the effect of under-representation of minority cultures in two image binary classification problems, and we explored Data Augmentation (DA) and Oversampling (OS) as preprocessing methods to address this issue. These two techniques result in improving cultural competence–measured using Cultural InCompetence (CIC) metric–especially when combined, demonstrating the validity of the proposals, regardless the use case considered.