When Airline Insurance Meets a Pandemic: Designing Machine Learning Resilience Under Volatile Drift and Data Deficiency

Abstract

Disaster environments create volatile data drifts and severe data deficiency that undermine machine learning (ML) resilience. We propose the Context-Enriched Drift-Aware Contextual Bandit (CEDA), a computational design artifact that addresses these challenges through two novel components. First, a Bayesian drift-point detector within a contextual bandit framework captures diverse drift patterns via Laplace approximation, supporting generalized reward functions and a broad range of exploration-exploitation policies. Second, a lightweight AutoEncoder-based context-enrichment embedding fuses heterogeneous spatial (e.g., socioeconomic) and temporal (e.g., mobility) disaster signals into a rich yet compact representation, enabling robust online updates and enhanced bandit performance. Evaluated on a proprietary airline-insurance dataset spanning pre- and post-COVID-19, CEDA significantly outperforms state-of-the-art drift-adaptive ML models. Ablation studies confirm that both drift detection and context enrichment contribute to performance and, when combined, produce synergistic gains. An economic benefit analysis further demonstrates how gains in conversion rate could translate into potential profit increases. When evaluated for generalizability in a non-disaster context using the public Yahoo! User Click Log dataset, CEDA outperforms leading bandit algorithms by achieving a significant lift in click-through rate. Overall, CEDA demonstrates robust performance across various settings and datasets, in both crisis and stable conditions. These findings offer important computational design implications for building ML resilience in sociotechnical contexts marked by volatile drift and data deficiency and offering practical implications for businesses seeking resilience strategies in ML and AI development.

Keywords: computational design, socio-technical environments, machine learning resilience, COVID-19 pandemic, non-stationary contextual bandit, Bayesian drift point detector, contextual enrichment