Predictive Multiplicity

Predictive multiplicity describes the phenomenon where multiple machine learning models, achieving near-identical performance on a given task, produce conflicting predictions for individual instances. Current research focuses on quantifying this multiplicity across various model types, including knowledge graph embeddings, linear models, and rule lists, and developing methods to mitigate its impact, such as ensemble techniques and improved model selection strategies. Understanding and addressing predictive multiplicity is crucial for building trustworthy AI systems, particularly in high-stakes applications like content moderation, loan applications, and healthcare, where inconsistent predictions can lead to unfair or arbitrary outcomes. The field is actively developing metrics to measure and visualize this phenomenon, aiming to improve model transparency and accountability.