Surprises in Proper Positive-Only Learning

Imagine you're trying to teach a model to recognize good products, but you only have examples of products that are good — no bad ones to compare against. This is tricky because, in real life, you need to know not just what’s good, but also what’s not. The traditional approach to learning assumes you have both good and bad examples, but when you only have positives, it can lead to confusion about what the model should learn. This is known as the challenge of positive-only learning. The problem is that without negative examples, the model might overfit to the positives and fail to generalize well to unseen data, which is a failure mode called improper learning. This paper addresses that gap by establishing a new condition that helps determine when a model can learn properly from just positive examples. They introduce a concept called uniform exterior separability, which, along with finite VC dimension, defines the boundaries of proper learning in this context. This means that now, when building models that rely on positive-only data, you have clearer guidelines on what can be learned effectively and what cannot. This is a significant step forward in understanding the landscape of learning theory, especially for applications where negative samples are hard to come by.

The paper introduces a new combinatorial condition for proper positive-only learning, expanding the understanding of learning theory.

The claims are supported by theoretical results and a clear characterization of learning conditions.