Automatic Detection of Learnability under Unreliable and Sparse User Feedback

Personalization for real-world machine-learning applications usually has to incorporate user feedback. Unfortunately, user feedback often suffers from sparsity and possible inconsistencies. Here we present an algorithm that exploits feedback for learning only when it is consistent. The user provides feedback on a small subset of the data. Based on the data representation alone, our algorithm employs a statistical criterion to trigger learning when user feedback is significantly different from random. We evaluate our algorithm in a challenging audio classification task with relevance to hearing aid applications. By restricting learning to an informative subset, our algorithm substantially improves the performance of a recently introduced classification algorithm.