Abstract
High-performing machine learning models often lack transparency and raise concerns about their trustworthiness, regardless of the data type they handle, whether tabular, textual, image, or audio data. This talk explores methods for investigating model behavior and improving trustworthiness from both subgroup and local perspectives. From a subgroup perspective, we focus on identifying and characterizing data subgroups where a model behaves differently. The identification of these data subgroups is relevant in many applications, such as model validation and testing, model comparison, error analysis, and identification of bias. We show how subgroup analysis can be a tool for model improvement, both in overall and subgroup performance, as well as drift detection. From a local perspective, we investigate how to explain individual predictions in speech models. The explanations analyze both word-level audio segments to understand the content and paralinguistic features to interpret how information is conveyed. Finally, the talk will cover the challenge of validating explanations, outlining a unified benchmarking suite to test and compare a wide range of state-of-the-art explainers for textual data.