Interpretability, then what? Edit machine learning models to reflect human knowledge and values
Recently, a lot of research in the field of machine learning has been done to create models can be explained. However, the interpretation that can be used to improve these models is still lacking in focus.
A recent arXiv.org paper by an interdisciplinary team presents the first interactive system that allows domain experts and data scientists to edit the weights of generalized additive models (a model with modern interpretation).
The researchers develop an easy-to-use and flexible user interface that supports multiple editing methods. That allows domain experts with less experience in machine learning to investigate and improve the model. Users are provided with ongoing feedback on the impact on different subgroups and correlations of the feature to avoid harmful edits. Transparent and reversible model modifications are also supported.
The proposed method has been successfully applied to fit the pneumonia and sepsis risk prediction model with the clinician’s clinical knowledge.
Machine learning (ML) interpretation techniques can reveal unexpected patterns in the data that models exploit to make predictions – potentially harmful once deployed. However, how to act to address these problems is not always clear. In collaboration between ML and human-computer interaction researchers, doctors and data scientists, we develop GAM Changer, the first interactive system that helps domain experts and scientists learn data easily and responsibly edit General Additive Models (GAM) and fix problematic patterns. With new interaction techniques, our tool puts interpretability into action – empowering users to analyze, validate, and align model behaviors with the knowledge and value of the model. surname. Physicians have begun using our tool to investigate and correct predictive models of pneumonia and sepsis risk, and a review with 7 data scientists working in the fields Different areas show that our tool is easy to use, meets their model editing needs and fits into their current workflow. Built using modern web technologies, our engine runs locally in the user’s web browser or compute notebook, reducing barriers to use. GAM Changer is available at the following public demo link: This https URL.
Research articles: Wang, ZJ, “Interpretability, Then What? Editing machine learning models to reflect human knowledge and values”, 2022. Link: https://arxiv.org/abs/2206.15465
