Browsing by Author "Rahikainen, Tintti"

Machine learning operations (MLOps) tools and practices help us continuously develop and de- ploy machine learning models as part of larger software systems. Explainable machine learning can support MLOps, and vice versa. The results of machine learning models are dependent on the data and features the models use, so understanding the features is important when we want to explain the decisions of the model. In this thesis, we aim to understand how feature stores can be used to help understand the features used by machine learning models. We compared two existing open source feature stores, Feast and Hopsworks, from an explainability point of view to explore how they can be used for explainable machine learning. We were able to use both Feast and Hopsworks to aid us in understanding the features we extracted from two different datasets. The feature stores have significant differences, Hopsworks being a part of a larger MLOps platform, and having more extensive functionalities. Feature stores provide useful tools for discovering and understanding the features for machine learning models. Hopsworks can help us understand the whole lineage of the data – where it comes from and how it has been transformed – while Feast focuses on serving the features consistently to models and needs complementing services to be as useful from an explainability point of view.