How do I use a feature store with Kumo?
In traditional machine learning systems, a feature store serves as a centralized hub for storing, processing, and accessing commonly used features. Feature engineering is by nature error-prone and time-intensive, requiring considerable time and effort to understand the problem space and the relevant data. Feature stores are therefore crucial for traditional MLOps workflows and ML pipelines, since they function as a central repository for storing and reusing commonly used features. Using a feature store allows for feature reuse, ensures consistency between serving and training pipelines, and helps monitor data quality.
Kumo’s state-of-the-art GNN architecture removes the need for computing feature stores and feature engineering pipelines. By leveraging the relational structure of the entities in the data to build a single enterprise graph, Kumo is able to achieve a comprehensive view of the dynamic interactions and relationships between the different entities in the raw data—without extensive feature engineering or the use of feature stores.
In short, with Kumo, you don’t need to use a feature store. However, if you want to use the features in your pre-existing feature store, it is fairly easy to do so. You can use your existing feature store to generate a training table consisting of the pre-materialized features from your feature store along with the label you are trying to predict. Then, you can register this table to your Kumo graph, and write predictive queries as you normally would. The resulting GNN model will use the features from your “imported” training table, as well signals from the broader connected graph, leading to better performance.
Learn More:
Updated 4 months ago