Telecom giant AT&T and Mountain View, California-headquartered AI cloud company H2O have jointly launched an artificial intelligence feature store for enterprises.
The repository, available as a paid software platform, enables data scientists, developers, and engineers to discover, share, and reuse machine learning features to speed up their AI project deployments. According to the companies, it can be used for work on personalization and recommendation engines as well as models that are aimed at forecasting, as well as the optimization of dynamic pricing, supply chain, and logistics and transportation.
The development comes as more and more organizations turn to AI implementation to generate actionable insights and predictions from a huge trove of data. AT&T originally had this solution in production use for network optimization, fraud prevention, tax calculations, predictive maintenance, among other things.
Features in machine learning
When it comes to building machine learning models, data is of the utmost importance. However, raw data is not the key to a well-performing algorithm. The information gathered first must be cleaned and enriched with features — individual independent variables or characteristics that act as the input for the system. The quality of these features defines the quality of the model as well as the accuracy of its predictions.
Typically, AI experts apply domain knowledge of the data and engineering tools to extract and create features. The entire process takes months and has to be repeated for every new AI project (adding to the cost), even if it is under the same organization.
A feature store strives to solve this challenge by serving as the home for commonly used features. With this solution, experts could create new features for a project and then add them into the store, ensuring they could be reused if required at a later stage. Databricks, Tecton, Molecula, Hopsworks, Splice Machine, and Amazon Web Services (AWS) are the leading players that offer feature stores to accelerate MLOps.
The latest offering from H2O and AT&T not only performs the core function of a feature store, but also comes equipped with multiple additional capabilities. The store offers integration with multiple data and machine learning pipelines, which can be applied to an on-premise data lake or by leveraging cloud and SaaS providers. It is integrated with Snowflake, Databricks, Apache Spark, H2O Sparkling Water, Python, Scala, and Java.
Then, the solution also offers an automatic recommendation engine that learns over time and recommends new features and feature updates to improve the AI model performance of the user. This way, data scientists can simply review the suggested updates and accept the recommendations best suited for their model. FeatureRank scores all the features on the store based on their popularity and value.
“We are building AI right into the feature store and have taken an open, modular, and scalable approach to tightly integrate into the diverse feature engineering pipelines while preserving sub-millisecond latencies needed to react to fast-changing business conditions,” Sri Ambati, CEO and founder of H2O.ai, said in a statement.
The company says any firm engaged in AI development can adopt its store, starting from financial services and health organizations to pharmaceutical makers, retail, and software developers.
“Feature stores are one of the hottest areas of AI development right now because being able to reuse and repurpose data engineering tools is critical as those tools become increasingly complex and expensive to build,” Andy Markus, chief data officer at AT&T, said.
“With our expertise in managing and analyzing huge data flows, combined with H2O.ai’s deep AI expertise, we understand what business customers are looking for in this space and our feature store offering meets this need,” Markus added.
According to a PwC study, AI will $15.7 trillion to the global economy by 2030. Leading this growth are China and North America, which will drive the greatest economic gains at $10.7 trillion.