Hema Raghavan, co-founder and head of engineering at Kumo AI, previously led AI initiatives at LinkedIn, including the iconic 'People You May Know' feature. In this conversation, she unveils how graph neural networks (GNNs) revolutionize automated machine learning by transitioning from CPU to GPU. Hema explains how GNNs enhance recommendation systems, simplifying feature engineering and adapting to user preferences. She also discusses the transformation of relational data into graph formats and the critical role of explainable AI in advancing technology.
Kumo's integration of graph neural networks with existing data warehouses allows enterprises to automate machine learning efficiently, enhancing predictive capabilities without extensive manual input.
The innovative query language developed by Kumo simplifies the process of defining machine learning problems, making advanced AI accessible to non-experts while empowering data scientists.
Deep dives
Innovative Approach to AutoML with Graph Neural Networks
Kumo AI focuses on automating machine learning (AutoML) specifically on GPUs, setting it apart from traditional AutoML solutions that rely on CPU-based models. This shift allows for building sophisticated AI models more efficiently, leveraging advanced graph neural networks (GNNs) to learn features automatically, without extensive manual feature engineering. As a result, enterprises can utilize their existing data warehouses—such as Snowflake and Databricks—to deploy predictive models rapidly and effectively. This innovation reduces the dependency on data science expertise while maximizing business value.
Graph Neural Networks and Their Versatility
Graph neural networks excel in modeling complex relationships, making them suitable for various domains beyond traditional social networks. For instance, in a YouTube scenario, GNNs can analyze user viewing habits along with content characteristics to recommend videos more accurately than conventional collaborative filtering methods. The ability to learn from a bidirectional graph involving users and content allows for deeper insights and connections between entities, enhancing predictive capabilities across different industries, including fintech and healthcare. Kumo aims to democratize the use of GNNs, allowing non-experts to harness their power for diverse machine learning problems.
Bridging the Gap Between Relational and Graph Data
Kumo innovates by simplifying the transformation of relational data into graph structures, enabling organizations to utilize existing tables effectively. This process eliminates the complexity typically associated with graph learning, which has historically required specialized knowledge. By employing predictive query language, Kumo provides an accessible interface that allows data scientists to define machine learning problems with ease, resembling SQL queries. As a result, companies can quickly leverage their relational data and gain valuable predictive insights without extensive training or restructuring efforts.
Kumo's Impact on AI Adoption and Development
Kumo fosters quicker adoption of AI by enabling organizations to achieve measurable results in a matter of weeks, alleviating the pressures faced by data teams under performance expectations. The platform empowers data scientists by automating feature engineering and simplifying model deployment, allowing them to focus on direct business impact rather than maintenance tasks. This shift encourages teams to experiment rapidly and innovate without the usual overhead associated with traditional AI workflows, making it an attractive solution for companies eager to stay competitive. Kumo's approach positions it as a key player in transforming the AI landscape, especially for enterprises transitioning to data-driven decision-making.
Hema Raghavan is co-founder of Kumo, a company that makes graph neural networks accessible to enterprises by connecting to their relational data stored in Snowflake and Databricks. Hema talks about how running GNNs on GPUs has led to breakthroughs in performance as well as the query language Kumo developed to help companies predict future data points. Although approachable for non-technical users, the product provides full control for data scientists who use Kumo to automate time-consuming feature engineering pipelines.
Mentioned in this episode:
Graph Neural Networks: Learning mechanism for data in graph format, the basis of the Kumo product
Graph RAG: Popular extension of retrieval-augmented generation using GNNs