Kumo’s ‘relational foundation model’ predicts the future your LLM can’t see

Publisher’s note: Kumo Ai was one of the finalists VB transform During our Annual innovation showcase and presented RFM from the main scene at VB transform Wednesday.

THE AI Generative Boom has given us powerful language models that can write, summarize and reason on large amounts of text and other types of data. But with regard to high -value predictive tasks such as prediction of customer unsubscribe or fraud detection from structured relational data, companies remain stuck in the world of traditional automatic learning.

Professor of Stanford and Soft The co-founder swears Leskovec maintains that it is the missing critical play. The tool of his company, a relational basic model (RFM), is a new type of pre-formulated AI which provides the “zero-shot” capacities of large language models (LLM) to structured databases.

“It is a question of making forecasts on something you do not know, something that has not yet happened,” Lesko de Venturebeat told. “And it is a fundamentally new capacity that is, I would say, missing from the current purpose of what we consider to be geni.”

Why the predictive ML is a “30 -year technology”

While the LLM and Recovery generation (RAG) Systems can answer questions about existing knowledge, they are fundamentally retrospective. They recover and reason on the information that is already there. For predictive commercial tasks, companies are still counting on conventional automatic learning.

For example, to build a model that predicts customers’ unsubscribe, a company must hire a team of data scientists that spend a lot of time making “functional engineering”, the process of manual creation of predictive signals from data. This implies a dispute of complex data to join information from different tables, such as the purchase history of a customer and website clicks, to create a single massive training table.

“If you want to do automatic learning (ML), sorry, you are stuck in the past,” said Leskvec. The expensive and long strangling bottlenecks prevent most organizations from being really agile with their data.

How Kumo generalizes the processors of the databases

Kumo’s approach, “Deep Learning”, goes around this manual process with two key information. First of all, it automatically represents any relational database as a single interconnected graph. For example, if the database has a “user” table to save customer information and a “command” table to save customer purchases, each line of the user table becomes a user node, each line of a control table becomes a control node, etc. These nodes are then automatically connected using existing relationships in the database, such as foreign keys, creating a rich card of the data set without manual effort.

Second, Kumo generalized Transformer architectureThe engine behind LLMS, to learn directly from this representation of graphics. Transformers excel in understanding the tokens sequences using a “attention mechanism” to weigh the importance of different tokens compared to each other.

Kumo RFM Apply this same mechanism of attention to the graph, allowing it to simultaneously learn models and complex relationships on several tables. Leskvec compares this jump to the evolution of computer vision. In the early 2000s, ML engineers had to manually design features such as edges and shapes to detect an object. But new architectures such as convolutional neurons (CNN) can take raw pixels and automatically learn the relevant features.

Likewise, the RFM ingests gross database tables and allows the network to discover the most predictive signals in itself without the need for manual effort.

The result is a pre-formulated foundation model that can instantly perform predictive tasks on a new database, which is called “zero-shot”. During a demo, Leskvec showed how a user could type a simple question to predict if a specific customer would place an order in the next 30 days. In a few seconds, the system returned a probability score and an explanation of the data points that led to its conclusion, such as the user’s recent activity or its absence. The model has not been formed on the basis of data provided and adapted in real time through learning in context.

“We have a pre-formulated model that you simply indicate your data, and this will give you a precise prediction 200 milliseconds later,” said Leskvec. He added that it can be “as precise as, say, weeks of the work of a data scientist”.

The interface is designed to be familiar to data analysts, not only to automatic learning specialists, democratizing access to predictive analysis.

Feed the agency future

This technology has important implications for the development of AI agents. For an agent to carry out significant tasks within a company, she must make more than simply treat the language; He must make intelligent decisions based on the company’s private data. RFM can serve as a predictive engine for these agents. For example, a customer service agent could question the RFM to determine the probability of a customer’s barattage or its potential future value, then use an LLM to adapt its conversation and therefore offers.

“If we believe in an agental future, agents will have to make rooted decisions in private data. And it’s the way for an agent to make decisions, “said Leskvec.

Kumo’s work points to a future where AI company is divided into two complementary areas: LLMS to manage retrospective knowledge in an unstructured text and RFM for predictive forecasts on structured data. By eliminating the bottleneck of the functionality engineering, the RFM promises to put powerful ML tools in the hands of more companies, which considerably reduces the time and the cost to pass data to the decision.

The company has published a public demo of the RFM and plans to launch a version which allows users to connect their own data in the coming weeks. For organizations that require maximum precision, Kumo will also offer a fine adjustment service to increase the performance of private data sets.