1 month ago
The Rise Of Large Database Models

Originally published in Forbes

Even as large language models have been making a splash with ChatGPT and its competitors, another incoming AI wave has been quietly emerging: large database models.

LDMs complement LLMs by capitalizing on the world’s other main data source: enterprise databases. Rather than tapping the great wealth of human writing such as books, documents and the web itself—as LLMs do—LDMs tap a company’s tabular data.

LDMs don’t power chatbots—instead of training on human language, they train on data records and transaction logs. So what capabilities do LDMs offer? Here’s an overview, as well as a specific example where LDMs power a predictive AI project at Swiss Mobiliar, Switzerland’s oldest private insurance company.

Uncovering Meaning Within Databases

Innovators at IBM’s illustrious Thomas J. Watson Research Center (where yours truly worked as a Research Affiliate during the summer of 1993) have led the charge in developing LDMs.

Like the endless reams of written word consumed by LLMs, enterprise databases represent a mammoth storehouse, a wellspring of facts and occurrences: every recorded purchase, transaction, click, credit application, customer profile and business record. While LLMs ascertain a certain amount about the meaning of words, LDMs discover the meaning of database values, such as the factors that define a customer record, including the customer’s location, buying history and demonstrated interests.

This empowers LDMs to offer a new kind of capability: database searches based on meaning, aka semantic queries. Traditionally, database queries must be issued in terms of explicit, unambiguous constraints expressed as specific value ranges—for example, “List all the customers who live in California, are more than 40 years old and have spent at least $2,000.” But with an LDM in place, you can ask a database to, “List all the customers most similar to Jane Doe,” or, “List all the cities most similar to Detroit in customer behavior.”

The use cases for semantic querying abound. What kind of food is nutritionally similar to toffee-covered almonds (answer: oatmeal)? Which products is this sort of customer also likely to buy? Which transactions deviate from the norm and therefore may be suspicious? Which variations on “TJ Watson Research” refer to the same thing—including “T.J. Watson” and “Thomas J. Watson Research Center,” but not “James Watson” (the co-discoverer of the double helix), “John Watson” (Sherlock’s sidekick) or “IBM’s Watson DeepQA” (the computer that bested the human Jeopardy! champions)?

IBM has launched LDMs from the research lab as a product called Db2 SQL Data Insights. The product is available as part of the Db2 database on the company’s z/OS operating system, which drives many real-time deployments of machine learning.

Let’s turn to a proven case study.

Insurance Sales: A Use Case For Large Database Models

Swiss Mobiliar handles sales with a personal touch. As with European insurance sales in general, the process leans more heavily on sales personnel than is typical in the U.S., where sales are often completed online. This leaves critical sales tactics in human hands. Before a salesperson issues an insurance quote, they must hand-craft it to increase the odds that it will be accepted and a contract will be signed.

Enter predictive AI, which tells you those odds. Given a prospective policyholder and a drafted quote, what are the chances the customer will sign? The answer to this question guides sales staff to adjust each quote. If they don’t like the odds, sales personnel can modify the quote with more aggressive coverage options or pricing, even issuing a special discount in certain cases, and then run the new quote through the predictive AI system for a new calculation of the odds. This empowers staff to navigate the balance between pricing and potential success rate by way of trial and error.

Typically, such predictive AI projects demand heavy involvement by experienced machine learning experts and a lengthy project lifecycle to define the requirements, prepare the data, train a model, evaluate it and integrate it for deployment.

But Swiss Mobiliar had other plans. The company wanted to find a quicker route to enterprise value.

A Turnkey Approach To Advanced Analytics

An avid user of IBM’s Db2 database solutions, Swiss Mobiliar Data Evangelist Thomas Baumann decided that his team would give SQL Data Insights a try for this project. “Our goal was to create quotes that perfectly fit our customers—not to complete a large-scale investigation into machine learning methods,” he told me. “I loved the possibility that we could get there by using my database staff and yet without also engaging senior data scientists.”

Baumann realized that the ability to find “similar” records, which SQL Data Insights builds in as a new SQL capability, gets you 95% of the way there for a predictive AI project like this one (as well as for projects that use clustering, aka unsupervised machine learning, which Baumann’s team pursues for some of his other projects).

Here’s how it works: Given a database record that defines the current situation—for this project, a prospective insurance policyholder and a candidate quote—simply pull out the most similar previous cases and count up how often those cases led to sales success. Voila, you have your odds.

Data scientists know this approach as k-nearest neighbors, a long-standing, classical machine learning method. “Nearest neighbor” refers to finding the previous cases that are the closest—that is, the most similar. Unlike most other ML methods, this one doesn’t require the model to be trained. Instead, a curated dataset of historical cases stands at the ready, with the most similar cases pulled out each time a prediction needs to be made for a new case.

The KNN method traditionally demands specialized ways to gauge “similarity” or “nearness” between database records. This typically must be hand-crafted by experts, taking into account the meanings of various values in each record and the significance that each one holds. Does the customer’s age matter more than their hometown or region? How about their financial history?

LDMs offer a saving grace. They’re turnkey for establishing a similarity metric, eliminating the need for experts to devise a customized definition for “nearest neighbors.”

Launching The Predictive Sales Tool

Baumann’s team implemented this approach across 15 million records of automobile insurance quote data, involving a couple dozen attributes for each one, including demographics, vehicle data, deductibles and price. The team discovered, after a bit of trial and error, that 43 was their sweet spot: Predictive performance peaked by pulling the 43 historical cases most “similar” to the current one and using that microcosmic track record to calculate the odds.

Next came deployment. Baumann and his team added the predictions to their sales team interface, which now displays the likelihood of closing each candidate quote.  A few hundred salespeople worked actively with this capability, using it to guide their formation of quotes by visiting the odds over multiple candidate quotes for each prospective policyholder before deciding on one.

Creating better quotes resulted in a remarkable sales boost: Over six months, the closings increased by seven percent, an increase that would have normally taken two years. As you might imagine, Baumann is actively pursuing other initiatives that also leverage SQL Data Insights.

In November 2024, Baumann presented on this project’s success at Machine Learning Week (see a PDF of his slides here), the conference series I founded, in a session called “Insurance Quote Recommendations at Swiss Mobiliar Powered By In-Database ML.” He’s set to present again at MLW 2025 and his writing does well to illustrate the value proposition by drawing intuitive analogies.

There’s a new kid on the block. LDMs introduce a new range of capabilities that complement those offered by LLMs. Just as LLMs train on language to deliver capabilities accessible to non-technical humans, LDMs train on enterprise databases to deliver capabilities accessible to database users who aren’t data scientists.

—

Financial disclosure: IBM has previously hired me for 18 speeches and two writing assignments, although never concerning the topic of this article.

About the author
Eric Siegel is a leading consultant and former Columbia University professor who helps companies deploy machine learning. He is the founder of the long-running Machine Learning Week conference series, the instructor of the acclaimed online course “Machine Learning Leadership and Practice – End-to-End Mastery,” executive editor of The Machine Learning Times and a frequent keynote speaker. He wrote the bestselling Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die, which has been used in courses at hundreds of universities, as well as The AI Playbook: Mastering the Rare Art of Machine Learning Deployment. Eric’s interdisciplinary work bridges the stubborn technology/business gap. At Columbia, he won the Distinguished Faculty award when teaching the graduate computer science courses in ML and AI. Later, he served as a business school professor at UVA Darden. Eric also publishes op-eds on analytics and social justice. You can follow him on LinkedIn.