Feature Discovery for Enterprise ML: dotData CEO at COLLIDE 2023
At COLLIDE 2023 in Atlanta, dotData CEO Ryohei Fujimaki argued that generating features automatically from raw business tables is faster and more reproducible than hand-built features. He also described three ways generative AI is being applied to unstructured text, conversational feature creation and feature interpretation.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Features drive both accuracy and trust. More table variety gives models more signal, and transparent features give business sponsors something to act on.
- Manual feature work is slow and error-prone. Fujimaki cited weeks-to-months turnaround, data leakage, poor reproducibility and limited scale.
- Let the data propose features. The framework generates a large feature space from raw tables and a target, then selects features and builds a pipeline.
- Case study claims. A lender's delinquency project went from about 150 hours with three data scientists to about 5 hours with one, he said.
- Generative AI helps at the edges. Uses included text featurization, conversational feature creation and explaining features in business terms.
At COLLIDE 2023 in Atlanta, Ryohei Fujimaki, CEO of dotData, presented a "data-centric" approach to feature engineering for enterprise machine learning. In a 33-minute session, he argued that generating candidate features automatically from raw business tables is faster, more reproducible and easier to move into production than hand-built features, and he described three ways his company was applying generative AI to the process.
Fujimaki said the point of the talk was to "let the data speak for itself." His remarks reflect the view of a vendor that sells this approach.
What is feature engineering in enterprise machine learning?
Fujimaki described feature engineering as the step that turns a business problem into something a model can use. A team starts with a goal such as predicting customer churn, loan risk or demand. The data sits in master repositories built for business operations, not analytics, so it has to be transformed into a single flat, aggregated feature table before modeling.
He used credit card churn as an example. A customer demographics table joins easily to the churn history. A payment history table has many rows per customer, so analysts must choose time windows and aggregation functions. Add online banking logs and campaign responses and the work becomes "messy" and "chaotic," he said, and real enterprises have far more than four tables.
Why does feature engineering matter for accuracy and trust?
Fujimaki gave two reasons. First, more variety in data, meaning more tables and not only more rows, gives a model more chances to find signal. Second, features carry insight. A feature such as a "credit card balance decrease 75% during the last three months" linked to a higher churn rate gives a business sponsor something to act on, he said, and transparent features build trust and ease deployment.
What goes wrong with the manual approach?
Fujimaki listed four problems he said he sees across data science teams:
- Slow turnaround. Domain experts, data engineers and data scientists pass ideas back and forth, and the cycle typically takes weeks to months.
- Errors. He cited a conversation with the head of data science at a U.S. hedge fund who described junior data scientists reporting strong signals that turned out to be data leakage, meaning future information unavailable at prediction time.
- Poor reproducibility. Experiments are spread across many notebook versions.
- Limited scale. The work depends on individuals, and countless feature ideas exist for any problem with 10 or 20 tables.
How does data-centric feature discovery work?
In the dotData framework, a team supplies raw source tables, a target to predict and the relationships between tables, such as a shared customer ID. Algorithms then generate a large feature space, from which features are selected and a production-ready feature pipeline is produced.
Fujimaki shared a project with a lender's loan delinquency model. The team's earlier approach drew on 2,622 manually built domain features, from which 42 were selected for the model, a process he said took about 150 hours with three data scientists. With the new approach, the team started from 15 source tables, generated a very large set of candidate features in a couple of hours, and had an algorithm select 40 features. He said it took about 5 hours with one data scientist and delivered a Gini index slightly higher than before.
He said time to value was 20 to 30 times faster, and that the model supports a million-dollar program. Captions garbled some of the figures, so the exact accuracy improvement is not reported here.
He also argued that the approach produces reusable assets. "Hundreds or thousands of Jupyter notebooks are created which is only one time used," he said, whereas a systematic process lets one person's discovery work carry to the next problem on the same data.
How is generative AI being used in feature discovery?
Fujimaki described three efforts, some still in progress:
- Unstructured data featurization. For free-form customer support text, calling a chat API on every record is slow and costly at millions of records. dotData combines a fast, cheap embedding API with a context label estimator trained on a subsample, then feeds the labels back as features. One resulting feature, for example, flagged premium subscribers expressing urgency, which he said carried a churn rate 30 times higher.
- Conversational feature creation. A domain expert describes a feature in plain language, such as average remaining time on customer loans. The system asks which tables and columns apply, writes the query, and remembers the answers for similar requests later. He called it an early-stage attempt.
- Feature interpretation. Generative AI proposes the business hypothesis behind a discovered feature, for example why low average balances correlate with default, to make findings easier to explain to sponsors.
Where to follow enterprise machine learning discussions
COLLIDE 2027, billed as the largest AI conference in the Southern U.S. with more than 1,000 enterprise data and AI leaders, takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta. Details are on the COLLIDE event page, and registration is coming soon.
Frequently asked questions
What is feature engineering in machine learning?
It is the data transformation step that turns raw business tables into a single flat, aggregated feature table a model can learn from, tailored to a specific business problem.
What is data leakage?
Fujimaki described it as using future information that would not be available when a prediction is actually made, which can make a model look far better than it is.
How can generative AI help with feature engineering?
He described using embeddings and LLM-derived labels to turn free-form text into features, letting experts request features conversationally, and generating business explanations for discovered features.
Why is feature engineering hard to scale?
According to Fujimaki, it depends on back-and-forth between domain experts, engineers and data scientists, and the number of possible features grows quickly with the number of tables.
This article is based on remarks by Ryohei Fujimaki, CEO, dotData at COLLIDE Data + AI Conference 2023 on October 3, 2023. More coverage: COLLIDE AI Conference.
Filed under: AI Use Cases & Applications, Data & AI Infrastructure
