How AI Accelerates

ML Development
Kumo’s approach to intelligent data science
Overview

The last two decades have created an explosion of diverse and complex data
that contain a wealth of insights waiting to be uncovered.

While analytics tools accessed only lagging, surface-level insights, machine learning unlocked deeper value with predictions.
Today’s methods, however, offer little more than incremental business benefits for months of effort. The natural limitations of
traditional data science have been reached, and a new approach is needed to deliver the next echelon of value.

Kumo uses artificial intelligence (AI) to accelerate the creation and performance of machine learning (ML) models. By
streamlining the machine learning lifecycle and eliminating the need for feature engineering, Kumo produces model
improvements up to 75% and growing. This approach uncovers more value from data and elevates the role of the data scientist,
unleashing their creativity and modeling expertise.

Traditional data science Kumo’s intelligent data science

Uses a representative subset of relational data Uses a graph data structure for efficient model
with complex joins to build and train models. construction and training.

Build custom models using complex Python Speeds up model development in an AI-powered,
libraries with tedious feature testing and low-code environment, bypassing complex feature
selection. engineering.

Takes months to clean and prep data, test features Empowers experts to enhance models and adjust hy
and bespoke models based on hypotheses. perparameters using their domain expertise, simplifying
the modeling process.

Kumo is the leader in intelligent data science because it uses the power of graph structure to provide the highest performing
models and most accurate predictions — down to the entity level. This is done by automatically converting relational data into
graphs and using graph neural networks (GNNs), a class of AI models designed for graphs, to learn from relationships in the
data. Kumo rapidly builds highly accurate models that keep pace with business questions with low-code, SQL-like queries that
predict things like segments, lifetime value, behaviors, and more to improve personalization and recommendations.

Deliver Accuracy with AI

AI is the centerpiece of intelligent data science solutions like Kumo. Kumo uses GNNs because of their exceptional accuracy in
making predictions across incredibly high volumes of relational data. This structured view of enterprise data allows for efficient
access and analysis of large datasets, enabling quick identification of the right data for model training.

Kumo automatically transforms relational databases into graph structure based on primary-foreign key relationships in the data.
The graph structure enables a deep understanding of interconnected relationships, specially designed to train GNNs. GNNs
operate by taking graph data as input, converting it into intermediate embeddings, and feeding them into a final layer designed
for specific prediction tasks.

01
This modeling approach outperforms many traditional approaches on tabular enterprise data, such as logistic regression,
gradient boosted decision trees, and Bayesian networks. Therefore, GNNs have widespread application in recommendation
systems, fraud detection, and forecasting, among others.

From relational data to GNN output

GRAPH layer GNN layer

Relational data GRAPH input Output layer Output Write results to data
warehouse
Activation Activation

Learn more about GNNs

Eliminate Feature Engineering

Kumo saves time during model development because it eliminates the complex process of feature engineering through
advanced deep learning methods. Keeping pace with business needs remains a significant challenge in traditional data science.
Often, by the time a model is developed or enhanced, the business context has changed, rendering the work obsolete.

Kumo's GNNs use a graph sampling and aggregation algorithm, to automatically learn complex features which would
traditionally need to be hand-crafted by a data scientist. Autonomous extraction of features from raw data, significantly
enhances performance and reduces the time to production down to hours

TRADITIONAL DATA SCIENCE KUMO

Business
Business
questions
question

Business Data
answers collection

Business Clean
Data Data cleaning answers data
visualization & preparation
6 - 9 months hours

Model Exploratory
evaluation data analysis

Fine tune AI
Machine Feature
learning engineering

90% OF TIME
SPENT HERE 90% OF TIME
SPENT HERE

02
Elevate Data and Domain Experts

Due to the time savings unlocked by Kumo’s GNN, data science experts can focus on higher-leverage work to maximize model
performance and drive business value. Kumo provides control over autoML search space, enabling data scientists to quickly
explore many GNN architectures. Data science teams can define custom model plans, leveraging their domain expertise to
fine-tune hyperparameters and refine models.

With Kumo, data scientists can take on more projects and answer questions faster to drive a measurable and sustainable
business impact.

Activate the Data Warehouse

The data warehouse contains all the carefully curated business data from potentially dozens of internal and external systems.
Unlike traditional ML approaches that only operate on single tables, GNNs can leverage the entirety of the data to produce the
best models and predictions. Kumo integrates directly with the tables in the client’s data warehouse, eliminating the need to
maintain complex data loading and transformation workflows.

Reduce latency

Deploy Kumo inside your

data warehouse Reduce cost
[mass data loads are expensive]

Customer data
stays in-boundary

Kumo Architecture Overview:

Interactions
Demographics

Models for
Business
Question AI fine-tuning
& predictions

Behavioral
Transactions

Data warehouse and more...

03
Get Better Results

Kumo makes its out-of-the-box predictions accurate by checking data quality on relevant ingested columns and tables,
supporting distributed model training, and providing MLOps tooling for detecting data drift. Further, Kumo’s streamlined retraining
pipelines ensure that the model stays responsive to changes in product, industry trends, or user behavior to help ensure the
highest accuracy possible.

Kumo can generate predictions multiple times a day from the most recently retrained models, integrating predictions directly into
platforms like Snowflake, Databricks, Amazon Redshift, or Google BigQuery for downstream use. Moreover, it facilitates
automated exports of batch predictions at predetermined intervals, providing a consistent flow of data for informed business
decisions and the best results in production.

Kumo accelerates time to value while delivering the highest performing models for exceptionally accurate predictions with model
performance improvements of up to 75% compared to client models that have been hand-tuned for years.

Kumo's Success Across Industries

On-demand food delivery service

The business goal: The approach:

Generate new sales and improve engagement Predict all restaurants each user is likely to
with targeted promotions. order from in the next 7 days.

The impact of Kumo:

Increased MAU Added $100 Cut serving

by 90,000 million in sales costs by 80%

Read the full story

A major grocery chain

The business goal: The approach:

Increase revenue from existing customers For each shopper, predict twelve new and
through personalized coupons. repeat products that drive conversion.

The impact of Kumo:

Increased new purchase Increased repeat purchase Improved models in

model by 24% model by 7% 5 days

Read the full story

Online bank

The business goal: The approach:

Increase revenue by steering customers toward Predict user intent across the top ten
the most likely opportunity at the right time. revenue generating opportunities.

The impact of Kumo:

Improved model Improved initial Achieved production

by 11x results by 73% ready model in 4 days

Read the full story

04
Top Use Cases for
Intelligent Data Science

Organizations onboard intelligent data science solutions like Kumo for revenue impacting KPIs like engagement,
sales, and conversion.

Recommendations for a
segment or individual
like item recommendation sent
by email or notification, in-app Personalized experiences
recommendation for items in- in app or website
cluding for cross-sell
like landing pages tailored to
user preferences, unique ranking
of products, responsive discov-
ery experiences, individualized
call to action and promotions

Next best action

like user behavior and marketing
campaign decision making
Segmentation
like target audience, cross-chan-
nel marketing campaigns, life-
time value modeling, and ad
campaign optimization

Funnel Acceleration
like lead ranking, conversion,
and churn

Get Started with Kumo

Kumo’s intelligent data science solution provides the combination of speed, accuracy, and continuous updates that empower
organizations to respond promptly and effectively to business-critical questions, enhancing decision-making processes and
operational efficiency.

Learn more at kumo.ai and get started with a free trial.

05