IAC 2024 - IA Fast Track to Search Focused AI Solutions
Getting Your Supply Chain Back on Track with AI
1. Getting your Supply Chain
Back on Track with AI
Meetup: Jun 2, 2020
Karthik Guruswamy
H2O.ai
2. Confidential2
• H2O.ai – Our Response to COVID-19
• Supply Chain in the COVID-19 Era
• Tech. Debt with Shortage of Time, Talent and Trust
• Augmenting COVID-19 data in AI/ML models
• Short Term Predictions vs Long Term What-if Scenarios
• Daily/Hourly Predictions - Demand Sensing
• Driverless AI – Time Series Forecasting
• Q – Demand Sensing
• Q & A
Topics
3. Confidential3
How H2O.ai is Contributing to Solving COVID-19
Expertise
H2O.ai’s data science experts
are contributing their
knowledge to solve pressing
problems with the pandemic
AI Platforms
H2O.ai is contributing its
Driverless AI and Q platform
to model, predict, and
visualize data sets
Sri Ambati
CEO and Founder, H2O.ai
1. Hospital staffing predictions
2. ICU transfers and triage
3. Population risk segmentation
4. Predicting the spread of COVID-19.
5. Predicting operational efficiency and
resilience during a pandemic
6. Hospital supply chain predictions
7. Predicting responses by city,
hospitals
8. Sepsis predictions
Problems we are solving
“
Data Sets
H2O.ai is evaluating global
and open health data sets to
determine patterns
“Data Science can save
lives today. AI is an
incredible force to do
good for humanity.”
AI Solutions
H2O.ai is creating pandemic
and health specific solutions
for general use
4. Confidential4
• Unprecedented disruption to Supply Chain happening last few
months
– Quick recovery very difficult, unless acted on it proactively and with the right
tools and approach
• Disruption across Models and Processes around:
– Manufacturing, Warehouse Mgmt, Transportation, Distribution, Logistics,
Inventory, Procurement, Demand sensing & Cash Flow
• Through and Post COVID-19 disruption
– Diminished or sudden increase in demand for consumption of
household/industrial
– More Household Supplies, Groceries, PPE, Baking Goods, Alcohol, Frozen Food …
– Less Automotive, Shaving Products, Flowers, Makeup
– Emerging demand for new type of products and services
• What happens if there is a second wave ?
Supply Chain in COVID-19 Era
5. Confidential5
• AI/ML models that were manually crafted with Agile projects are
increasingly subject to stress
– Underlying assumptions/ecosystem/macro-economic environment has changed.
– New features have to be now created/back tested
– Never one size fits all – COVID cases change from state to state, country to
country!
– Where is the historical data ?
• Cannot afford to build new AI/ML models with the same methods and
timelines as situation changes every other week
• Becomes even more difficult to:
– Find time to train and deploy short-lived models with HIGHEST ACCURACY
– Talent to find and test new features as domain expertise is not fully reusable
– Trust the models and get insights each week on what’s impacting predictions!
Tech. Debt with Shortage of Time, Talent and Trust
6. Confidential6
• Kaggle GMs who work for H2O.ai have created power-growth models and made it
to the leaderboard here over multiple weeks of competition.
Incorporating COVID-19 data in AI/ML Models
7. Confidential7
• Power Growth models are great for forecasting and incorporates various
free parameters to model the infection rate.
• Epidemiological Models such as SEIRD can also be used on existing data
using different parameters
– SEIRD – Susceptible, Exposed, Infected, Recovered, Dead
• The next version of Driverless AI 1.9.0 will incorporate a SEIRD
transformer, where you can supply a range of parameters to optimize and
have a Driverless AI fit a model:
– SEIRD predicted - actual (residuals)
• Short term historical data + future predictions of infections etc., can be
incorporated into time series models.
• Other aspects of govt policy on lock downs, mandatory social distancing,
quarantines for the affected etc., can be creatively factored into the models
Incorporating COVID-19 data in AI/ML Models
8. Confidential8
• Data such as
– Unemployment Claims, Consumer Price Index, Producer
Price Index etc.,
– Mobility data – Migration data of city dwellers to rural areas
– Any other data that that your organization already has
access to, that will apply to supply chain
Incorporating Additional Macroeconomic Data
10. Confidential10
• Forecast models for short term are great (next week
forecast for instance)
• But what if situation changes Daily or Hourly ?
– We need Demand Sensing!
• Demand Sensing can alter existing short-term time
series predictions and adjust for day to day, hour to
hour or even near real-time
Daily/Hourly Predictions -> Demand Sensing
14. Confidential14
• Account for the forecasting gap (if any) and forecast horizon
• Robust time-series cross validation scheme – back testing
• Automatic handling of time group columns
• Time-series specific feature transformations
• Incorporate exogenous variables – COVID #s as an example
• Integration with your own forecasting feature, model, cost function
• Explain a forecast value with Shapley
Time Series Forecasting with Driverless AI
23. Confidential23
Feature Transformations for Forecasting
• Aggregate lag based features, eg. Exponential Moving Average
• Interaction of lags: lag1 - lag2
• Target Transformations: e.g., sqrt, log
• Decomposing the Date Column, e.g. Day of the Month, isHoliday
• Drop-out regularization : randomly remove lag features
27. Confidential27
Demand Sensing
Time Series forecast provides predictions for the sales next X number of weeks and can run every week. The
current crisis will almost always affect the forthcoming sales. Forecasts #s can be updated to produce a second
prediction/correction of the sales of the Y number of weeks and can be run daily/hourly. These daily/hourly data
could be new COVID cases, POS sales, social media data (tweets) etc.
30. Confidential30
Q Capabilities
• Analytical Database - Store data in a format that supports ad-hoc interactive
querying
• Data Prep - Merge/Join multiple data sources or transform data
• Search/Visualization - Find, filter, understand data. Visualize in ad-hoc ways
• Dashboards/Reports - Save or bookmark findings and share with others
• Integration with Driverless AI
• App SDK/Store – Author interactive apps or use existing apps
31. Confidential31
Q Data Prep
• Data pre-processing pipeline:
• Formulas (mathematical, string etc.) to create new columns.
• Text cleansing – stop word removal, stemming, punctuation and symbol removal etc.
• Merge and join tables.
35. Confidential35
Q Apps
• Programmed in 100% pure Python
• No front-end programming (HTML/Javascript/CSS) required
• No need to reason about client-server / distributed architecture
• Apps run in parallel, managed by Q scheduler
• Full-fledged workflow engine
• Apps run in isolated venv, light on resources (no Kubernetes / Docker required)
36. Confidential36
Q Demand Sensing for sales
forecasting uses augmented COVID-
19 data. It can help a company
forecast the sales for different SKUs
filtered by region/customer/brands.
The forecasting model uses sensing
variables such as – COVID-19
cases, social sentiment and more.
The Q Demand Sensing app then
shows the impact on sales with or
without sensing variables.
Demand Sensing with COVID-19 Data: Q AI App
A New AI App that Forecasts Sales with Augmented Data
37. Confidential37
Mortgage Lending with COVID-19 Data: Q AI App
One-Click
Scenario
Analysis &
Catastrophe
Modeling
The new Q
Mortgage Lending
app, uses a banks
data and
augments that
with
unemployment or
demographic data
to quickly
determine and
predict default on
loans – in other
words, a new risk
prediction with
COVID-19 data.