(video summary at: https://conversionhotel.com/session/keynote-2019-ai-personalization-demystified/)
Guy has a background that is quite impressive. From Princeton to 3 years Boston Consulting Group in the 90’s and then on for his MBA at Stanford and MBA internship at Microsoft. Guy did his time at an online marketplace start-up that got acquired before he moved to Yahoo! to have several product marketing positions for 7 years in a row. He has been the head of global product and industry marketing at Twitter and was member of the executive team at Brightroll. Nowadays he is the founder and CEO of a start-up that uses machine learning to optimize websites.
The number 1 outcome of the #CH2019 attendee question on what the hot topics in our industry are for the coming year, was: personalization. AI and Machine Learnings were also high on this list, the combination of these topics is the sweet-spot where we all know of that the companies we work for are going to invest heavenly in the next coming years. But as true data driven optimizers we also know that many companies won’t have the users to do proper personalization and/or machine learning. We also believe that they should fix the basics first…
Nevertheless we know we are going to have to deal a lot with personalization and machine learning in our jobs. It may sound scary, because you don’t exactly know what it is and how to apply it, but you also know that this is the way to the future (or was McDonald’s buying Dynamic Yield a big fail?). But how to take the first steps in a proper way? When I thought of a potential speaker that could demystify AI and personalization for us – I thought of Guy.
Happy learning,
Ton Wesseling
Founder & host of The Conference formerly known as Conversion Hotel
12. Intellimize Confidential
Belief #1: Manage both ABM and anonymous traffic
ABM for named accounts
using rules + ML
Welcome visitor by
company name
Mention account
executive by name
Iterate on calls to
action and pieces of
content
Rest of your anonymous
traffic using ML
Show an industry
specific case study
Iterate headlines, calls
to action, and pieces
of content
Tailor content based
on previous
behavior on the site
Note: This is an example. These are not things Okta is doing
Note: ABM = Account Based Marketing. ML = machine learning
13. Intellimize Confidential
Belief #2: Testing and personalization go together
Optimize to deliver more revenue /
customers / leads
A/B testing
Find a better way to engage everyone
Personalization
Find a better way to engage each
segment or unique visitor
14. Intellimize Confidential
Belief #2: Testing and personalization go together
Optimize to deliver more revenue /
customers / leads
A/B testing
Find a better way to engage everyone
Rules based personalization
Find a better way to engage each
segment
Faster
Find a winner for everyone
Slower
Test to find a winner for each segment
15. Intellimize Confidential
Belief #2: Testing and personalization go together
Optimize to deliver more revenue /
customers / leads
A/B testing
Find a better way to engage everyone
Machine learning personalization
Find a better way to engage each
segment or unique visitor
Slower
Find a winner for everyone
Faster
Find a winner for each unique visitor
16. Intellimize Confidential
Using ML to learn about prospects
25x
faster
learning
Value
messaging
Speed
messaging
Innovation
messaging
18. Intellimize Confidential
Machine learning branches: supervised learning
We teach the computer how to predict something based on input
● Most common machine learning problem
● Trained on historical data where the “right answer” is known
○ These are called ‘labeled’ training examples
● Maps inputs (what you know now) to outputs (success or failure) using
historical and accurate data
● Uses this mapping to guess the value of future events that the model does not
already know about, based on what the model has already learned
19. Intellimize Confidential
Machine learning branches: unsupervised learning
The computer teaches itself how to predict something. You don’t yet know what
that something is
● Trained on historical examples where the outcome not known (ie you don’t
know what success looks like)
○ These are called ‘unlabeled’ training examples
● The objective is to discover structure in the data (eg through a cluster analysis)
○ This objective is not to map inputs to outputs
20. Intellimize Confidential
Machine learning branches: reinforcement learning
You train the computer to achieve some eventual goal
● Only know the final outcome (aka unlabeled intermediate steps)
● Only get to learn based on the choices model makes
● Explore-exploit tradeoff to balance
21. Intellimize ConfidentialIntellimize Confidential
Type of ML problem
What is machine learning good at in marketing?
● Lead scoring
● Ideal price
● Ideal promotion amount
Regression
(predicting a continuous value or number)
● Will this person click on this ad (yes or no)
● High / medium / low lead score
● Email is spam or not spam
Classification
(predicting among discrete options)
● Which product to show
● Which content to show
Recommendation
● Understand speech (ie NLP)
● Recognize the content of an image
● Write an email subject line
Speech / image recognition
● Customer segmentation
● Find business insights from data
Clustering
● Fraud detection
● Outlier detection
Anomaly detection
● Sequence of emails to send
● CRO
Reinforcement learning
Problem
Unsupervised
learningRL
Supervised
learning
21
22. Intellimize ConfidentialIntellimize Confidential
Type of ML problem
Theory vs realistic application
Regression
(predicting a continuous value or number)
Classification
(predicting among discrete options)
Recommendation
Speech / image recognition
Clustering
Anomaly detection
Reinforcement learning
22
Realistic implementation: content reco
Classification: offensive / non-offensive
Classification: popular / not popular
Collaborative filtering: ranked list
Rule: no more than 2 items from each category
23. Intellimize Confidential
Type of ML problem
Theory vs realistic application
Regression
(predicting a continuous value or number)
Classification
(predicting among discrete options)
Recommendation
Speech / image recognition
Clustering
Anomaly detection
Reinforcement learning
23
Realistic implementation: ensemble learning
Regression #1
Regression #2
Regression #3
Regression #4
Regression #5
Score
25. Intellimize Confidential
1. Most AI you see in market is “if this… then that…” rules based logic
2. Testing and personalization should be used together for revenue/customers/leads
3. The best outcomes happen when we combine optimizers and AI
4. AI can accelerate your testing and personalization materially
5. b2b marketers can tailor prospects’ journeys using both ABM and AI together
6. Practically applying AI is much more than picking the right algorithm
7. You can and should meet your prospects where they are in their journey with you
Top 7 takeaways