Machine Learning at Groupon: A Product Perspective

Transcript

1. Machine Learning at Groupon: A Product perspective Laura Hamilton, Group Product Manager June 8, 2017
2. Agenda - What is machine learning? - What do product managers need to know? - Applications at Groupon - The future
3. What is machine learning? Machine learning is using any of a variety of models to ingest data and make predictions. In 1959, Arthur Samuel defined machine learning as the “field of study that gives computers the ability to learn without being explicitly programmed.” (Source)
4. How powerful is machine learning? - Handwriting recognition with 99.79% accuracy (Source) - Facial recognition with 97.25% accuracy (Source) - Self-driving cars (Courses: MIT and Udacity) - Self-flying helicopters (Source) - Medical image recognition (cancer detection, tumor classification, plaque detection, etc.) - Voice interfaces (Google Home, Alexa) - Breaking Captcha (92% accuracy, comparable to human)
5. Models: A crash course - Logistic Regression - Simple and fast predictor of binary outcomes. Good for spam classification, fraud prediction, computational marketing, etc. - Linear Regression - For prediction of continuous values. A good first stab to better understand the data. Easy to interpret with the naked eye. Has a closed-form solution that is easy to program. Difficult to overfit. - Naïve Bayes - Good for multi-class classification, particularly with large datasets. Good with categorical input variables.
6. Models: A crash course...continued - Decision Trees - Use branching to model decisions and consequences. - Random Forests - What it sounds like: A whole bunch of decision trees. Less likely to overfit than a decision tree. Image: Wikipedia Image: Public Domain
7. Models: A crash course...continued - Collaborative Filtering - Good for creating a recommendation system. Used by Netflix, Groupon, and Amazon - Neural Networks - Extremely powerful, very complex. Good for image classification, facial recognition, and other complex tasks. Image: Wikipedia
8. Getting started: Tools - R - Wide variety of simple and complex machine learning packages. Great for data visualization and preprocessing. - Python - Elegant language with lots of packages. In particular check out scikit-learn. - H 2 O - Writes the code in Java for you. Excels at highly parallelized deep learning. Has a web GUI too.
9. What do PMs need to know? - What types of problems machine learning can solve - Which algorithm is best suited for which goal - What data inputs will drive the best performance of the algorithm - When a model is likely to be overfit or underfit - Warning signs if there may be a bug
10. What do PMs need to know? Cont’d An architect needs to know what materials are available to build a house, even if she isn’t doing the building herself. So too does a product manager need to know the machine learning tools that are used in modern technology products today.
11. Machine Learning at Groupon Scale - One million merchants worked with to date - 48.3 million active customers - 150 million downloads of the mobile app - One billion Groupons sold - $20 billion saved by consumers - Tens of billions of user actions per month - Decisions made in fractions of a second
12. Machine Learning at Groupon Supply Intelligence - Random forest model used to identify best sales leads Data quality - Random forest model used to dedupe database records Fraud prevention - Logistic regression model used to flag fraudulent transactions Discovery and personalization - Identify the most relevant and personalized deals for the user
13. Discovery and personalization - Tens of billions of customer activities per month - Personalize browse feed based on product views, clicks, purchases, and other features - Naïve Bayes model used to predict the probability that a user will be interested in a particular deal - Collaborative filtering used to group users with similar preferences together and personalize suggestions - Freshness algorithm penalizes multiple reimpressions
14. The Future - The “interface” will go away. It will be replaced by voice conversations. - “Ok Google, find me someplace I can go Skydiving today with Jill, Aaron, and Katie.” - “I found a Groupon for 27% off Skydive Windy City. Would you like me to make a reservation for 4 at 2:30?” - “Yes” - “Meet at Skydive Windy City at 2:15. They recommend wearing shorts and a T-shirt. Would you like me to schedule an Uber?”
15. Contact Email: Lahamilton@Groupon.com Twitter: @lauradhamilton Website: lauradhamilton.com