- 1. PredictionIO
Adam Wang
Annie Cheng
Brett Clancy
Erin Chen
Han-Wen Chen
Siyuan Wang
Yiting Wang
Koober–
- 2. Koober
Meet Tim
- 3. Koober
Tim’s Challenge
+
Provide an efficient service for customers and taxi
drivers by…
Wait Times Revenue
- 4. Tim has some questions
How should taxis be allocated
among different locations?
Koober
What factors drive customers
to use taxis more?
- 5. But most importantly…
Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
- 6. Current Solutions
Koober
iTaxi
Shiny
TaxiPrediction
- 7. Our Solution
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Koober
- 8. Why Koober?
Demand Analysis + Prediction. Empower taxi
owners with valuable rider information backed by
past trends.
Koober
Anyone in Ride Sharing Industry. Plan more
appropriate routes and make better transportation
decisions.
- 9. So how does it work?
Koober
1. Choose your info need: analysis or prediction.
OR
- 10. So how does it work?
Koober
2. Adjust your input query parameters.
- 11. So how does it work?
Koober
3. Visualize taxi demand on the heat maps.
- 12. So how does it work?
Koober
4. Make more informed taxi resource decisions.
- 13. Let’s try it out!
https://koober-dashboard.herokuapp.com/
Koober
- 14. How does it really work?
Koober
Train and test machine learning models.
Display demand on interactive heat maps.
- 15. What have we done?
Koober
Task 1: Generate dataset and extract features.
Task 2: Build various machine learning models.
Task 3: Develop the website and dashboard interface.
Task 4: Integrate Mapbox data visualization.
- 16. Raw Data
NYC Taxi & Limousine Commission:
Yellow Taxi Cab Dataset
Koober
Weather Underground API:
• Historical Weather Data
• Current Conditions
• 10-Day Forecast
Task 1
- 17. Demo Data Loader
Koober
Task 1
Development Production
• Generate fake traffic data
based on user-defined
parameters
• Simplify debugging and
validating the model
prediction result
• Process NYC Taxi Data and
integrate with the historical
weather data
• Incorporate Kafka to
facilitate importing large
amount of training data
- 18. Feature Extraction
Koober
Task 1
{
"eventTime": "2017-01-20T18:54:07.000-05:00",
"lat": 40.713802337646484,
"lng": -77.0088882446289,
"temperature": 0,
"clear": 1,
"fog": 0,
"rain": 0,
"snow": 0,
"hail": 0,
"thunder": 0,
"tornado": 0
}
Weather
Location
Event Time
Temperature
- 19. Location Clustering
Koober
Preparator:
• K-Means Location Clustering
(200 clusters)
• Each cluster represents a
neighborhood geographically
• Demand Aggregation per Unit
Time Interval and Location
Cluster
Task 1
- 20. What have we done?
Koober
Task 1: Generate dataset and extract features.
Task 2: Build various machine learning models.
Task 3: Develop the website and dashboard interface.
Task 4: Integrate Mapbox data visualization.
- 21. Koober
Task 2
Gradient-Boosted Trees
Linear Regression with Stochastic Gradient Descent
Neural Network
Random Forest
Ridge Regression
Multiple Models. Supports many classic individual
and combined machine learning models.
Models
- 22. Koober
Task 2
PredictionIO - Train
- 23. Koober
Task 2
PredictionIO - Predict
- 24. What have we done?
Koober
Task 1: Generate dataset and extract features.
Task 2: Build various machine learning models.
Task 3: Develop the website and dashboard interface.
Task 4: Integrate Mapbox data visualization.
- 25. What have we done?
Koober
Task 1: Generate dataset and extract features.
Task 2: Build various machine learning models.
Task 3: Develop the website and dashboard interface.
Task 4: Integrate Mapbox data visualization.
- 26. Koober
Task 4
User
Interface
Time
Location
Weather Predict Query
Predict Result
for Each
Algorithm
Map GL
component
Mapbox PredictionIO
Engine
Web App
Map Visualization
- 27. Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
Looking Back
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Website Data
Machine
Learning
Visualization
- 28. Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
Looking Back
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Website Data
Machine
Learning
Visualization
- 29. Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
Looking Back
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Website Data
Machine
Learning
Visualization
- 30. Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
Looking Back
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Website Data
Machine
Learning
Visualization
- 31. Koober
How can the taxi industry utilize past data
to predict the future needs of their riders?
Looking Back
Koober. An open-source interactive website for
visualizing past and predicting future taxi demand.
Website Data
Machine
Learning
Visualization
- 32. Looking Forward
Koober
ModelsTraining Data Prediction Modes
More Cities
Uber/Lyft
Better Tuning
New Models
Wait Times
Prices
- 33. Our Mentor
Engineering and Open Source Ambassador at
Salesforce.com
Koober
James Ward
www.jamesward.com
- 34. Our Professor
Assistant Professor in Computer Science at
Cornell University
Koober
Ross Tate
http://www.cs.cornell.edu/~ross/
- 35. Contribute at
Koober
https://github.com/jamesward/
koober
- 36. Questions?
Koober