MongoDB & Machine Learning - Thomas Maiaroto, Union of RAD

MongoSV 2011

Using some clever functions within MongoDB you can implement various algorithms to create trainable learning systems within your app's database. You can apply these algorithms to systems such as spam filtering, content auto-tagging, social analytics, and other classification applications. Why build these systems using MongoDB? Other than the performance benefits of Mongo's aggregation systems, you can simplify your workflow, and improve the portability of your business logic. Finally, MongoDB offers many tools that your language or toolkit of choice may not.

Transcript

1. Machine Learning Tom Maiaroto @shift8creative
2. What is Machine Learning?
3. Algorithms & Approaches Decision trees Random forests Artificial neural networks k-NN (nearest neighbour) Naive Bayesian classifier
4. Algorithms & Approaches Decision trees Random forests Artificial neural networks k-NN (nearest neighbour) Naive Bayesian classifier
5. So could machines one day rule the earth?
6. So could machines one day rule the earth? Maybe (ok probably not)
7. What can Machine Learning do for Apps? Spam filtering
8. What can Machine Learning do for Apps? Auto-tagging
9. What can Machine Learning do for Apps? All Sorts of Categorization
10. What can Machine Learning do for Apps? Sentiment Analysis
11. Languages Commonly Used • Java o Java-ML, WEKA, Apache Mahout, many more... • Python o NLTK, scikit-learn, PyML, a good deal more... • C++ o libDAI, Armadillo, Orange, tons more... and then some others...
12. Languages Commonly Used http://www.mloss.org
13. MongoDB Too! • Map/Reduce • Stored JavaScript • Geo-spatial Indexing • Replication
14. Geo-spatial Indexing Did someone say nearest neighbour?
15. Geo-spatial Indexing Did someone say nearest neighbour? Design geeks, imagine the visualizations...
16. Replication • Store massive amounts of data • Distributed performance benefits • Dedicated databases for calculations All the obvious benefits.
17. Map/Reduce It's the brain.
18. Map/Reduce It's the brain. It's not just for aggregation.
19. Map/Reduce It's the brain. It's not just for aggregation. It's faster than you might think.
20. Map/Reduce It's the brain. It's not just for aggregation. It's faster than you might think. It runs in the database.
21. Map/Reduce In the computer...
22. Example Time! It's simple...Just take this...
23. Example Time! It's simple...Just take this...
24. Example Time! Just kidding... Let's Break Down a Naive Bayes Classifier
25. Classification/Naive Bayes Training the System
26. Classification/Naive Bayes Training the System Simple... $inc
27. Classification/Naive Bayes Just Keep Count of Words per Category
28. Classification/Naive Bayes Reduce:
29. Classification/Naive Bayes Reduce:
30. Classification/Naive Bayes Finalize:
31. Classification/Naive Bayes Finalize:
32. Classification/Naive Bayes Call the Command:
33. Classification/Naive Bayes Results: Can see total words. Can also see word counts per category.
34. Classification/Naive Bayes Results: ...and of course the scores per category... cae = arts and entertainment cs = science ...
35. Classification/Naive Bayes • Accurate even with little training • MongoDB on a small VM Took 1.7 seconds • Compared to say PHP 33 seconds and timed out • More training data == exponentially faster than PHP
36. Classification/Naive Bayes • This wasn't even a full map/reduce • Your mileage will vary based on formula • You can cache certain values for speed • Don't forget about stored JavaScript (but use it wisely)
37. Porter Stemming Algorithm Thank You Martin Porter http://tartarus.org/martin/PorterStemmer
38. Porter Stemming Algorithm • Exists for nearly every language • MongoDB will use JavaScript of course • Decent execution time
39. Porter Stemming Algorithm • About 2.5x faster than PHP class • 663x faster than a web browser
40. Porter Stemming Algorithm • About 2.5x faster than PHP class • 663x faster than a web browser • 7x slower than PHP PECL extension
41. Real World Application Social Harvest Analyzes social data from the internet to determine languages spoken, gender, age, sentiment analysis, and categories. www.social-harvest.com
42. Real World Application Social Harvest Who doesn't like pie charts?
44. Follow Tom @shift8creative www.shift8creative.com www.social-harvest.com www.union-of-rad.com Thank You!