1. Overview of Machine Learning & Feature Engineering Machine Learning 101 Tutorial Strata + Hadoop World, NYC, Sep 2015 Alice Zheng, Dato 1
2. 2 About us Chris DuBois Intro to recommenders Alice Zheng Overview of ML Piotr Teterwak Intro to image search & deep learning Krishna Sridhar Deploying ML as a predictive service Danny Bickson TA Alon Palombo TA
3. 3 Why machine learning? Model data. Make predictions. Build intelligent applications.
4. Classification Predict amongst a discrete set of classes 4
5. 5 Input Output
6. 6 Spam filtering data prediction Spam vs. Not spam
7. Text classification EDUCATION FINANCE TECHNOLOGY
8. Regression Predict real/numeric values 8
9. 9 Stock market Input Output
10. Similarity Find things like this 10
11. 11 Similar products Product I’m buying Output: other products I might be interested in
12. 12 Given image, find similar images http://www.tiltomo.com/
13. Recommender systems Learn what I want before I know it 13
14. 14
15. 15 Playlist recommendations Recommendations form coherent & diverse sequence
16. 16 Friend recommendations Users and “items” are of the same type
17. Clustering Grouping similar items 17
18. 18 Clustering images Goldberger et al. Set of Images
19. 19 Clustering web search results
20. 20 Machine learning … how? Data Answers I fell in love the instant I laid my eyes on that puppy. His big eyes and playful tail, his soft furry paws, … Many systems Many tools Many teams Lots of methods/jargon
21. 21 The machine learning pipeline I fell in love the instant I laid my eyes on that puppy. His big eyes and playful tail, his soft furry paws, … Raw data Features Models Predictions Deploy in production
22. 22 Three things to know about ML • Feature = numeric representation of raw data • Model = mathematical “summary” of features • Making something that works = choose the right model and features, given data and task
23. Feature = numeric representation of raw data
24. 24 Representing natural text It is a puppy and it is extremely cute. What’s important? Phrases? Specific words? Ordering? Subject, object, verb? Classify: puppy or not? Raw Text {“it”:2, “is”:2, “a”:1, “puppy”:1, “and”:1, “extremely”:1, “cute”:1 } Bag of Words
25. 25 Representing natural text It is a puppy and it is extremely cute. Classify: puppy or not? Raw Text Bag of Words it 2 they 0 I 1 am 0 how 0 puppy 1 and 1 cat 0 aardvark 0 cute 1 extremely 1 … … Sparse vector representation
26. 26 Representing images Image source: “Recognizing and learning object categories,” Li Fei-Fei, Rob Fergus, Anthony Torralba, ICCV 2005—2009. Raw image: millions of RGB triplets, one for each pixel Classify: person or animal? Raw Image Bag of Visual Words
27. 27 Representing images Classify: person or animal? Raw Image Deep learning features 3.29 -15 -5.24 48.3 1.36 47.1 - 1.92 36.5 2.83 95.4 -19 -89 5.09 37.8 Dense vector representation
28. 28 Feature space in machine learning • Raw data high dimensional vectors • Collection of data points point cloud in feature space • Feature engineering = creating features of the appropriate granularity for the task
29. Crudely speaking, mathematicians fall into two categories: the algebraists, who find it easiest to reduce all problems to sets of numbers and variables, and the geometers, who understand the world through shapes. -- Masha Gessen, “Perfect Rigor”
30. 30 Algebra vs. Geometry a b c a2 + b2 = c2 Algebra Geometry Pythagorean Theorem (Euclidean space)
31. 31 Visualizing a sphere in 2D x2 + y2 = 1 a b c Pythagorean theorem: a2 + b2 = c2 x y 1 1
32. 32 Visualizing a sphere in 3D x2 + y2 + z2 = 1 x y z 1 1 1
33. 33 Visualizing a sphere in 4D x2 + y2 + z2 + t2 = 1 x y z 1 1 1
34. 34 Why are we looking at spheres? = = = = Poincaré Conjecture: All physical objects without holes is “equivalent” to a sphere.
35. 35 The power of higher dimensions • A sphere in 4D can model the birth and death process of physical objects • High dimensional features can model many things
36. Visualizing Feature Space
37. 37 The challenge of high dimension geometry • Feature space can have hundreds to millions of dimensions • In high dimensions, our geometric imagination is limited - Algebra comes to our aid
38. 38 Visualizing bag-of-words puppy cute 1 1 I have a puppy and it is extremely cute I have a puppy and it is extremely cute it 1 they 0 I 1 am 0 how 0 puppy 1 and 1 cat 0 aardvark 0 zebra 0 cute 1 extremely 1 … …
39. 39 Visualizing bag-of-words puppy cute 1 1 1 extremely I have a puppy and it is extremely cute I have an extremely cute cat I have a cute puppy
40. 40 Document point cloud word 1 word 2
41. Model = mathematical “summary” of features
42. 42 What is a summary? • Data point cloud in feature space • Model = a geometric shape that best “fits” the point cloud
43. 43 Clustering model Feature 2 Feature 1 Group data points tightly
44. 44 Classification model Feature 2 Feature 1 Decide between two classes
45. 45 Regression model Target Feature Fit the target values
46. Visualizing Feature Engineering
47. 47 When does bag-of-words fail? puppy cat 2 1 1 have I have a puppy I have a cat I have a kitten Task: find a surface that separates documents about dogs vs. cats Problem: the word “have” adds fluff instead of information I have a dog and I have a pen 1
48. 48 Improving on bag-of-words • Idea: “normalize” word counts so that popular words are discounted • Term frequency (tf) = Number of times a terms appears in a document • Inverse document frequency of word (idf) = • N = total number of documents • Tf-idf count = tf x idf
49. 49 From BOW to tf-idf puppy cat 2 1 1 have I have a puppy I have a cat I have a kitten idf(puppy) = log 4 idf(cat) = log 4 idf(have) = log 1 = 0 I have a dog and I have a pen 1
50. 50 From BOW to tf-idf puppy cat1 have tfidf(puppy) = log 4 tfidf(cat) = log 4 tfidf(have) = 0 I have a dog and I have a pen, I have a kitten 1 log 4 log 4 I have a cat I have a puppy Decision surface Tf-idf flattens uninformative dimensions in the BOW point cloud
51. 51 Entry points of feature engineering • Start from data and task - What’s the best text representation for classification? • Start from modeling method - What kind of features does k-means assume? - What does linear regression assume about the data?
52. Dato’s Machine Learning Platform
53. 53 Dato’s machine learning platform Raw data Features Models Predictions Deploy in production GraphLab Create Dato Distributed Dato Predictive Services
54. 54 Data structures for feature engineering Features SFrames User Com. Title Body User Disc. SGraphs
55. 55 Machine learning toolkits in GraphLab Create • Classification/regression • Clustering • Recommenders • Deep learning • Similarity search • Data matching • Sentiment analysis • Churn prediction • Frequent pattern mining • And on…
56. Demo
57. 57 Dimensionality reduction Feature 1 Feature 2 Flatten non-useful features PCA: Find most non-flat linear subspace
58. 58 PCA : Principal Component Analysis Center data at origin
59. 59 PCA : Principal Component Analysis Find a line, such that the average distance of every data point to the line is minimized. This is the 1st Principal Component
60. 60 PCA : Principal Component Analysis Find a 2nd line, - at right angles to the 1st - such that the average distance of every data point to the line is minimized. This is the 2nd Principal Component
61. 61 PCA : Principal Component Analysis Find a 3rd line - at right angles to the previous lines - such that the average distance of every data point to the line is minimized. … There can only be as many principle components as the dimensionality of the data.
62. Demo
63. 63 Coursera Machine Learning Specialization • Learn machine learning in depth • Build and deploy intelligent applications • Year long certification program • Joint project between University of Washington + Dato • Details: https://www.coursera.org/specializations/machine-learning
64. 64 Next up today alicez@dato.com @RainyData, #StrataConf 11:30am - Intro to recommenders Chris DuBois 1:30pm - Intro to image search & deep learning Piotr Teterwak 3:30pm - Deploying ML as a predictive service Krishna Sridhar