know Machine Learning Basic Concepts.pdf
- 2. Terminology Machine Learning, Data Science, Data Mining, Data Analysis, Sta- tistical Learning, Knowledge Discovery in Databases, Pattern Dis- covery.
- 3. Data everywhere! 1. Google: processes 24 peta bytes of data per day. 2. Facebook: 10 million photos uploaded every hour. 3. Youtube: 1 hour of video uploaded every second. 4. Twitter: 400 million tweets per day. 5. Astronomy: Satellite data is in hundreds of PB. 6. . . . 7. “By 2020 the digital universe will reach 44 zettabytes...” The Digital Universe of Opportunities: Rich Data and the Increasing Value of the Internet of Things, April 2014. That’s 44 trillion gigabytes!
- 4. Data types Data comes in different sizes and also flavors (types): ⊠ Texts ⊠ Numbers ⊠ Clickstreams ⊠ Graphs ⊠ Tables ⊠ Images ⊠ Transactions ⊠ Videos ⊠ Some or all of the above!
- 5. Smile, we are ’DATAFIED’ ! • Wherever we go, we are “datafied”. • Smartphones are tracking our locations. • We leave a data trail in our web browsing. • Interaction in social networks. • Privacy is an important issue in Data Science.
- 6. The Data Science process T i m e DATA COLLECTION Static Data. Domain expertise 1 3 4 5 ! DB% DB EDA MACHINE LEARNING Visualization Descriptive statistics, Clustering Research questions? Classification, scoring, predictive models, clustering, density estimation, etc. Data-driven decisions Application deployment Model%(f)% Yes!/! 90%! Predicted%class/risk% A!and!B!!!C! Dashboard Static Data. 2 DATA PREPARATION Data!cleaning! + + + + + - + + - - - - - - + Feature/variable! engineering!
- 7. Applications of ML • We all use it on a daily basis. Examples:
- 8. Machine Learning • Spam filtering • Credit card fraud detection • Digit recognition on checks, zip codes • Detecting faces in images • MRI image analysis • Recommendation system • Search engines • Handwriting recognition • Scene classification • etc...
- 10. ML versus Statistics Statistics: • Hypothesis testing • Experimental design • Anova • Linear regression • Logistic regression • GLM • PCA Machine Learning: • Decision trees • Rule induction • Neural Networks • SVMs • Clustering method • Association rules • Feature selection • Visualization • Graphical models • Genetic algorithm http://statweb.stanford.edu/~jhf/ftp/dm-stat.pdf
- 11. Machine Learning definition “How do we create computer programs that improve with experi- ence?” Tom Mitchell http://videolectures.net/mlas06_mitchell_itm/
- 12. Machine Learning definition “How do we create computer programs that improve with experi- ence?” Tom Mitchell http://videolectures.net/mlas06_mitchell_itm/ “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. ” Tom Mitchell. Machine Learning 1997.
- 13. Supervised vs. Unsupervised Given: Training data: (x1, y1), . . . , (xn, yn) / xi ∈ Rd and yi is the label. example x1 → x11 x12 . . . x1d y1 ← label . . . . . . . . . . . . . . . . . . example xi → xi1 xi2 . . . xid yi ← label . . . . . . . . . . . . . . . . . . example xn → xn1 xn2 . . . xnd yn ← label
- 14. Supervised vs. Unsupervised Given: Training data: (x1, y1), . . . , (xn, yn) / xi ∈ Rd and yi is the label. example x1 → x11 x12 . . . x1d y1 ← label . . . . . . . . . . . . . . . . . . example xi → xi1 xi2 . . . xid yi ← label . . . . . . . . . . . . . . . . . . example xn → xn1 xn2 . . . xnd yn ← label
- 15. Supervised vs. Unsupervised Unsupervised learning: Learning a model from unlabeled data. Supervised learning: Learning a model from labeled data.
- 16. Unsupervised Learning Training data:“examples” x. x1, . . . , xn, xi ∈ X ⊂ Rn • Clustering/segmentation: f : Rd −→ {C1, . . . Ck} (set of clusters). Example: Find clusters in the population, fruits, species.
- 19. Unsupervised learning Feature'2' Feature'1' Methods: K-means, gaussian mixtures, hierarchical clustering, spectral clustering, etc.
- 20. Supervised learning Training data:“examples” x with “labels” y. (x1, y1), . . . , (xn, yn) / xi ∈ Rd • Classification: y is discrete. To simplify, y ∈ {−1, +1} f : Rd −→ {−1, +1} f is called a binary classifier. Example: Approve credit yes/no, spam/ham, banana/orange.
- 23. Supervised learning !"#$%&"'( ' !"#$%&"')' *"+,-,./'0.%/1#&2' Methods: Support Vector Machines, neural networks, decision trees, K-nearest neighbors, naive Bayes, etc.
- 25. Supervised learning Non linear classification
- 26. Supervised learning Training data:“examples” x with “labels” y. (x1, y1), . . . , (xn, yn) / xi ∈ Rd • Regression: y is a real value, y ∈ R f : Rd −→ R f is called a regressor. Example: amount of credit, weight of fruit.
- 27. Supervised learning Regression: !" #$%&'($")" Example: Income in function of age, weight of the fruit in function of its length.
- 33. K-nearest neighbors • Not every ML method builds a model! • Our first ML method: KNN. • Main idea: Uses the similarity between examples. • Assumption: Two similar examples should have same labels. • Assumes all examples (instances) are points in the d dimen- sional space Rd.
- 34. K-nearest neighbors • KNN uses the standard Euclidian distance to define nearest neighbors. Given two examples xi and xj: d(xi, xj) = v u u u t d X k=1 (xik − xjk)2
- 35. K-nearest neighbors Training algorithm: Add each training example (x, y) to the dataset D. x ∈ Rd, y ∈ {+1, −1}.
- 36. K-nearest neighbors Training algorithm: Add each training example (x, y) to the dataset D. x ∈ Rd, y ∈ {+1, −1}. Classification algorithm: Given an example xq to be classified. Suppose Nk(xq) is the set of the K-nearest neighbors of xq. ŷq = sign( X xi∈Nk(xq) yi)
- 37. K-nearest neighbors 3-NN. Credit: Introduction to Statistical Learning.
- 38. K-nearest neighbors 3-NN. Credit: Introduction to Statistical Learning. Question: Draw an approximate decision boundary for K = 3?
- 39. K-nearest neighbors Credit: Introduction to Statistical Learning.
- 40. K-nearest neighbors Question: What are the pros and cons of K-NN?
- 41. K-nearest neighbors Question: What are the pros and cons of K-NN? Pros: + Simple to implement. + Works well in practice. + Does not require to build a model, make assumptions, tune parameters. + Can be extended easily with news examples.
- 42. K-nearest neighbors Question: What are the pros and cons of K-NN? Pros: + Simple to implement. + Works well in practice. + Does not require to build a model, make assumptions, tune parameters. + Can be extended easily with news examples. Cons: - Requires large space to store the entire training dataset. - Slow! Given n examples and d features. The method takes O(n × d) to run. - Suffers from the curse of dimensionality.
- 43. Applications of K-NN 1. Information retrieval. 2. Handwritten character classification using nearest neighbor in large databases. 3. Recommender systems (user like you may like similar movies). 4. Breast cancer diagnosis. 5. Medical data mining (similar patient symptoms). 6. Pattern recognition in general.
- 44. Training and Testing !"#$%$%&'()*' +,'-.&/"$*01' +/2).'345' 6%7/1)8'' &)%2)"8'' #&)8'' 4#1$.9'(*#*:(8' ;$<7/2)' =")2$*'#1/:%*'>' =")2$*'9)(?%/' Question: How can we be confident about f?
- 45. Training and Testing • We calculate Etrain the in-sample error (training error or em- pirical error/risk). Etrain(f) = n X i=1 ℓoss(yi, f(xi))
- 46. Training and Testing • We calculate Etrain the in-sample error (training error or em- pirical error/risk). Etrain(f) = n X i=1 ℓoss(yi, f(xi)) • Examples of loss functions: – Classification error: ℓoss(yi, f(xi)) = ( 1 if sign(yi) 6= sign(f(xi)) 0 otherwise
- 47. Training and Testing • We calculate Etrain the in-sample error (training error or em- pirical error/risk). Etrain(f) = n X i=1 ℓoss(yi, f(xi)) • Examples of loss functions: – Classification error: ℓoss(yi, f(xi)) = ( 1 if sign(yi) 6= sign(f(xi)) 0 otherwise – Least square loss: ℓoss(yi, f(xi)) = (yi − f(xi))2
- 48. Training and Testing • We calculate Etrain the in-sample error (training error or em- pirical error/risk). Etrain(f) = n X i=1 ℓoss(yi, f(xi)) • We aim to have Etrain(f) small, i.e., minimize Etrain(f)
- 49. Training and Testing • We calculate Etrain the in-sample error (training error or em- pirical error/risk). Etrain(f) = n X i=1 ℓoss(yi, f(xi)) • We aim to have Etrain(f) small, i.e., minimize Etrain(f) • We hope that Etest(f), the out-sample error (test/true error), will be small too.
- 51. Structural Risk Minimization Predic'on*Error * Low*******************************************Complexity*of*the*model*************************************High* ____Test*error**** ____Training*error* High*Bias****** * * * * * * * ***Low*Bias** Low*Variance* * * * * * * * *High*Variance* UnderfiAng****************Good*models** * * *OverfiAng * ***********
- 53. Training and Testing !"#$%&' ()&' !"#$%&' ()&' High bias (underfitting) !"#$%&' ()&'
- 54. Training and Testing !"#$%&' ()&' !"#$%&' ()&' High bias (underfitting) !"#$%&' ()&' High variance (overfitting)
- 55. Training and Testing !"#$%&' ()&' !"#$%&' ()&' High bias (underfitting) Just right! !"#$%&' ()&' High variance (overfitting)
- 56. Avoid overfitting In general, use simple models! • Reduce the number of features manually or do feature selec- tion. • Do a model selection (ML course). • Use regularization (keep the features but reduce their impor- tance by setting small parameter values) (ML course). • Do a cross-validation to estimate the test error.
- 57. Regularization: Intuition We want to minimize: Classification term + C × Regularization term n X i=1 ℓoss(yi, f(xi)) + C × R(f)
- 58. Regularization: Intuition !"#$%&' ()&' !"#$%&' ()&' !"#$%&' ()&' f(x) = λ0 + λ1x ... (1) f(x) = λ0 + λ1x + λ2x2 ... (2) f(x) = λ0 + λ1x + λ2x2 + λ3x3 + λ4x4 ... (3) Hint: Avoid high-degree polynomials.
- 59. Train, Validation and Test TRAIN VALIDATION TEST Example: Split the data randomly into 60% for training, 20% for validation and 20% for testing.
- 60. Train, Validation and Test TRAIN VALIDATION TEST 1. Training set is a set of examples used for learning a model (e.g., a classification model).
- 61. Train, Validation and Test TRAIN VALIDATION TEST 1. Training set is a set of examples used for learning a model (e.g., a classification model). 2. Validation set is a set of examples that cannot be used for learning the model but can help tune model parameters (e.g., selecting K in K-NN). Validation helps control overfitting.
- 62. Train, Validation and Test TRAIN VALIDATION TEST 1. Training set is a set of examples used for learning a model (e.g., a classification model). 2. Validation set is a set of examples that cannot be used for learning the model but can help tune model parameters (e.g., selecting K in K-NN). Validation helps control overfitting. 3. Test set is used to assess the performance of the final model and provide an estimation of the test error.
- 63. Train, Validation and Test TRAIN VALIDATION TEST 1. Training set is a set of examples used for learning a model (e.g., a classification model). 2. Validation set is a set of examples that cannot be used for learning the model but can help tune model parameters (e.g., selecting K in K-NN). Validation helps control overfitting. 3. Test set is used to assess the performance of the final model and provide an estimation of the test error. Note: Never use the test set in any way to further tune the parameters or revise the model.
- 64. K-fold Cross Validation A method for estimating test error using training data. Algorithm: Given a learning algorithm A and a dataset D Step 1: Randomly partition D into k equal-size subsets D1, . . . , Dk Step 2: For j = 1 to k Train A on all Di, i ∈ 1, . . . k and i 6= j, and get fj. Apply fj to Dj and compute EDj Step 3: Average error over all folds. k X j=1 (EDj)
- 65. Confusion matrix !"#$%$&' (')*%$&' !"#$%$&' !"#$%&'()*)+$%,!&- ./0($%&'()*)+$%,.&- (')*%$&' ./0($%1$2/*)+$%,.1- !"#$%1$2/*)+$%,!1- 34*#/0%;/<$0% &"$=)4*$=%;/<$0
- 66. Evaluation metrics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
- 67. Terminology review Review the concepts and terminology: Instance, example, feature, label, supervised learning, unsu- pervised learning, classification, regression, clustering, pre- diction, training set, validation set, test set, K-fold cross val- idation, classification error, loss function, overfitting, under- fitting, regularization.
- 68. Machine Learning Books 1. Tom Mitchell, Machine Learning. 2. Abu-Mostafa, Yaser S. and Magdon-Ismail, Malik and Lin, Hsuan-Tien, Learning From Data, AMLBook. 3. The elements of statistical learning. Data mining, inference, and prediction T. Hastie, R. Tibshirani, J. Friedman. 4. Christopher Bishop. Pattern Recognition and Machine Learn- ing. 5. Richard O. Duda, Peter E. Hart, David G. Stork. Pattern Classification. Wiley.
- 69. Machine Learning Resources • Major journals/conferences: ICML, NIPS, UAI, ECML/PKDD, JMLR, MLJ, etc. • Machine learning video lectures: http://videolectures.net/Top/Computer_Science/Machine_Learning/ • Machine Learning (Theory): http://hunch.net/ • LinkedIn ML groups: “Big Data” Scientist, etc. • Women in Machine Learning: https://groups.google.com/forum/#!forum/women-in-machine-learning • KDD nuggets http://www.kdnuggets.com/
- 70. Credit • The elements of statistical learning. Data mining, inference, and prediction. 10th Edition 2009. T. Hastie, R. Tibshirani, J. Friedman. • Machine Learning 1997. Tom Mitchell.