BSSML16 L1. Introduction, Models, and Evaluations
0 likes700 views
The document provides an overview of machine learning concepts, platforms, and evaluation techniques, focusing on BigML's approach to making machine learning accessible. It discusses the evolution of BigML's platform, various tools, methodologies for model evaluation, and the significance of data in decision-making. Key components like decision trees, accuracy metrics, and automation in model selection are highlighted to enhance understanding of machine learning applications.
![BigML, Inc 44Introduction, Models, and Evaluations
Phi Coeļ¬cient
__________TP*TN_-_FP*FN__________
SQRT[(TP+FP)(TP+FN)(TN+FP)(TN+FN)]
ā¢ Returns a value between -1 and 1
ā¢ -1 then predictions are opposite reality
ā¢ 0 no correlation between predictions
and reality
ā¢ 1 then predictions are always correct](https://image.slidesharecdn.com/bssml16-introduction-models-evaluations-161211190259/85/BSSML16-L1-Introduction-Models-and-Evaluations-44-320.jpg)
BSSML16 L1. Introduction, Models, and Evaluations
- 1. D E C E M B E R 8 - 9 , 2 0 1 6
- 2. BigML, Inc 2 Poul Petersen CIO, BigML, Inc. Intro, Models & EvaluationGetting Started with Machine Learning
- 3. BigML, Inc 3Introduction, Models, and Evaluations Audience Diversity Expert: Published papers at KDD, ICML, NIPS, etc or developed own ML algorithms used at large scale. Aļ¬cionado: Understands pros/cons of different techniques and/or can tweak algorithms as needed. Newbie: Just taking Coursera ML class or reading an introductory book to ML. Absolute beginner: ML sounds like science ļ¬ction Practitioner: Very familiar with ML packages (Weka, Scikit, R, etc).
- 4. BigML, Inc 4Introduction, Models, and Evaluations Building BigMLās Platform 2011 Prototyping and Beta API-first Approach 2013 Evaluations, Batch Predictions, Ensembles, Sunburst 2015 Association Discovery, Correlations, Samples, Statistical Tests 2014 Anomaly Detection, Clusters, Flatline 2016 Scripts, Libraries, Executions, WhizzML, Logistic Regression 2012 Core ML workflow: source, dataset, model, prediction
- 5. BigML, Inc 5Introduction, Models, and Evaluations time Automation Paving the Path to Automatic Machine Learning A REST API Programmable Infrastructure Sauron ā¢ Automatic deployment and auto-scaling Data Generation and Filtering C Flatline ā¢ DSL for transformation and new field generation B Wintermute ā¢ Distributed Machine Learning Framework 2011 2016 Automatic Model Selection E SMACdown ā¢ Automatic parameter optimization Workflow Automation D WhizzML ā¢ DSL for programmable workflows BigML Vision
- 6. BigML, Inc 6Introduction, Models, and Evaluations BigML Architecture Tools REST API Distributed Machine Learning Backend Web-based Frontend Visualizations Smart Infrastructure (auto-deployable, auto-scalable) SOURCE SERVER DATASET SERVER MODEL SERVER PREDICTION SERVER EVALUATION SERVER SAMPLE SERVER WHIZZML SERVER - https://bigml.com/tools - https://bigml.com/api SERVERS EVENTS GEARMAN QUEUE DESIRED TOPOLOGY AWS COSTS RUNQUEUE SCALER BUSY SCALER AUTO TOPOLOGY AUTO TOPOLOGY AUTO TOPOLOGY AUTO TOPOLOGY ACTUAL TOPOLOGY
- 7. BigML, Inc 7Introduction, Models, and Evaluations SOURCE DATASET CORRELATION STATISTICAL TEST MODEL ENSEMBLE LOGISTIC REGRESSION EVALUATION ANOMALY DETECTOR ASSOCIATION DISCOVERY PREDICTION BATCH PREDICTIONSCRIPT LIBRARY EXECUTION Data Exploration Supervised Learning Unsupervised Learning Automation CLUSTER Scoring BigMLās Platform
- 8. BigML, Inc 8Introduction, Models, and Evaluations What is ML? ā¢ You are looking to buy a house ā¢ Recently found a house you like ā¢ Is the asking price fair? Imagine: What Next?
- 9. BigML, Inc 9Introduction, Models, and Evaluations What is ML? Why not ask an expert? ā¢ Experts can be rare / expensive ā¢ Hard to validate experience: ā¢ Experience with similar properties? ā¢ Do they consider all relevant variables? ā¢ Knowledge of market up to date? ā¢ Hard to validate answer: ā¢ How many times expert right / wrong? ā¢ Probably canāt explain decision in detail ā¢ Humans are not good at intuitive statistics
- 10. BigML, Inc 10Introduction, Models, and Evaluations Human Intuition Consider the following two cities: Common Intuition: People in Cloud City never need sunglasses since itās so cloudy Did it occur to you: Sun City sells more sunglasses per-capita than LA Cloud City 350 grey and rainy days 15 sunny days Sun City 15 grey and rainy days 350 sunny days Question: Where is the number of sunglasses sold (per-capita) bigger?
- 11. BigML, Inc 11Introduction, Models, and Evaluations Human Intuition Imagine Mr. FernĆ”ndez is selected at random Is Mr. FernĆ”ndez more likely to be a librarian or a farmer? Did it occur to you that worldwide there is an estimatedāØ 1 billion people oļ¬cially employed in agriculture? Mr. FernĆ”ndez http://www.globalagriculture.org/report-topics/industrial-agriculture-and-small-scale-farming.html
- 12. BigML, Inc 12Introduction, Models, and Evaluations Intuitive Statistics Madrid 81 87 93Ā % 234 270 87Ā % Barcelona 192 263 73Ā % 55 80 69Ā % John Frank Wins Total Success Wins Total Success Trials 273 350 78Ā % 289 350 83Ā % John and Frank are both practicing litigation law in Madrid and Barcelona. Simpsonās Paradox A trend that appears in diļ¬erent groups of data disappears when these groups are combined, and the reverse trend appears for the aggregate data. Which attorney will you choose?
- 13. BigML, Inc 13Introduction, Models, and Evaluations What is ML? Replace the expert with data? ā¢ Intuition: square footage relates to price. ā¢ Collect data from past sales SQFT SOLD 2424 360000 1785 307500 1003 185000 4135 600000 1676 328500 1012 247000 3352 420000 2825 435350 PRICE = 125.3*SQFT + 96535 PREDICT 400262 320195 222211 614651 306538 223339 516541 450508
- 14. BigML, Inc 14Introduction, Models, and Evaluations What is ML? Price?
- 15. BigML, Inc 15Introduction, Models, and Evaluations What is ML? Price? SQFT relates to Price? SQFT SALE PRICE 2424 360000,0 1785 307500,0 1003 185000,0 4135 600000,0 1676 328500,0 1012 247000,0 3352 420000,0 2825 435350,0 PRICE = 125.3*SQFT + 96535
- 16. BigML, Inc 16Introduction, Models, and Evaluations What is ML? Replace the expert scorecard ā¢ Experts can be rare / expensive ā¢ Hard to validate experience: ā¢ Experience with similar properties? ā¢ Do they consider all relevant variables? ā¢ Knowledge of market up to date? ā¢ Hard to validate answer: ā¢ How many times expert right / wrong? ā¢ Probably canāt explain decision in detail ā¢ Humans are not good at intuitive statistics
- 17. BigML, Inc 17Introduction, Models, and Evaluations What is ML? Replace the expert with data ā¢ Intuition: square footage relates to price. ā¢ Collect data from past sales SQFT SOLD 2424 360000,0 1785 307500,0 1003 185000,0 4135 600000,0 1676 328500,0 1012 247000,0 3352 420000,0 2825 435350,0 PRICE = 125.3*SQFT + 96535
- 18. BigML, Inc 18Introduction, Models, and Evaluations More Data! SQFT BEDS BATHS ADDRESS LOCATION LOT SIZE YEAR BUILT PARKING SPOTS LATITUDE LONGITUDE SOLD 2424 4 3,0 1522 NW Jonquil Timberhill SE 2nd 5227 1991 2 44,594828 -123,269328 360000 1785 3 2,0 7360 NW Valley Vw Country Estates 25700 1979 2 44,643876 -123,238189 307500 1003 2 1,0 2620 NW Chinaberry Tamarack Village 4792 1978 2 44,593704 -123,295424 185000 4135 5 3,5 4748 NW Veronica Suncrest 6098 2004 3 44,5929659 -123,306916 600000 1676 3 2,0 2842 NW Monterey Corvallis 8712 1975 2 44,5945279 -123,291523 328500 1012 3 1,0 2320 NW Highland Corvallis 9583 1959 2 44,591476 -123,262841 247000 3352 4 3,0 1205 NW Ridgewood Ridgewood 2 60113 1975 2 44,579439 -123,333888 420000 2825 3,0 411 NW 16th Wilkins Addition 4792 1938 1 44,570883 -123,272113 435350 Uhhhhā¦ā¦..
- 19. BigML, Inc 19Introduction, Models, and Evaluations This is MLā¦ Price? SQFT relates to Price? SQFT SALE PRICE 2424 360000,0 1785 307500,0 1003 185000,0 4135 600000,0 1676 328500,0 1012 247000,0 3352 420000,0 2825 435350,0 PRICE = 125.3*SQFT + 96535 DATA MODELINSTANCE PREDICTION āa ļ¬eld of study that gives computers the ability to learn without being explicitly programmedā Professor Arthur Samuel, 1959
- 20. BigML, Inc 20 Model Demo #1
- 21. BigML, Inc 21Introduction, Models, and Evaluations Supervised Learning animal state ā¦ proximity action tiger hungry ā¦ close run elephant happy ā¦ far take picture ā¦ ā¦ ā¦ ā¦ ā¦ Classiļ¬cation animal state ā¦ proximity min_kmh tiger hungry ā¦ close 70 hippo angry ā¦ far 10 ā¦ ā¦. ā¦ ā¦ ā¦ Regression animal state ā¦ proximity action1 action2 tiger hungry ā¦ close run look untasty elephant happy ā¦ far take picture call friends ā¦ ā¦ ā¦ ā¦ ā¦ ā¦ Multi-Label Classiļ¬cation label(s)
- 22. BigML, Inc 22Introduction, Models, and Evaluations Decision Trees
- 23. BigML, Inc 23Introduction, Models, and Evaluations Decision Trees Website Visits > 0
- 24. BigML, Inc 24Introduction, Models, and Evaluations Decision Trees Minutes Used > 200
- 25. BigML, Inc 25Introduction, Models, and Evaluations Decision Trees Last Bill > $180
- 26. BigML, Inc 26Introduction, Models, and Evaluations Decision Trees Last Bill > $180 and Support Calls > 0
- 27. BigML, Inc 27Introduction, Models, and Evaluations Why Decision Trees ā¢ Works for classiļ¬cation or regression ā¢ Easy to understand: splits are features and values ā¢ Lightweight and super fast at prediction time ā¢ Relatively parameter free ā¢ Data can be messy ā¢ Useless features are automatically ignored ā¢ Works with un-normalized data ā¢ Works with missing data ā¢ Resilient to outliers ā¢ Well suited for non-linear problems ā¢ Top performer when combined into ensemblesā¦
- 28. BigML, Inc 28Introduction, Models, and Evaluations Handling Missing Data Missing@ Decision Trees KNN Logistic Regression Naive Bayes Neural Networks SVM Training Yes No No Yes Yes* No Prediction Yes No No Yes No No
- 29. BigML, Inc 29Introduction, Models, and Evaluations Data Types numeric 1 2 3 1, 2.0, 3, -5.4 categoricaltrue, yes, red, mammal categoricalcategorical A B C DATE-TIME2013-09-25 10:02 DATE-TIME YEAR MONTH DAY-OF-MONTH YYYY-MM-DD DAY-OF-WEEK HOUR MINUTE YYYY-MM-DD YYYY-MM-DD M-T-W-T-F-S-D HH:MM:SS HH:MM:SS 2013 September 25 Wednesday 10 02 text / items Be not afraid of greatness: some are born great, some achieve greatness, and some have greatness thrust upon 'em. text āgreatā āafraidā ābornā āsomeā appears 2 times appears 1 time appears 1 time appears 2 times
- 30. BigML, Inc 30Introduction, Models, and Evaluations Text Analysis Be not afraid of greatness: some are born great, some achieve greatness, and some have greatness thrust upon 'em. great: appears 4 times Bag of Words
- 31. BigML, Inc 31Introduction, Models, and Evaluations Text Analysis great afraid born achieve 4 1 1 1 ā¦ ā¦ ā¦ ā¦ Be not afraid of greatness: some are born great, some achieve greatness, and some have greatness thrust upon āem. Model The token āgreatā does not occur The token āafraidā occurs more than once
- 32. BigML, Inc 32 Model Demo #2
- 33. BigML, Inc 33Introduction, Models, and Evaluations Learning Problems (ļ¬t) ā¢ Model does not ļ¬t well enough ā¢ Does not capture the underlying trend of the data ā¢ Change algorithm or features Under-ļ¬tting Over-ļ¬tting ā¢ Model ļ¬ts too well does not āgeneralizeā ā¢ Captures the noise or outliers of the data ā¢ Change algorithm or ļ¬lter outliers
- 34. BigML, Inc 34Introduction, Models, and Evaluations Why Not Decision Trees ā¢ Slightly prone to over-ļ¬tting ā¢ But weāll ļ¬x this with ensembles ā¢ Splitting prefers decision boundaries that are parallel to feature axes ā¢ More data ā¢ Predictions outside training data can be problematic ā¢ Weāll ļ¬x this with model competence ā¢ Can be sensitive to small changes in training data
- 35. BigML, Inc 35Introduction, Models, and Evaluations Evaluation DATASET TRAIN SET TEST SET PREDICTIONS METRICS
- 36. BigML, Inc 36Introduction, Models, and Evaluations Accuracy TP + TN Total ā¢ āPercentage correctā - like an exam ā¢ = 1 then no mistakes ā¢ = 0 then all mistakes ā¢ Intuitive but not always useful ā¢ Watch out for unbalanced classes!
- 37. BigML, Inc 37Introduction, Models, and Evaluations Accuracy Classiļ¬ed as Fraud Classiļ¬ed as Not Fraud TP = 0 FP = 0 TN = 7 FN = 3 ACC = 70% =Fraud =Not FraudPositive Class Negative Class
- 38. BigML, Inc 38Introduction, Models, and Evaluations Precision __TP__ TP + FP ā¢ āaccuracyā of positive class ā¢ = 1 then no FP ā¢ = 0 then no TP
- 39. BigML, Inc 39Introduction, Models, and Evaluations Precision Classiļ¬ed as Fraud Classiļ¬ed as Not Fraud TP = 2 FP = 2 TN = 5 FN = 1 P = 50% =Fraud =Not FraudPositive Class Negative Class
- 40. BigML, Inc 40Introduction, Models, and Evaluations Recall __TP__ TP + FN ā¢ percentage of positive class correctly identiļ¬ed ā¢ = 1 then no FN ā¢ = 0 then no TP
- 41. BigML, Inc 41Introduction, Models, and Evaluations Recall Classiļ¬ed as Fraud Classiļ¬ed as Not Fraud TP = 2 FP = 2 TN = 5 FN = 1 R = 66% =Fraud =Not FraudPositive Class Negative Class
- 42. BigML, Inc 42Introduction, Models, and Evaluations f-Measure 2 * Recall * Precision Recall + Precision ā¢ harmonic mean of Recall & Precision ā¢ = 1 then Recall = Precision = 1 ā¢ If Precision OR Recall is small then f-measure is small
- 43. BigML, Inc 43Introduction, Models, and Evaluations f-Measure Classiļ¬ed as Fraud Classiļ¬ed as Not Fraud R = 66% P = 50% f = 57% =Fraud =Not FraudPositive Class Negative Class
- 44. BigML, Inc 44Introduction, Models, and Evaluations Phi Coeļ¬cient __________TP*TN_-_FP*FN__________ SQRT[(TP+FP)(TP+FN)(TN+FP)(TN+FN)] ā¢ Returns a value between -1 and 1 ā¢ -1 then predictions are opposite reality ā¢ 0 no correlation between predictions and reality ā¢ 1 then predictions are always correct
- 45. BigML, Inc 45Introduction, Models, and Evaluations Phi Coeļ¬cient Classiļ¬ed as Fraud Classiļ¬ed as Not Fraud TP = 2 FP = 2 TN = 5 FN = 1 Phi = 0.356 =Fraud =Not FraudPositive Class Negative Class
- 46. BigML, Inc 46 Model Demo #3
- 47. BigML, Inc 47Introduction, Models, and Evaluations Evaluations
- 48. BigML, Inc 48Introduction, Models, and Evaluations Evaluations
- 49. BigML, Inc 49Introduction, Models, and Evaluations Evaluations
- 50. BigML, Inc 50Introduction, Models, and Evaluations Mean Absolute Error e1 e2 e7 e6 e5 e4 e3 MAE = |e1| + |e2| + ā¦ + |en| n
- 51. BigML, Inc 51Introduction, Models, and Evaluations Mean Squared Error e1 e2 e7 e6 e5 e4 e3 MSE = (e1)2 + (e2)2 + ā¦ + (en)2 n
- 52. BigML, Inc 52Introduction, Models, and Evaluations MSE / MAE ā¢ For both MAE & MSE: Smaller is better, but values are unbounded ā¢ MSE is always larger than or equal to MAE
- 53. BigML, Inc 53Introduction, Models, and Evaluations R Squared Error e1 e2 e7 e6 e5 e4 e3 Mean v1 v2 v3 v4 v5 v7 v6 MSEmodel MSEmean RSE = 1 -
- 54. BigML, Inc 54Introduction, Models, and Evaluations R-Squared Error ā¢ RSE: measure of how much better the model is than always predicting the mean ā¢ < 0 model is worse then mean ā¢ = 0 model is no better than the mean ā¢ = 1 model ļ¬ts the data perfectly
- 55. BigML, Inc 55 Model Demo #3