Workshop on Machine Learning
1 like660 views
The document provides an overview of machine learning, including its definitions, history, and various cultural approaches within the field, such as statistics and AI/ML cultures. It outlines key concepts like the machine learning process, data exploration techniques using R, and types of machine learning models. Additionally, it emphasizes the importance of evaluation metrics and the challenges associated with processing large datasets.
Workshop on Machine Learning
- 1. Machine Learning what, why and how
- 2. me ... Harshad ➔ Senior Data Scientist @ Sokrati ➔ Spent last 4 years trying to understand and apply machine learning Sokrati is a digital advertising startup based out of Pune
- 3. what are we going to do ?
- 4. get a 10000 feet view ...
- 5. then go to specifics ...
- 7. what is ML ? ➔ Too many definitions! ➔ Too much debate ➔ Analytics vs ML vs data mining vs Big Data vs Statistics vs next buzzword in the market
- 8. what is ML ? Teaching machines to take decisions with the help of data practical man’s definition of machine learning!
- 9. bit of history ...
- 10. That’s too ancient! That’s not ML...
- 11. bit of (relevant) history ... ➔ Insurance, Banking industry ◆ Credit scoring ◆ mathematical models in finance ➔ Artificial Intelligence and other fancy ideas ◆ deep blue, Samuel’s checkers machine ◆ IBM watson computer
- 12. Find ‘f’ => endeavour to understand the world! g( y ) = f( X )
- 13. the two cultures in ML Stats Culture Vs AI / ML Culture
- 14. the two cultures in ML Stats Culture ➔ Focus on ‘why this model’ ➔ Goodness of fit, hypothesis tests, residuals ➔ or MCMC methods, bayesian modeling ➔ regression models, survival analysis AI / ML Culture ➔ Focus on ‘good predictions’ ➔ Cross validations, ensemble of models ➔ Focus on underfit vs overfit analysis ➔ neural nets, tree based models (random forests et. al.)
- 15. but let’s build bridges Stats ➔ Focus on basics, sound theory ➔ Exploration, summaries ➔ Models ML / AI ➔ Focus on predictions ➔ Model evaluation ➔ Feature selection Computations ➔ Focus on application ➔ Achieving scale and usability ➔ Hadoop , Storm and friends.. Business Knowledge ➔ Focus on interpretation ➔ Visualizations ➔ Creating stories out of data
- 16. typical ML process ➔ Objective ➔ Source data ➔ Explore ➔ Model ➔ Evaluate ➔ Apply ➔ Validate
- 18. bottom line : not in terms of algorithm but outcome probability of customer churn group set of emails by topic predict rainfall recommend item to a consumer to increase likelihood of click
- 19. sourcing data to be covered at end!
- 20. explore R you ready ?
- 21. introduction to R ➔ Starting R ➔ Data Structures ◆ Atomic Vectors, matrix ◆ Lists ◆ Data frames ➔ Data types ◆ usual suspects in numerics (int, double, character) ◆ Factors ◆ logical
- 22. data frame , the workhorse ➔ Load sample data frame ➔ Explore data frame (head, tail) ➔ Access elements by index ◆ access rows ◆ access columns ● single ● multiple (by name, by index, by -ve index) ➔ Find metadata ◆ names ◆ dimensions ➔ Explore using plot (pairs)
- 23. broadcasting / vectorization ➔ Very important concept ➔ Subsetting ◆ vectors ◆ data frames ➔ Applying operations ◆ operations on entire column
- 24. data exploration ➔ summarizing using mean ➔ quantiles, when mean is not enough ◆ outlier detection ➔ functional roots of R : sapply summary ➔ summary function ◆ on numerical ◆ on factors ➔ plots (basic) ➔ histograms ➔ correlations
- 25. models simple & real world
- 26. basic types of models ➔ Supervised learning ➔ Unsupervised learning ➔ Semi-supervised learning ➔ Reinforcement learning
- 27. linear regression ➔ load sample dataset (cars) ➔ build linear regression model ➔ understand the output ◆ summary ◆ plots ➔ understanding train vs predict cycle ◆ most important idea!
- 28. demo on real world dataset data exploration, classification
- 29. hierarchical clustering ➔ basic idea of clustering ◆ distance as a proxy for similarity, group by distance ◆ group anything as long as distance can be calculated ➔ load and explore eurodist data ➔ fit hierarchical cluster ➔ plot dendrogram
- 30. demo on real world data and the most important idea!
- 31. concept of vector space model ➔ Words as axis ➔ Bag of words defines vector space ➔ Document as a point in space ➔ We can ◆ define distance ◆ measure similarity (cosine similarity) ◆ group documents ➔ what can be a document ?
- 32. evaluation
- 33. evaluation metrics ➔ depends on type of model ◆ regression : MAPE, MSSE ◆ classification : accuracy, precision, recall, F score ◆ clustering : within vs between variance ➔ ML world (ref : two cultures) has much better story ➔ Not enough to perform well on training set
- 34. brief intro regularization ➔ Bias vs Variance problem ➔ We want to be ‘just right’ ➔ Concept of regularization ➔ Intro Cross validation
- 35. demo of evaluation and fantastic Scikits API
- 36. sourcing and applying and the great ML divide
- 37. the great ML divide Lab Culture ➔ Theory ➔ Small Datasets ➔ In memory ➔ Not live ➔ R, Octave, Python.. Source and Apply ➔ The practice ➔ Huge datasets ➔ Live in production ➔ Hadoop and friends, Python ? , R ?
- 38. processing data at scale ➔ Data is not available in final form ➔ Non standard data ◆ click streams ◆ event logs ◆ free form text ➔ Process at scale ➔ Transform , clean, group in final form
- 39. 5 min intro to Hadoop Ecosystem ➔ Write in assembly ◆ java ➔ DSLs ◆ Pig ◆ hive ◆ impala ➔ Functional Languages to rescue!
- 40. Introduction to Cascalog processing data at scale
- 41. Recap ➔ Pragmatic ML ➔ Key phases ➔ Supervised learning ➔ Unsupervised learning ➔ Evaluation ➔ Application issues ➔ Processing data @ scale
- 42. Let’s discuss...