Machine Learning From Raw Data To The Predictions
1 like98 views
The document discusses machine learning concepts and Azure Machine Learning. It begins with an introduction to Azure ML services and environment. It then covers important steps for data preparation, including checking for missing values and outliers, performing feature engineering and selection. Common machine learning algorithms are also outlined, such as regression, classification, clustering and anomaly detection. The presentation concludes with a demonstration of how to build a predictive model for price elasticity using Azure ML Studio.
Machine Learning From Raw Data To The Predictions
- 1. Machine Learning From Raw Data To The Predictions Luca Zavarella
- 4. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Who Am I? Data Science Microsoft Professional Program Microsoft SQL Server BI MCTS & MCITP Working with SQL Server since 2007 BI & Advanced Analytics Technical Director @ Email: lzavarella@solidq.com Twitter: @lucazav LinkedIn: http://it.linkedin.com/in/lucazavarella Survey: http://bit.do/ml-survey
- 5. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Agenda • What’s Azure Machine Learning • How to Prepare you Data • Missing Values • Outliers • Feature Engineering • Feature Selection • ML Algorithms • Demo
- 7. #SqlSat675 – 18/11/2017http://bit.do/ml-survey The Azure Machine Learning Offer Azure ML Services Experimentation Service Model Management Service Workbench Azure ML Studio Deploy on Web Service DSVM
- 8. #SqlSat675 – 18/11/2017http://bit.do/ml-survey The Azure ML Studio Environment Data Azure Machine Learning Consumers Cloud storage Azure Storage Azure Table Hive etc. Local storage Upload data from PC… Business Apps Excel Business problem Modeling Business valueDeployment API Azure Marketplace (Applications store) Azure ML Gallery (community) ML Web Services (REST API Services) ML Studio (Web IDE) Workspace: Experiments Datasets Trained models Notebooks Access settings Data Model Manage API
- 9. DEMO 1 The Azure ML Studio Environment
- 10. Data Preparation
- 11. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Why we Need to Prepare Data? • Feed ML algorithms the right data for the problem to solve • It has to be in a useful scale • It has to be in the right format • Meaningful features are to be included
- 12. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Useful Steps for Data Preparation • Here some steps to follow: • Check missing values • Check outliers • Do some Feature Engineering • Do some Feature Selection
- 13. Missing Values
- 14. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Missing Values - Ask Your Self Some Questions • What data is missing? • Which variables and how much is missing? • Do we know why it is missing? • It’s missing because it should be missing or are cases incomplete? • What would be implications for modeling? • Do we have enough samples to model without it? • Do we have bias due to missing values?
- 15. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Missing Values - Three Options, No Silver Bullets • Do nothing! • Data is truly missing and should not exist • The lack of information can become a new feature (true/false) • Data is absent when it should be there in a random way • Weight the data • Create a pseudo-population of weighted copies of the complete cases to remove selection bias introduced by the missing data • Imputing Missing Values • Impute the mean, median or most common value • Relationships with other variables are lost • Imputation through Predictive Models • Impute several values for each missing case • Return several imputed data sets • Allow you to run multiple models and pool them together at the end
- 16. Outliers
- 17. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Outliers – What They Are • An outlier is an observation that lies an abnormal distance from other values in a random sample from a population • The “abnormal” distance hasn’t a fixed measure! It depends on data • The analyst will decide what will be considered abnormal • Extreme values sometimes have a big effect on statistical operations • That effect is not necessarily a good effect
- 18. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Outliers – Why They Are There Data Entry Errors Measurement Error There are 10 weighing machines. 9 of them are correct, 1 is faulty Experimental Error In a 100m sprint of 7 runners, one runner missed out on concentrating on the ‘Go’ call which caused him to start late Intentional Outlier Teens would typically under report the amount of alcohol that they consume Data Processing Error Some manipulation or extraction errors may lead to outliers in the dataset Sampling error We have to measure the height of athletes; by mistake, we include a few basketball players in the sample Natural Outlier Malls sell more products in the Christmas period.
- 19. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Outliers – Extreme and Mild Ones The Boxplot is your friend! Just for univariate analysis and for bivariate analysis with a categorical variable…
- 20. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Multivariate Outliers – 2D • Cases with an unusual combination of values on different variables • Finding outlier in 2d is more tricky than in univariate analysis
- 21. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Multivariate Outliers – More Than 2D Finding outliers in >2d can be well hidden!
- 22. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Outliers – Remove Them or What? • Generally, it’s a bad idea to remove them • It decreases the ability of the model to generalize if the cause of this type of variance is unknown (it can be not just noise) • The outliers effect can be often mitigated by appropriate math transformations (log! ☺) • Remove them only if they are truly “abnormal” • Sometimes you can substitute them with NA and then impute them, or replace them with the mean/median
- 24. #SqlSat675 – 18/11/2017http://bit.do/ml-survey What is Feature Engineering • It’s the process of using domain knowledge of the data to create features that make ML algorithms work in a better way • Eg. Annual Debt Ratio = Monthly Debt * 12 / Annual Income • FE is sometimes necessary in order to turn raw input data into things the ML algorithm can understand • It’s often an art and it’s the most creative phase of Data Processing
- 25. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Examples of Simple Feature Engineering ItemID ItemColor 1 Yellow 2 Green 3 Unknown 4 Green 5 Yellow ItemId HasColor IsYellow IsGreen 1 1 1 0 2 1 0 1 3 0 0 0 4 1 0 1 5 1 1 0 One-hot encoding or Dummy variables ItemID Price 1 10 2 12 3 5 4 23 5 76 ItemID PriceRange 1 Low 2 Medium 3 Low 4 Medium 5 High Binning / Bucketing 1-10 Low 11 – 25 Medium 26 – 80 High
- 26. #SqlSat675 – 18/11/2017http://bit.do/ml-survey An Example of “Exotic” Feature Engineering • Cyclical features need to be treated carefully • Hours: the model can’t understand that 24h == 0h • In a linear model the “length” (t0 – t23) = 23 is “huge” • Use the clock hand’s projections to convert hours to smooth linear variables: 𝑥 = 𝑐𝑜𝑠 𝜋 2 − 2𝜋ℎ 24 = 𝑠𝑖𝑛 2𝜋ℎ 24 and 𝑦 = 𝑐𝑜𝑠 2𝜋ℎ 24 ItemID Hours 1 22 2 23 3 0 4 1 5 2 X_Hours Y_Hours ItemId X_Hours Y_Hours 1 -0.5000 0.8660 2 -0.2588 0.9659 3 0.0000 1.0000 4 0.2588 0.9659 5 0.5000 0.8660
- 27. DEMO 2 Data Preparation on Azure ML Studio Data Exploration with IDEAR
- 29. #SqlSat675 – 18/11/2017http://bit.do/ml-survey What is Feature Selection • It’s the process of selecting a subset of relevant features for use in model construction • FS will include and exclude attributes present in the data without changing them • Dimensionality reduction is a different thing • New combination of attributes maximizing total variance • Methods: Principal Component Analysis, Singular Value Decomposition, Sammon’s Mapping • Reduce the noise due to irrelevant features • Avoid to over-fit the training set • Make the model generalized
- 30. ML Algorithms
- 31. #SqlSat675 – 18/11/2017http://bit.do/ml-survey ML Algorithms by Learning Style • Supervised Learning • Predicting the future • Learn from known past examples • Labels provided • Unsupervised Learning • Making sense of data • Understanding the past • Learning the structure of data • Labels not provided • Reinforcement Learning • Machine trained to make decisions using trial and error • Maximize a cumulative reward • Exploration vs exploitation trade-off Broadly split into three main categories:
- 32. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Common Supervised Learning Algorithms 1 • Regression • Modeling and predicting continuous, numeric variables • Linear Regression • Straightforward to understand • Not flexible enough to capture more complex patterns • Regression Tree • Can learn non-linear relationships • Fairly robust to outliers (using ensembles) • Ensembles perform very well • They combine predictions from many individual trees
- 33. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Common Supervised Learning Algorithms 2 • Classification • Modeling and predicting categorical variables • Logistic Regression • Yes! “Regression” but not for continuous variables • Predictions ∈ 0, 1 class probabilities • Not flexible enough to capture complex relationships • Classification (or Decision) Tree • Robust to outliers and scalable • Naturally model non-linear decision boundaries • Individual trees are prone to overfitting, not if in ensembles
- 34. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Common Unsupervised Learning Algorithms • Clustering • Find natural groupings of observations based on the inherent structure within your dataset • K-Means • Make clusters based on geometric distance • Grouped around centroids in a globular form • Simple and flexible • Number of clusters to be specified • Anomaly Detection • Identify unusual patterns not conform to expected behavior • Find outliers! ☺ • Used when there are lot of "normal" data and not many cases of anomalies • One-class Support Vector Machine (SVM) • Points in space, mapped so that separate categories are divided by a clear gap that is as wide as possible • PCA-Based • First computes its projection on the eigenvectors, then computes the normalized reconstruction error • The higher the error, the more anomalous the instance is
- 35. #SqlSat675 – 18/11/2017http://bit.do/ml-survey No Free Lunch! • The “No Free Lunch” theorem in ML • No one algorithm works best for every problem • Many factors at play, e.g. size and structure of dataset • Try many different algorithms for your problem!
- 36. DEMO 3 Price Elasticity in Azure ML
- 37. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Lesson Learned • It’s easy to create a predictive model with Azure ML Studio • But you have to know how to model data before put it in the pot to avoid bad predictions!
- 38. #SqlSat675 – 18/11/2017http://bit.do/ml-survey Links • What is Azure ML (https://is.gd/zUaGXs) • Price Elasticity Experiment (https://is.gd/nq8jJs) • Elasticity in Economics (https://is.gd/652EGX) • TDSP Utility IDEAR (https://is.gd/5KKdPM) • How to Better Evaluate the Goodness-of-Fit of Regressions (https://is.gd/2z30H4)