Introduction to Data Analytics with R
1 like1,044 views
The document provides an introduction to data analytics using R given by Wei Zhong from NUS. It begins with an overview of Wei Zhong and his background in computational biology. The agenda is then outlined, covering key concepts in data analytics like logistic regression, decision trees, random forests, and evaluation metrics. An overview of data analytics is presented, distinguishing descriptive, predictive, and prescriptive analytics. Statistical learning techniques like linear models, tree-based models, and clustering are introduced. Key concepts like cross-validation that will be used in the hands-on session are then defined.
Introduction to Data Analytics with R
- 1. NUS DataScience Introduction to Data Analytics with R TOH Wei Zhong 31/10/2015
- 2. A little bit about me • Graduated from NUS, Computational Biology • Statistics and computing onto biology and healthcare • E.g. –omics • Data Scientist in NCS • Smart Nation projects (defense and public safety) NUS-DataScience TOH Wei Zhong 231/10/2015
- 3. Agenda for this afternoon • Overview of data analytics • Introduce key concepts for hands-on session • Logistic regression • Decision tree • Random forest • Evaluation metrics • Cross-validation • Short break • Hands-on NUS-DataScience TOH Wei Zhong 331/10/2015
- 4. Overview of data analytics NUS-DataScience TOH Wei Zhong 431/10/2015
- 5. What is data analytics? • A collection of established methods/techniques that • Seeks to make sense of and generate insights and knowledge from collected data (Big Data or otherwise) • Is statistically sound and rigorous • Preferably scalable • Is used to support decision making NUS-DataScience TOH Wei Zhong 531/10/2015
- 6. Data Science Data analytics Data visualization Big Data technologies Text and network analytics NLP Semantics Social media Hadoop Data streaming Spark Tableau Communication Grammar of Graphics NUS-DataScience TOH Wei Zhong 631/10/2015
- 7. A common way to think about data analytics Given existing data, generate some form of summary / aggregated view so that data can be consumed Given existing data, construct models so that predictions on future, yet-to-be collected data can be made Given constructed models, recommend future decisions Descriptive Predictive Prescriptive NUS-DataScience TOH Wei Zhong 731/10/2015
- 8. Key aspects that businesses are concerned about Accuracy Value-adding Interpretability “Factors associated” NUS-DataScience TOH Wei Zhong 831/10/2015
- 9. Key techniques in data analytics • Feature selection • Clustering • Linear models • Tree-based models • Evaluation metrics • Resampling methods • Hypothesis testing • Association rule mining • Time series analysis • Feature engineering Statistical learning Sometimes neglected, but nonetheless powerful NUS-DataScience TOH Wei Zhong 931/10/2015
- 10. Statistical learning • Supervised and unsupervised learning (also, semi- supervised learning) • Supervised learning: learning with ground truth/answers available (response variable) • Classification: response variable is categorical • Regression: response variable is continuous or numerical • Unsupervised learning: finding intrinsic relationships between samples in the dataset • Clustering algorithms NUS-DataScience TOH Wei Zhong 1031/10/2015
- 11. Supervised learning: linear models • Generalized linear models (GLM): mainstay tool in data analytics • Generalized in the sense of the type of response variable: • Continuous response variable: ordinary least squares (OLS) regression • Binary / multinomial response: logistic regression • Discrete response: Poisson regression • Gives an equation: y = β0 + β1x1 + β2x2 + … + ε • Regularization (ridge regression, LASSO regression) NUS-DataScience TOH Wei Zhong 1131/10/2015
- 12. Supervised learning: tree-based models • Models that uses decision trees as fundamental building blocks • Random forest • Gradient boosting machines • Rotation forest • More on decision trees and random forest later NUS-DataScience TOH Wei Zhong 1231/10/2015
- 13. Unsupervised learning: clustering • Clustering: empirically grouping observations / samples / rows in a dataset together in different groups (cluster), such that the more similar observations are grouped together • Unsupervised because there is no ground truth to guide the process, unlike e.g. regression NUS-DataScience TOH Wei Zhong 1331/10/2015
- 14. Feature selection • Feature: a variable / attribute in the dataset • Feature selection: the process of selecting relevant features that aids in the modelling process, used especially when there are too many features in the dataset to work with • Curse of dimensionality: the more irrelevant features are used in a model, the weaker the model NUS-DataScience TOH Wei Zhong 1431/10/2015
- 15. Evaluation metrics • Measures using which constructed models are assessed • Examples include accuracy and ROC-AUC • Later NUS-DataScience TOH Wei Zhong 1531/10/2015
- 16. Key concepts NUS-DataScience TOH Wei Zhong 1631/10/2015
- 17. Hands-on session • For the hands-on session, we will be look at a dataset of emails, consisting of both spam and non- spam • The objective is to construct models that can predict whether a given email is spam or non-spam NUS-DataScience TOH Wei Zhong 1731/10/2015
- 18. NUS-DataScience TOH Wei Zhong 1831/10/2015
- 19. A bit on R • R is a statistical computing language that was developed with statistical analysis in mind • One of the most popular tools in the data science community • R scripts: sequence of procedures that enables step-by- step customized data crunching • R packages: collations of R scripts (functions) that we can leverage on to do various, more complex tasks easily, e.g. manipulate data and construct models • R and Rstudio NUS-DataScience TOH Wei Zhong 1931/10/2015
- 20. Key concepts to be used • Logistic regression • Decision tree • Random forest • Cross-validation • Evaluation metrics: accuracy and ROC-AUC NUS-DataScience TOH Wei Zhong 2031/10/2015
- 21. Logistic regression • A type of generalized linear model (GLM) • Assigns each variable used in the model with a coefficient that can be used in summation to predict log-odds • y = β0 + β1x1 + β2x2 + … + ε • Log-odds = log(odds) ∝ probability • In our case, probability of an email being a spam email NUS-DataScience TOH Wei Zhong 2131/10/2015
- 22. Pros and cons of logistic regression • Pros: • Easy to interpret – the idea of regression is familiar and intuitive • Cons: • Requires certain statistical assumptions to hold true in the data • Generally low predictive accuracy NUS-DataScience TOH Wei Zhong 2231/10/2015
- 23. Decision trees • A simple model used in supervised learning • CART, C4.5 – amongst top 10 most popular data mining algorithms • Can handle both classification and regression • The tree package that we are using uses the recursive partitioning algorithm NUS-DataScience TOH Wei Zhong 2331/10/2015
- 24. Equivalents • Tree == Binary partitioning of dataset • Each partition is represented by the mode (classification) or mean (regression) NUS-DataScience TOH Wei Zhong 2431/10/2015
- 25. Terminologies • Depth • Node • Leaf nodes • Non-leaf nodes • The size of a tree sometimes refers to the number of leaf nodes • Parents and children • Branching factor NUS-DataScience TOH Wei Zhong 2531/10/2015
- 26. Pruning • Typically after the construction of a decision tree, we would want to prune the tree, because the tree may be overly complicated NUS-DataScience TOH Wei Zhong 2631/10/2015
- 27. Pruning (2) • Pruning refers to the process of trimming the tree to a more compact and concise one, without sacrificing much performance • The tree package uses cost-complexity pruning • Comparing the relationship between number of leaf nodes and performance of model NUS-DataScience TOH Wei Zhong 2731/10/2015
- 28. Pros and cons of decision trees • Pros: • Very easy to interpret and communicate to others, because it is similar to how humans think and make decisions • Easy to construct • Cons: • Generally unstable • Generally low predictive accuracy NUS-DataScience TOH Wei Zhong 2831/10/2015
- 29. Random forest • In the RF model, instead of using one decision tree to do predictions, we use multiple of them • The idea is to build decision trees on different subsets of the training data • Each subset is known as a “bag” • Each bag yields one decision tree • To make a prediction, we ask each tree to make a predictions • To get the overall prediction of the RF model, we take a majority vote NUS-DataScience TOH Wei Zhong 2931/10/2015
- 30. Pros and cons of random forest • Pros: • One of the top-performing models in supervised learning • With some basic understanding of sampling and bootstrapping, RF can be easy to communicate. The intuition of voting as a mechanism to make decisions is simple • Able to derive variable importance measures • Cons: • Computationally intensive NUS-DataScience TOH Wei Zhong 3031/10/2015
- 31. Evaluation metrics: assessing the performance of a supervised learning model • In order to know whether the models constructed can perform well in reality, we need to assess some metrics to assess their performance • Classification: accuracy / error rate • Sensitivity, specificity etc. • Regression: mean squared error • 𝑀𝑆𝐸 = 1 𝑛 (𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 − 𝑎𝑐𝑡𝑢𝑎𝑙)2 • Also, there are two types of classification models: (1) Those that output classes / categories as predictions (2) Those that output probabilities as predictions • (2): can use ROC-AUC as a measure of performance NUS-DataScience TOH Wei Zhong 3131/10/2015
- 32. Cross-validation • Gives rise to the idea of training and testing datasets • Rationale: • Recall that the constructed models are ultimately meant to do predictions on future, unknown observations • Models are constructed/trained using input datasets. We call them training data • If the models constructed are too attuned to the training data => overfitting NUS-DataScience TOH Wei Zhong 3231/10/2015
- 33. NUS-DataScience TOH Wei Zhong 3331/10/2015 http://www.turingfinance.com/regression- analysis-using-python-statsmodels-and- quandl/
- 34. Cross-validation • In order to know whether our models are overfitted to the data, we use cross-validation • Split the dataset in two parts: training and testing • Use the training set to build the models • Use the models to make predictions on the testing set • A way to think about this: studying for an examination NUS-DataScience TOH Wei Zhong 3431/10/2015
- 35. Hands-on NUS-DataScience TOH Wei Zhong 3531/10/2015