Art of Feature Engineering for Data Science with Nabeel Sarwar
4 likes2,074 views
This document discusses feature engineering in data science, particularly in the context of machine learning at Comcast. It covers various processes involved in feature engineering such as selecting, cleaning, transforming, and enriching data, as well as methodologies for handling categorical and numerical variables. Additionally, it highlights the importance of domain knowledge, statistical correlation, and the challenges of dimensionality in optimizing machine learning models.
Art of Feature Engineering for Data Science with Nabeel Sarwar
- 1. Art of Feature Engineering For Data Science Nabeel Sarwar Machine Learning Engineer at Comcast Spark Summit Europe - October 26th, 2017
- 2. What is Comcast NBCUniversal? • Much like Ireland’s Sky • Provide Broadband Services in the United States - Internet - Video - Media - Mobile • ~25 million customers on combination of broadband services • Use Case: machine learning to optimize customer experience - While complying with all applicable privacy policies and laws
- 3. XFINITY TV XFINITY Internet XFINITY Voice XFINITY Home ® Digital & OtherOther *Minority interestand/or non-controlling interest. Slide is notcomprehensiveofall ComcastNBCUniversal assets Updated:December 22,2015
- 4. Where in the process? Obtain Select Clean Transform & Enrich Define Model Selection Tuning & Eval Feedback Imagine Arrows Everywhere
- 5. Feature Engineering 101 • Cheat sheet to some of the terminology and best practices • Why is feature engineering necessary? • How and how much feature engineering? • Feature Engineering Process variations • Select • Clean • Normalize and Transform • Philosophies - Confirm a hypothesis - Explore trends - Middle ground approach
- 6. Features • ML algorithms need pieces of information • Each individual piece in a sample: feature • Encoded in all sort of different ways • Differentiate by use case - Insurance: Model of car - Imaging: Color of pixels - General Relativity (black hole detection): coordinates in space - Online Food Ordering: what you ate last • Differentiate by model - Sequencing: Last event states - Imaging: Channels of the pixels - Reinforcement Learning: Actions from state to state • Differentiate by available data: - How do you partition? - Feature selection
- 7. Example: Titanic Survival Prediction • Who survives on the titanic? • Select the features • Prioritize who survives - Age - Role on ship - Speaking tongue of passenger • Pick Decision Tree • Bucket Age • Keep role as is • Encode speaking tongue into binary variable whether they speak English or not
- 8. How to actually do it
- 9. Categorical Variables • Discrete (but not necessarily disjoint) levels • Can often seem numeric • Encoding methods - Leave as is (but encode into some number - StringIndexer) - One Hot Encoding - Binary Encoding • Text: Tokenizing & Word2Vec • Decision Trees and variants can often do worse with one hot encoding
- 10. Numerical Variables • Can be continuous or discrete • Can become categorical through binning and bucketting • To scale or not variables - Some variables might be weighted higher than others (KNN?) - Per mini-batch in neural network training - Normalize or max-min scaling • Always determine relevant stats: - Mean - Variance - Kurtosis - Mode - Median • Specifically for words: term frequency-inverse documentfrequency (tf-idf)
- 11. Cleaning Data • Correcting for input mistakes or known mistakes in the data (normalization) - If predicting text: Knowing the last few sequence was Iceland instead of ice land • Transforming missing data and outliers - Missing Data -> Imputation - Mean - Mode - Matrix Factorization (A = LDU) • Converting data into known formats for rest of pipeline - Can be very simple: Mapping categories to different numbers - Or very difficult: Cleaning irregularly sampled data
- 12. Selecting Features • Domain Knowledge • Statistically correlate features - Weights - Gini - Correlation - Information Gain - Information Criteria (AIC, BIC, Bayes Factors) • Global Gains vs Local Gains • Throw all the data and then down sample • Throw in little data and then increase until acceptable tolerance • Noisy or dirty data • Some models prefer no intra feature other interactions - Ordinary Least Squares and some exponential families
- 13. Transformations and Enrichment • Build Features from Features - Add - Subtract (the mean) - Multiply - Ratios - Polynomials (kernel trick in SVMs) - Rational Differences - Logarithms, Exponentials, Sigmoids - Fourier or Laplace Transforms • Encode categorical variables • Scaling numerical variables • Look up tables to bring in additional features • Transforming dates - Year, month, day, hour, minute, second - Holidays and Special Events val encoder = new OneHotEncoder() .setInputCol(”categories") .setOutputCol(”oneHot") val encodedDF = encoder.transform(categoryDF)
- 14. Dimension Reduction • Curse of Dimensionality - Unhealthy prunes - Distances look the same - Slow training • Optimization in production • Reduce number of dimensions by learning the most important “directions” or combination of features - Finding the right basis (eigen) vectors • Principal Component Analysis - Singular Value Decomposition on Covariance Matrix • Factor Analysis - Expectation-Maximization to find basis vectors • Non-negative Matrix Factorization - Poisson-gamma factorization • Just remove features or transform better!
- 15. A Harder Use Case: Star-Galaxy Separation • Stars and galaxies look like little points in space, hard to differentiate on images • Can read temperatures • Know that Stars are one-dimensional based on temperature • Galaxies are composed of various stars • Bin the temperatures • Logarithm of temperatures • Form Features from the combination of differences • Logarithms on same magnitudes - No need to scale
- 18. Philosophy • Explore: - Not sure what exactly is there - Try out different combinations until you see trends - Select all and throw out as needed • Confirmation: - Have some hypothesis on how data is supposed to behave - Statistical tests - Add Feature if it passes your confirmation test • Hybrid: - Data sets often have a combination of features that fit both philosophies
- 19. New Horizons • Deep Learning: Feature Engineering -> Architecture building - Still need good data and good features • Meta Features - Models building features for other models • Automation! - Combinatorial complexity