top 10 data science practitioner pitfalls

H2O.ai Machine Intelligence

Top 10 Data Science Practitioner Pitfalls

Erin LeDell and Mark Landry

Silicon Valley Big Data Science September 2015


H2O.ai

H2O Company

H2O Software

• Team: ~35. Founded in 2012, Mountain View, CA• Stanford Math & Systems Engineers

• Open Source Software (Apache 2.0 License)• Ease of Use via Web Interface• R, Python, Scala, Spark & Hadoop Interfaces• Distributed Algorithms Scale to Big Data


Scientific Advisory CouncilDr. Trevor Hastie

Dr. Rob Tibshirani

Dr. Stephen Boyd

• John A. Overdeck Professor of Mathematics, Stanford University• PhD in Statistics, Stanford University• Co-author, The Elements of Statistical Learning: Prediction, Inference and Data Mining• Co-author with John Chambers, Statistical Models in S• Co-author, Generalized Additive Models • 108,404 citations (via Google Scholar)

• Professor of Statistics and Health Research and Policy, Stanford University• PhD in Statistics, Stanford University• COPPS Presidents’ Award recipient• Co-author, The Elements of Statistical Learning: Prediction, Inference and Data Mining• Author, Regression Shrinkage and Selection via the Lasso• Co-author, An Introduction to the Bootstrap

• Professor of Electrical Engineering and Computer Science, Stanford University• PhD in Electrical Engineering and Computer Science, UC Berkeley• Co-author, Convex Optimization• Co-author, Linear Matrix Inequalities in System and Control Theory• Co-author, Distributed Optimization and Statistical Learning via the Alternating Direction

Method of Multipliers


What is Data Science?

Problem Formulation

• Identify an outcome of interest and the type of task: classification / regression / clustering

• Identify the potential predictor variables• Identify the independent sampling units

• Conduct research experiment (e.g. Clinical Trial)• Collect examples / randomly sample the population• Transform, clean, impute, filter, aggregate data• Prepare the data for machine learning — X, Y

• Modeling using a machine learning algorithm (training)• Model evaluation and comparison• Sensitivity & Cost Analysis

• Translate results into action items• Feed results into research pipeline

Collect & Process Data

Machine Learning

Insights & Action


Classification

Clustering

Machine Learning Task Overview

• Predict a real-valued response (viral load, weight)• Gaussian, Gamma, Poisson and Tweedie • MSE and R^2

• Multi-class or Binary classification• Ranking• Accuracy and AUC

• Unsupervised learning (no training labels)• Partition the data / identify clusters• AIC and BIC

Regression


Machine Learning Workflow

Source: NLTK

Example of a supervised machine learning workflow.


Train vs Test1 of 10



1. Train vs TestTraining Set vs.

Test Set• Partition the original data (randomly or stratified) into

a training set and a test set. (e.g. 70/30)

• It can be useful to evaluate the training error, but you should not look at training error alone.

• Training error is not an estimate of generalization error (on a test set or cross-validated), which is what you should care more about.

• Training error vs test error over time is an useful thing to calculate. It can tell you when you start to overfit your model, so it is a useful metric in supervised machine learning.

• Be careful of data leakage (from the training set into the test set).

• If you are using pooled repeated measures data (vs iid data), you must ensure that all rows associated with a cluster/individual are either in train or test, but not in both.

Training Error vs.Test Error

Data Leakage


1. Train vs Test Error

Source: Elements of Statistical Learning


Validation Set2 of 10



2. Train vs Test vs ValidTraining Set vs.

Validation Set vs.Test Set

• If you have “enough” data and plan to do some model tuning, you should really partition your data into three parts — Training, Validation and Test sets.

• There is no general rule for how you should partition the data and it will depend on how strong the signal in your data is, but an example could be: 50% Train, 25% Validation and 25% Test

• The validation set is used strictly for model tuning (via validation of models with different parameters) and the test set is used to make a final estimate of the generalization error.

Validation is for Model Tuning


Model Performance3 of 10



3. Model Performance

Test Error • Partition the original data (randomly) into a training set and a test set. (e.g. 70/30)

• Train a model using the training set and evaluate performance (a single time) on the test set.

• Train & test K models as shown.

• Average the model performance over the K test sets.

• Report cross-validated metrics.

• Regression: R^2, MSE, RMSE• Classification: Accuracy, F1, H-measure, Log-loss• Ranking (Binary Outcome): AUC, Partial AUC

K-foldCross-validation

Performance Metrics


Class Imbalance4 of 10



4. Class Imbalance

Imbalanced Response Variable

• A dataset is said to be imbalanced when the binomial or multinomial response variable has one or more classes that are underrepresented in the training data, with respect to the other classes.

• This is incredibly common in real-word datasets.• In practice, balanced datasets are the rarity, unless

they have been artificially created.• There is no precise definition of what defines an

imbalanced vs balanced dataset — the term is vague.• My rule of thumb for binary response: If the minority

class makes <10% of the data, this can cause issues.

• Advertising — Probability that someone clicks on ad is very low… very very low.

• Healthcare & Medicine — Certain diseases or adverse medical conditions are rare.

• Fraud Detection — Insurance or credit fraud is rare.

Very common

Industries


4. Simple RemediesArtificial Balance • You can balance the training set using sampling.

• Notice that we don’t say to balance the test set. The test set represents the true data distribution. The only way to get “honest” model performance on your test set is to use the original, unbalanced, test set.

• The same goes for the hold-out sets in cross-validation. For this, you may end up having to write custom code, depending on what software you use.

• H2O has a “balance_classes” argument that can be used to do this properly & automatically.

• You can manually upsample (or downsample) your minority (or majority) class(es) set either by duplicating (or sub-sampling) rows, or by using row weights.

• The SMOTE (Synthetic Minority Oversampling Technique) algorithm generates simulated training examples from the minority class instead of upsampling.

Potential Pitfalls

Solutions


4. Advanced Remedies

AUC-MaximizingAlgorithms

• There are ways to tackle this issue more directly.• By using algorithms that optimize a metric that is

insensitive to prior class probabilities — for example, Area Under the ROC Curve (AUC).

• Many algorithms work by optimizing a metric equivalent or similar to accuracy. If your data is imbalanced, this will not produce a good model since you can have excellent accuracy and poor AUC.

Cost-Sensitive Training

• Use a cost function to penalize the types of errors you care about most more harshly.

• Cost Matrix:


Categorical Data5 of 10



5. Categorical DataReal Data • Most real world datasets contain categorical data.

• Problems can arise if you have too many categories.• A lot of ML software will place limits on the number of

categories allowed in a single column (e.g. 1024) so you may be forced to deal with this whether you like it or not.

• When there are high-cardinality categorical columns, often there will be many categories that only occur a small number of times (not very useful).

• If you have some hierarchical knowledge about the data, then you may be able to reduce the number of categories by using some sensible higher-level mapping of the categories.

• Example: ICD-9 codes — thousands of unique diagnostic and procedure codes. You can map each category to a higher level super-category to reduce the cardinality.

Too Many Categories

Solutions


5. Missing CategoriesMissing Data • There are many approaches to imputing categorical

data. The simplest approach is to impute all missing values with the mode (the category that occurs most).

• When your data is split into training and test sets, there may be categories that are represented in the training set but not in the test set and vice versa.

• If you have expanded your categorical variable into a group of binary indicator columns equal to the number of categories, then new categories in the test set should not cause any problems. Example: If you expand a categorical (cat, dog) into “cat” and “dog” indicator columns and your test set has a “rat” in it, then the value in each of those columns will be 0 — Neither cat nor dog.

• If the algorithm you are using (e.g. Random Forest) implicitly uses the categories then you may want to add an “Other” column that all new categories will be grouped into.

Training vs.Test Categories

New Categories in Test Set


Missing Data6 of 10



6. Missing Data

Types of Missing Data

• Unavailable: Valid for the observation, but not available in the data set.

• Removed: Observation quality threshold may have not been reached, and data removed

• Not applicable: measurement does not apply to the particular observation (e.g. number of tires on a boat observation)

• It depends! Some options:• Ignore entire observation.• Create an binary variable for each predictor to

indicate whether the data was missing or not• Segment model based on data availability.• Use alternative algorithm: decision trees accept

missing values; linear models typically do not.

What to Do


Outliers7 of 10



7. Outliers

Types of Outliers• Outliers can exist in response or predictors• Valid outliers: rare, extreme events • Invalid outliers: erroneous measurements

• Remove observations.• Apply a transformation to reduce impact: e.g. log or

bins.• Choose a loss function that is more robust: e.g. MAE

vs MSE.• Impose a constraint on data range (cap values).• Ask questions: Understand whether the values are

valid or invalid, to make the most appropriate choice.

What to Do

What Can Happen

• Outlier values can have a disproportionate weight on the model.

• MSE will focus on handling outlier observations more to reduce squared error.

• Boosting will spend considerable modeling effort fitting these observations.


Data Leakage8 of 10



8. Data Leakage

What Is It • Leakage is allowing your model to use information that will not be available in a production setting.

• Obvious example: using the Dow Jones daily gain/loss as part of a model to predict individual stock performance (even if that symbol is not part of the Dow)

• Model is overfit. • Will make predictions inconsistent with those you

scored when fitting the model (even with a validation set).

• Insights derived from the model will be incorrect.

• Understand the nature of your problem and data.• Scrutinize model feedback, such as relative influence

or linear coefficient.

What Happens

What to Do


Useless Models9 of 10



9. Useless Models

What is a “Useless” Model?

• Solving the Wrong Problem.• Not collecting appropriate data.• Not structuring data correctly to solve the problem.• Choosing a target/loss measure that does not

optimize the end use case: using accuracy to prioritize resources.

• Having a model that is not actionable.• Using a complicated model that is less accurate than

a simple model.

• Understand the problem statement.• Solving the wrong problem is an issue in all problem-

solving domains, but arguably easier with black box techniques common to ML

• Utilize post-processing measures• Create simple baseline models to understand lift of more

complex models• Plan on an iterative approach: start quickly, even if on

imperfect data• Question your models and attempt to understand them

What To Do


No Free Lunch10 of 10



10. No Free Lunch

No Such Thing as a Free Lunch

• No general purpose algorithm to solve all problems.• No right answer on optimal data preparation.• General heuristics are not always true:

• Tree models solve problems equivalently with any order-preserving transformation.

• Decision trees and neural networks will automatically find interactions.

• High number of predictors may be handled, but lead to a less optimal result than fewer key predictors.

• Models can not find relative information that span multiple observations.

• Model feedback can be misleading: relative influence, linear coefficients

• Understand how the underlying algorithms operate• Try several algorithms and observe relative performance

and the characteristics of your data• Feature engineering & feature selection• Interpret and react to model feedback

What To Do


Where to learn practical tips?

• WinVector Blog (Nina Zumel & John Mount): http://win-vector.com/blog

• Practical Data Science With R (book by Nina Zumel & John Mount): https://www.manning.com/books/practical-data-science-with-r

• Elements of Statistical Learning (book by Trevor Hastie, Robert Tibshirani & Jerome Friedman): http://statweb.stanford.edu/~tibs/ElemStatLearn

• Machine Learning Mastery Blog (Jason Brownlee): http://machinelearningmastery.com


Where to learn more about H2O?

• H2O Online Training (free): http://learn.h2o.ai• H2O Slidedecks: http://www.slideshare.net/0xdata• H2O Video Presentations: https://www.youtube.com/user/0xdata• H2O Community Events & Meetups: http://h2o.ai/events• Machine Learning & Data Science courses: http://coursebuffet.com

Customers ! Community ! Evangelists !

November 9, 10, 11 Computer History Museum

H 2 O W O R L D . H 2 O . A I

!

20% off registration using code:

h2ocommunity !


@ledell on Twitter, GitHub [email protected]

http://www.stat.berkeley.edu/~ledell

@Mark_a_Landry on Twitter [email protected]