Last Updated on August 15, 2020

You often jump from problem-to-problem in applied machine learning and you need to get up to speed on a new dataset, fast.

A classical and under-utilised approach that you can use to quickly build a relationship with a new data problem is **Exploratory Data Analysis**.

In this post you will discover Exploratory Data Analysis (EDA), the techniques and tactics that you can use and why you should be performing EDA on your next problem.

## Build a Relationship with the Data

The process of classical statistics is to test hypotheses already held about the problem.

This is done by fitting specific models and demonstrating specific relationships in the data. It’s an effective approach, but it assumes you already have hypotheses about the problem, that you already understand the data. This is rarely the case in applied machine learning.

Before you can model the data and test your hypotheses, you need to build a relationship with the data. You can build this relationship by spending time **summarizing**, **plotting** and **reviewing** actual real data from the domain.

This approach of analysis before modeling is called Exploratory Data Analysis.

In spending time with the data up-front you can build an intuition with the data formats, values, and relationships that can help to explain observations and modeling outcomes later.

It is called exploratory data analysis because you are **exploring your understanding of the data**, building an intuition for how the underlying process that generated it works and provoking questions and ideas that you can use as the basis for your modeling.

The process can be used to sanity check the data, to identify outliers and come up with specific strategies for handling them. In spending time with the data, you can spot corruption in the values that may signal a fault in the data logging process.

## Origin of Exploratory Data Analysis

Exploratory Data Analysis was developed by John Tukey at Bell Labs as a way of systematically using the tools of statistics on a problem before a hypotheses about the data wereÂ developed. It is an alternative or opposite approach to “*confirmatory data analysis*“.

The seminal description of the process was in Tukey’s 1977 book Exploratory Data Analysis.

The objective is to understand the problem in order to generate testable hypotheses. As such, the outcomes likeÂ the graphs and summary statistics are only for youÂ to improve *your*Â understanding, not to demonstrate a relationship in the data to a general audience. This gives the agile flavor to the process.

The S language was developed in the same laboratory and was used as the tool for EDA. The use of scripts to generate data summaries and views is a natural and intentional fit for the process.

Wikipedia provides a nice short list of the objectives of EDA:

- Suggest hypotheses about the causes of observed phenomena
- Assess assumptions on which statistical inference will be based
- Support the selection of appropriate statistical tools and techniques
- Provide a basis for further data collection through surveys or experiments

## Techniques for Exploratory Data Analysis

Exploratory Data Analysis is often performed with a representative sample of the data. You do not need to use all data available nor big data infrastructure.

Spend time with the raw data.

Starting with eyeballing tables of numbers is wise. Skimming through tables can quickly highlight the form of each data attribute, obvious perversions and large outliners in the values and start to suggest candidate relationships to explore between attributes. Take notes.

Simple univariate and multivariate methods that give a view on the data can be used.

For example, five methods that I would consider must have are:

- Five number summaries (mean/median, min, max, q1, q3)
- Histogram graphs
- Line Charts
- Box and Whisker plots
- Pairwise Scatterplots (scatterplot matrices)

In addition to summaries, also look at transforms of the data and re-scalings of the data. Flush out interesting structures that you can describe.

Take notes. Take lots of notes.

Ask lots of questions of the data, for example:

- What values do you see?
- What distributions do you see?
- What relationships do you see?
- What relationships do you think might benefit the prediction problem?
- What ideas about the domain does the data spark?
- and on…

## Focus on Understanding

You are not creating a report, you are trying to understand the problem.

The results are ultimately throw-away, and all that you should be left with is a greater understanding and intuition for the data and a long list of hypotheses to explore whenÂ modeling.

The code does not need to be beautiful (but they need to be correct). Use reproducible scripts and standard packages.

You do not need to dive into advanced statistical methods or plots. Keep it simple and spend time with the data.

A query interface like SQL can help you play a lot of what-if scenarios very quickly with a sample of your data.

The models will only be as good as the questions and understanding you have of the data and the problem.

## Resources

The book Doing Data Science: Straight Talk from the Frontline has a short section on EDA and provides a great reading list for more information:

- Exploratory Data Analysis
- The Visual Display of Quantitative Information (highly recommended)
- The Elements of Graphing Data
- Statistical Graphics for Visualizing Multivariate Data

TryÂ exploratory data analysis on your current or next project.

If you already do it, try some methods you have not used beforeÂ or try to beÂ systematic, even sketch out a checklist of things to look at to cover your bases as a first pass on the data.

I love this post, as per usual. As I read I found myself asking “are there any heuristics for interpretation?” For instance, you mentioned that EDA can help suggest hypotheses or assess assumptions for potential inference methods–can we get some very clear examples?

I wish I knew all the rules-of-thumb that great data scientists already follow, esp. re: EDA. I can visualize data very well and apply all the techniques described, but I want to know what to do or what has been done with some of the output from data exploration. Do you have any examples from your experience?

Thanks so much!

Great suggestion, thanks.

EDA is super specific to each problem, so I cannot give general heuristics. Perhaps picking up a good book on the topic might help?

https://amzn.to/2I1S4J2

Is it critical to NOT perform EDA on the test or holdout set ? That is to only run EDA on the train set ? I’ve heard some folks say using EDA on the test or holdout can cause leakage, misleading results,etc…

Yes, ideally you would focus on the training dataset.

Dear Dr Jason,

Just to let you know, that in the 1970s, there was an Australian working at Princeton with John W Tukey, Prof. Don McNeil BSc (Hons) (Tasmania), PhD (ANU) on the topic of interactive data analysis. He published a book “Interactive Data Analysis”, 1977, reference https://catalogue.nla.gov.au/Record/656675.

He is a professor emeritus at Macquarie, https://researchers.mq.edu.au/en/persons/don-mcneil.

Thank you,

Anthony of Sydney