With the advent of tools for automated model training (autoML), building predictive models is becoming easier, more accessible and faster than ever. Tools for R such as mlrMBO (Bischl et al., 2017), parsnip (Kuhn and Vaughan, 2019); tools for python such as TPOT (Olson et al., 2016), auto-sklearn (Feurer et al., 2015), autoKeras (Jin et al., 2018) or tools for other languages such as H2O Driverless AI (H2O.ai, 2019; Cook, 2016) and autoWeka (Kotthoff et al., 2017) supports fully- or semi-automated feature engineering and selection, model tuning and training of high-performing models.
Yet, model building is always preceded by a phase of understanding the problem, understanding of a domain and exploration of a data set. Usually, in the process of the data analysis much more time is spent on data preparation and exploration than on model tuning. This is why the current bottleneck in data analysis is in the EDA phase. Recently, a number of tools were developed to automate or speed up the part of the summarizing data and discovering patterns. Since the process of building predictive models automatically is referred to as autoML, we will dub the automation of data exploration autoEDA. The surge in interest in autoEDA tools111Access the raw data with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/52ec") is evident in the Figure 1. Table 1 describes the popularity of autoEDA tools measured as the number of downloads from CRAN and usage statistics from Github222Access the raw data with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/006d").
There is an abundance of R libraries that provide functions for both graphical and descriptive data exploration. Here, we restrict our attention to packages that aim to automatize or significantly speed up the process of exploratory data analysis for tabular data. Such tools usually work with full data frames, which are processed in an automatic or semi-automatic manner, for example by guessing data types and dropping variables that do not satisfy some criteria and return summary tables, groups of plots or full reports.
This paper has two main goals. Firstly, to characterize existing R tools for automated exploratory data analysis and their range of capabilities. To our best knowledge, this is first such a review. Previously, a smaller comparison of seven packages was done in Putatunda et al. (2019). Secondly, based on this summary, to identify areas, where automated data exploration could be improved. In particular, we are interested in gauging the potential of AI-assisted EDA tools.
The first goal is addressed in Sections 0.2 R packages for automated EDA and 0.3 Feature comparison where we first briefly describe each package and the compare, how are different EDA tasks are tackled by these packages. Then, in Section 0.4 Summary, we compile a list of strong and weak points of the automated EDA software and detail some open problems.
0.1.1 The tasks of Exploratory Data Analysis
The CRISP-DM standard (Wirth, 2000) lists the following phases of a data mining project:
Automated EDA tools aim to make the Data understanding phase as fast and as easy as possible. This part of a project can be further divided into smaller tasks. These include a description of a dataset, data exploration, and data quality verification. All these tasks can be achieved both by providing descriptive statistics and numerical summaries and by visual means. AutoEDA packages provide functions to deal with these challenges. Some of them are also concerned with simple feature engineering and data cleaning. Both these tasks belong in the Data preparation phase, which precedes and supports the model building phase. Let us notice that business understanding is affected by the data understanding, which makes this part of the analysis especially important.
Goals of autoEDA tools are summarized in Table 2. The Phase and Tasks columns are taken from the CRISP-DM standard, while Type and Examples columns provide examples based on current functionalities of autEDA packages.
|Data understanding||Data description||dimensions||variables number|
|meta-data||size in RAM|
|Data validity||invalid values||typos|
|missing values||NA count|
|multivaraite||Parallel coord. plot|
|Data preparation||Data cleaning||Imputation||Impute mean|
|Outlier treatment||Impute median|
|Derived attributes||Dimension reduction||PCA|
|Categorical||Merge rare factors|
Each task should be summarized in a report, which makes reporting another relevant problem of autoEDA. Uni- and bivariate data exploration is a part of the analysis that is most thoroughly covered by the existing autoEDA tools. The form of univariate summaries depends on the variable type. For numerical variables, most packages provide descriptive statistics such as centrality and dispersion measures. For categorical data, unique levels and associated counts are reported. Bivariate relationships descriptions display either dependency between one variable of interest and all other variables, which includes contingency tables, scatter plots, survival curves, plots of distribution by values of a variable (histograms, bar plots, box plots), or between all pairs of variables (correlation matrices and plots), or chosen pairs of variables.
0.2 R packages for automated EDA
In this section, twelve R libraries are shortly summarized. One of them is only available on GitHub (autoEDA), other packages are CRAN-based. The list is not exhaustive, but these are the most matured general-purpose packages. There are other libraries that are either restricted to one area of application, like the RBioPlot (Zhang and Storey, 2016) package for biomolecular data or ExPanDaR Gassen (2018) for panel data; designed for one specific task, for example creating tables, some of these packages are briefly discussed in Section 0.2.13; or not mature enough. The exact versions of packages that were used to create examples can be found in the reference. All examples are based on a subset of typical_data333Access the data with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/278c7") dataset from visdat package. Whenever possible, archivist (Biecek and Kosinski, 2017) hooks are provided for easy access to the presented objects. When a function call only gives side-effects, a link is provided to the full result (PDF/PNG files). Tables were prepared with the xtable package (Dahl et al., 2018).
0.2.1 The dataMaid package
The dataMaid (Petersen and Ekstrom, 2018) package has two central functions: the check function, which performs checks of data consistency and validity, and summarize, which summarizes each column. Another fmakeDataReport, which automatically creates a report in PDF, DOCX or HTML format. The goal is to detect missing and unusual - outlying or incorrectly encoded - values. The report contains whole dataset summary: variables and their types, number of missing values and if a problem was detected and univariate summaries in the form of descriptive statistics, histograms/bar plots and an indication of possible problems.
User-defined checks and summaries can be also included in the analysis. The vignette Extending dataMaid explains how to define them. It is also possible to customize the report, in particular, it can only present variables with identified issued. An example report444Find the full report at https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/dataMaid/dataMaid_report.pdf can be found in Figure 2.
0.2.2 The xray package
The xray (Seibelt, 2017) package has three functions for the analysis of data prior to statistical modeling:
detecting anomalies: missing data, zero values, blank strings, and infinite numbers (anomalies function),
drawing and printing univariate distributions of each variable through histograms, bar plots and quantile tables (distributions),
drawing plots of variables over time for a specified time variable (timebased).
Examples are presented in the readme file in the GitHub repository of the project (https://github.com/sicarul/xray), but no vignette is attached to it. Plots555Access the associated table with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/a3a3") generated by the package are presented in Figure 3.
0.2.3 The visdat package
The package visdat (Tierney, 2017) is maintained by rOpenSci. It consists of six functions that help visualize:
variables types and missing data (vis_dat function),
types of each value in each column (vis_guess),
clusters of missing values (vis_miss),
differences between the two datasets (vis_compare),
where given conditions are satisfied in the data (vis_expect),
correlation matrix for the numerical variables (vis_cor).
Each of these functions returns a single ggplot2 (Wickham, 2016) plot that shows a rectangular representation of the dataset where the expected information is denoted by colors. An example of this visualization666Access the plot object with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/3cfd") can be seen in Figure 4.
The package includes a vignette Using visdat that provides examples for all package options. Interestingly, it is the only packages that use solely visual means of exploring the data.
0.2.4 The dlookr package
The dlookr (Ryu, 2019)
package provides tools for 3 types of analysis: data diagnosis including correctness, missing values, and outlier detection, exploratory data analysis, and feature engineering: imputation, dichotomization, and transformation of continuous features. It can also automatically generate a PDF report for all these analyses.
For data diagnosis, types of variables are reported along with counts of missing values and unique values. Variables with a low proportion of unique values are described separately. All the typical descriptive statistics are provided for each variable. Outliers are detected and distributions of variables before and after outlier removal are plotted. Both missing values and outliers can be treated using impute_na and impute_outlier functions.
In the EDA report, descriptive statistics are presented along with normality tests, histograms of variables and their transformations that reduce skewness: logarithm and root square. Correlation plots are shown for numerical variables. If the target variable is specified, plots that show the relationship between the target and each predictor are also included.
Transformation report compares descriptive statistics and plots for each variable before and after imputation, skewness-removing transformation and binning. If the right transformation is found among the candidate transformations, it can be applied to the feature through one of the binning, binning_by, transform functions.
Every operation or summary presented in the reports can be also performed manually. A dedicated vignette explains each of the main functionalities (Data quality diagnosis, Data Transformation, Exploratory Data Analysis vignettes). An example777Access the full report at https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/dlookr/dlookr_eda.pdf taken from one of the reports can be found in Figure 5.
0.2.5 The DataExplorer package
DataExplorer (Cui, 2018) is a recent package that helps automatize EDA and simple feature engineering. It provides functions for:
whole dataset summary: dimensions, types of variables, missing values etc (introduce and plot_intro functions),
missing values visualization as missing fraction per column and analysis (plot_missing and profile_missing),
plotting distributions of variables, separately numerical and categorical (plot_histogram and plot_bar),
QQ Plots (plot_qq),
plotting correlation matrices (plot_correlation function),
visualizing PCA results by plotting percentage of explained variance and correlations with each original feature for every principal component (plot_prcomp),
plotting relationships between the target variable and predictors - scatterplots and boxplots (plot_scatterplot and plot_boxplot functions),
replacing missing values by a constant (set_missing),
grouping sparse categories (group_category),
creating dummy variables and dropping features (dummify, drop_features).
The create_report function generates a report. By default, it consists of all the above points except for feature engineering and it can be further customized. An introductory vignette Introduction to DataExplorer that showcases all the functionalities is included in the package. It is noticeable that the package almost entirely relies on visual techniques. Plots taken from an example report888Access the full report https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/DataExplorer/dataexplorer_example.pdf are presented in Figure 6.
0.2.6 The funModeling package
dataset summary (df_status function),
plots and descriptive statistics for categorical and numerical variables (plot_num, profiling_num and freq),
classical and information theory-based correlation analysis for target variable vs other variables - numerical in the first case, all in the second case (correlation_table and var_rank_info functions),
plots of distribution of target variables vs predictors (bar plots, box plots and histograms via cross_plot and plotar functions),
quantitative analysis for binary target variables (categ_analysis),
different methods of binning continuous features (discretize_df, convert_df_to_categoric and discretize_rgr),
variable scaling (range01),
outlier treatment (prep_outliers, tukey_outlier and hampel_outlier functions),
gain and lift curves (gain_lift).
It is the only library that encompasses visualizations related to predictive models and non-standard correlation analysis. The range of tools contained by funModeling is very wide. The package includes an exhaustive introduction vignette called funModeling quick-start. One of the bivariate visualizations999Find all the plots at https://github.com/mstaniak/autoEDA-resources/tree/master/autoEDA-paper/plots/funmodeling offered by the package can be found in Figure 7.
0.2.7 The autoEDA package
autoEDA package (Horn, 2018a) is a GitHub-based tool for univariate and bivariate visualizations and summaries. The dataOverview function returns a data frame that describes each feature by its type, number of missing values, outliers and typical descriptive statistics. Values proposed for imputation are also included. Two outlier detection methods are available: Tukey and percentile-based. A PDF report can be created using the autoEDA function. It consists of the plots of distributions of predictors grouped by outcome variable or distribution of outcome by predictors.
The package can be found on Xander Horn’s GitHub: https://github.com/XanderHorn/autoEDA. It does not include a vignette, but a short introduction article was published to LinkedIn (Horn, 2018b) and similar examples can be found in the readme of the project. Plots from a report101010Find the full report at https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/autoEDA/autoEDA_report.pdf generated by autoEDA are displayed in Figure 8.
0.2.8 The arsenal package
The arsenal package (Heinzen et al., 2019) is a set of four tools for data exploration:
table of descriptive statistics and p-values associated statistical tests grouped by levels of a target variable (the so-called Table 1), also for paired observation, for example longitudinal data, via functions tableby and paired, which is limited to comparisons at two times points,
comparison of two data frames that can detect shared variables (compare function),
frequency tables for categorical variables (freqlist function),
Results of each function can be saved to a short report using the write2 function. An example111111Access the table with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/d951") can be found in 9.
A separate vignette is available for each of the functions. arsenal is the most statistically-oriented package among reviewed libraries. It borrows heavily from SAS-style procedures used by the authors at the Mayo Clinic.
0.2.9 The SmartEDA package
The SmartEDA package (Kondapalli, 2018), is focused entirely on data exploration through graphics and descriptive statistics with no view on feature engineering. The range of tools it includes is wide:
dataset summary (ExpData function),
descriptive statistics that may include correlation with target variable and density or bar plots (ExpNumStat, ExpNumViz, ExpCatStat and ExpCatViz functions, all visualizations may include the target variable),
QQ plots (ExpOutQQ),
contingency tables (ExpCTable function),
information value and Weight of the Evidence coding (ExpWoETable, ExpInfoValue),
parallel coordinate plot for multivariate visualization (ExpParcoord).
Plotting functions return grids of ggplot2 object. The results can be written to a HTML report (ExpReport function). There are also additional functionalities dedicated to data.table objects from data.table package (Dowle and Srinivasan, 2019). An introductory vignette called Explore data using SmartEDA (Intro) is attached to the library. Another vignette Custom summary statistics describe cutomizing output tables. The package is also described in the Putatunda et al. (2019) paper. Examples121212A full report is available at https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/SmartEDA/smarteda_report_target.pdf can be found in Figure 10.
0.2.10 The summarytools package
The summarytools package (Comtois, 2019) builds tables with whole data or univariate summaries, frequency tables or cross-tabulations. In addition, the output can be formatted to be included in knitr(Xie, 2015) or plain documents, HTML files and shiny apps (Chang et al., 2018). The are four main functionalities:
whole dataset summary including variable types and a limited number of descriptive statistics, counts of unique values and missing values and univariate plots within the output table (dfSummary function),
descriptive statistics, including skewness and kurtosis, for numerical variables, possibly grouped by levels of a factor (descr, stby),
counts and proportions for levels of categorical features (freq),
contingency tables for pairs of categorical variables (ctab).
All results can be saved and displayed in different formats. The package includes a vignette titled Introduction to summarytools. An example of univariate summaries131313Access the R object with archivist::aread("mstaniak/autoEDA-resources/autoEDA-paper/9e12"). can be found in Figure 3.
0.2.11 The exploreR package
The exploreR package (Coates, 2016) takes a unique approach to data exploration compared to other packages. The analysis is based on linear regression. There are three functionalities:
fitting univariate regression model for each independent variable and summarizing the results in a table that consists of estimated parameters, p-values, and values (masslm function),
plotting target variable against each independent variable along with the fitted least squares line (massregplot),
feature standardization by scaling to the interval
or subtracting mean and dividing by standard deviation.
Regression plots can be saved to a PDF file. A vignette called The How and Why of Simple Tools explains all the functions and provides examples. One of the regression plots141414A PDF file with all the plots can be found at https://github.com/mstaniak/autoEDA-resources/blob/master/autoEDA-paper/plots/exploreR.pdf is presented in Figure 11.
0.2.12 The RtutoR package
The RtutoR package (Nair, 2018a) is a tool for automated reporting. There are three options for creating a report that contains univariate and bivariate data summaries:
plots can be created interactively in a shiny app (launch_plotter function),
the whole report can be generated from a shiny app that allows the user to tweak the report (gen_exploratory_report_app),
the report can be created by a direct call to the generate_exploratory_analysis_ppt function.
The report is saved in the PPTX format. Notably, this package can identify the top k relevant variables based on a chosen criterion, for example, information gain, and display only plots for these variables. An example report can be found in the GitHub repository of the package151515Find the report at https://github.com/anup50695/RtutoR/blob/master/titanic_exp_report_2.pptx. The package was introduced in an R-Bloggers blog post Nair (2018b).
0.2.13 Other packages
As mentioned before, there are numerous R packages that aim to make data exploration faster or make the outputs more polished.
For table summaries of data that often include statistical tests, there are a few packages worth mentioning. The package tableone (Yoshida and Bohn., 2018) provides a CreateTableOne function to make publication-ready tables referred to as Table 1 - traditional name of tables that describe patients’ characteristic, usually stratified and including p-values from significance tests. The describe function from describer package (Hendricks, 2015)
prints a summary of a data.frame or a vector which includes data types, counts and descriptive statistics. A function of the same name from prettyR(Lemon and Grosjean, 2018) returns descriptive statistics for each column in a data.frame. This package is focused on improving the aesthetics of R statistical outputs. Similarly, the package Hmisc (Harrell Jr et al., 2019) includes a describe function that displays typical descriptive statistics and number of unique and missing values for each column. The plot method called on the result of the describe function returns a dot plot for each categorical and a spike histogram for each continuous column. The scope of this package is bigger than just Exploratory Data Analysis, as it includes many tools related to regression models.
There are also many packages related to data visualization. Two of them are particularly worth mentioning. The ggfortify package(Tang et al., 2016) serves as a uniform interface to plots of different statistical objects, including PCA results that can be used for data exploration and time series plots. The autoplotly library (Tang, 2018) was built on top of ggfortify to provide automatically generated, interactive visualizations of many statistical models. While these two packages are focused on statistical modeling, they can be helpful in exploratory analysis and exemplify the potential of quick and interactive visualization in R.
Two more packages are relevant to our interest. gpairs (Emerson and Green, 2014) and GGally (Schloerke et al., 2018) packages implement the generalized pairs plot (Emerson et al., 2013). This type of plot extends well known scatter plot matrices, that visualize bivariate relationships for many variables, by handling both numerical and categorical variables. It is helpful in data exploration and shares similarities to walls of histograms that can be found in automated EDA libraries.
0.3 Feature comparison
|Bar plots by target||x||x||x||x||x||x|
|Histograms by target||x||x||x||x|
|Parallel coord. Plot||x|
0.3.1 Data description
Almost all packages contain functions for summarizing the dataset. Tools that support data validity analysis are less common.
Whole dataset summaries
visdat package introduces the most original summaries of full dataset. The drawback of this approach is that it is not well suited for high dimensional data. But for a smaller number of variables, it gives a good overview of the dataset.
Most packages that provide a whole dataset summary take a similar approach and present names and types of variables, number of missing values and sometimes unique values or other statistics. This is true for summarytools (dfSummary function), autoEDA (dataOverview function), dataMaid (makeDataReport result), funModeling (df_status function) and DataExplorer (introduce function), which provides the information separately on two plots - one for dataset structure, one for missing data. In the dlookr package, summaries for numerical variables and categorical variables are only presented separately in the report (describe function).
Some packages can perform automated checks for the data, including at least outlier detection. The dataMaid package’s main purpose is to find inconsistencies and errors in the data. It finds possible outliers, missing values, low-frequency and possibly miscoded levels of factors. All this information can be summarized in a quality report. The dlookr package covers similar functionality. There are two main differences, the report does not describe possible miscoded factors, but outlier analysis is supplemented with plots showing variable distribution before and after removing the outliers. In all cases, the analysis is rather simple, for example in zero-inflated variables non-zero values are treated as outliers (dlookr). Other packages only provide information about the number of missing values, outlier and identify columns that consist of a single value.
0.3.2 Data exploration
While multivariate analysis is rarely supported, there are many tools for descriptive and graphical exploration uni- and bivariate patterns in the data.
All the tools that support univariate analysis take a similar approach to univariate analysis. For categorical variables, counts are reported and bar plots are presented, while histogram or boxplots and typical descriptive statistics (including quantiles, sometimes skewness) are used for continuous variables.
In dataMaid and dlookr packages, these plots are presented variable-by-variable in the report. In other packages (DataExplorer, funModeling, SmartEDA) groups of plots are shown together - as a wall of histograms or bar plots. Notably, dlookr reports skewness of variables and in case a skewed variable is found, it shows the distribution after some candidate transformations to reduce the skewness. This package also reports normality. The SmartEDA package also reports skewness and displays QQ plots against normal distribution, but it does not provide any means of reducing skewness.
The funModeling and SmartEDA packages only support calculating correlations between variables and a specified target. DataExplorer, visdat and DataExplorer packages can plot correlation matrices. They differ in categorical variables treatment. Some packages require only numerical features (visdat). Interestingly, in DataExplorer161616Access the plot with archivist::ared("mstaniak/autoEDA-resources/autoEDA-paper/0526"), low-cardinality categorical features are converted to 0-1 variables and plotted alongside numerical variables, as seen in Figure 12.
The arsenal package only presents variable summaries by levels of a chosen categorical variable. The report from autoEDA package consists of a limited number of bar plots/boxplots with target variable as one of the dimensions. Similarly, in DataExplorer, dlookr, funModeling and SmartEDA, scatter plots and box plots or histograms with a specified target variable on one of the axis can be plotted. Additionally, funModeling and dlookr draw histograms/densities of continuous features by the target. The funModeling package also has unique options: drawing bar plots of discretized variables by the target and quantitative analysis for binary outcome based on representativeness and accuracy. arsenal, summarytools and SmartEDA also feature contingency tables. Moreover, the exploreR package uses linear regression plots and statistics to find relationships between the target and other variables.
0.3.3 Data cleaning and feature engineering
The dataMaid package assumes that every decision regarding the data should be made by the analyst and does not provide any tools for data manipulation after diagnosis. Most of the packages only provide exploration tools. Exceptions are dlookr, funModeling, DataExplorer and exploreR. DataExplorer tools are limited to normalization, imputation by a constant, merging levels of factors and creating dummy variables.
The dlookr package can create a report that presents different possible transformations of features. Missing values can be imputed by mean/median/mode and distributions of variables before and after the procedure compared. The same is done for imputation of outliers. Logarithmic and root square transforms are proposed for skewed variables. Different methods of binning continuous variables are also presented, including Weight of the Evidence.
The funModeling package can perform discretization of a variable using an equal frequency criterion or gain ratio maximization. It can also scale variables to the interval . Outliers can be treated using the Tukey or Hampel method.
DataExplorer, dlookr, dataMaid, SmartEDA and RtutoR have an option of generating a report and saving it to a file. They consist of all or most possible outputs of the package which are organized either by variable (dataMaid, dlookr) or by type of variable (DataExplorer, SmartEDA) and the task. autoEDA package generates a minimal report with bivariate plots. Packages arsenal, funModeling, xray, summarytools and exploreR have an option of saving outputs - plots or tables - to files.
Automated EDA can be either directed towards a general understanding of a particular dataset or be more model-oriented, serving as a foundation for good modeling. While presented packages include some tools related to simple feature engineering, they are more focused on data understanding. For this task, they have many advantages. In this section, we summarize the strong points of existing tools and point out some possible improvements and new directions for autoEDA.
The packages dlookr, dataMaid, DataExplorer, SmartEDA are capable of creating good quality reports.
DataExplorer has very good visualizations for PCA.
DataExplorer handles categorical variables on correlation plots by creating dummy features, which is a unique idea compared to other packages.
The visdat package, while probably not the best choice for high dimensional data, features interesting take on initial whole dataset exploration.
The dlookr package is capable of selecting skewed variables and proposing transformations. Some of the other packages display binned continuous variables, which can also help in seeing visualizing dependencies.
dataMaid is a good tool for finding problems in the data. Thanks to the structure of check and summarize functions results, discovered issues can be treated effectively.
For datasets with a moderate number of features, DataExplorer, funModeling, dlookr and SmartEDA give a reasonable insight into variables distributions and simple relationships.
SmartEDA package provides a method of visualizing multivariate relationships - parallel coordinate plot.
The exploreR package provides usefuls tool for assessing bivariate relationship through linear regression.
We can see that tasks related to data quality and whole dataset summary are well by the existing libraries. Getting the big picture of the data and finding possible data quality problems is easy, especially due to the dataMaid package. Univariate exploration For classical applications, for example, statistical analyses in medicine, the current tools provide very good tables, such as the ones from tableone package, and uni-/bivariate plots. Univariate analysis can be performed either variable-after-variable (dlookr, dataMaid), where we can see the statistical properties of each variable, or as groups of plots based on variable type (DataExplorer, funModeling). Both ways can be useful for a reasonable number of predictors. While multivariate tools are scarce, the available tools, PCA in DataExplorer and PCP in SmartEDA, are very well done.
All the presented tools are likely to fail in typical situations with imperfect data. In particular, they are usually not robust to issues like zero-variance/constant variables (DataExplorer can’t generate a report in this case). Error messages in some cases not uninformative.
In some situations, they lack flexibility. For example, in DataExplorer arguments can be passed to cor function, but not to corrplot function.
In case of walls of histograms (or bar plots), no selection is being done and no specific order is chosen to promote most interesting distributions. The same is true for automatically created reports. Moreover, for high-dimensional data or high-cardinality factors, the plots are unreadable or impractical. This is especially true for DataExplorer and funModeling functions (e.g. cross_plot), even though DataExplorer removes too large factors from the panels. This problem is only addressed by the RtutoR package, which allows to select top k relevant variables.
Plots are limited to bivariate relationships. Exploring higher dimensional dependencies would be interesting, for example by adding color and size dimensions to the plots. Since such an addition would result in a large number of new plots, it would require a proper method of finding the most relevant visualization.
Support for time-varying variables and non-classical (not IID) problems such as survival analysis is limited or non-existent. For survival analysis, the automation level is low, but there are two notable tools for summarizing dependencies. First is the recognized package survminer (Kassambara and Kosinski, 2018), which helps visualize survival curves, while also displaying survival tables and other information. The other tool is a no longer developed cr17 package (Młynarczyk and Biecek, 2017), which includes summarizeCR function that returns several tables and plots for competing risks analysis. More tools for fast visualization of at least bivariate relationships in such problems would be a big help for analysts.
PCA, parallel coordinate plots and model summaries are supported, but each by a single, separate package. It is evident that there is a shortage of multivariate tools. Cluster analysis is not available in any of the packages.
Automated reports could be enriched by textual annotations and descriptions, either built from simple templates or from a generative model.
Only one of the packages addresses the issue of skewed variables. Proposing transformations of continuous features other than binning would be helpful and could improve visualizations, for example, scatter plots with skewed variables.
Univariate regression models can be plotted by the exploreR package. Exploration based on simple statistical models, for example, scatter plot smoothing, is not an option in any of the other packages. Using regression models and feature transformations to identify and measure relevant relationships could improve bivariate analyses supported by automated EDA.
None of the packages addresses issues such as multicollinearity.
Missing data imputation more advanced than imputing a constant is delegated to other packages, although, it is known that imputation by a constant is usually not the best method of missing values treatment.
Some of the above issues limit the packages’ usefulness in iterative work. Though, the comparisons of transform and original features and the possibility of applying discovered transformations to data in dlookr package are steps in the right direction.
The tools available in R have similar range to other languages’ libraries, for example from Python. Python packages such as Dora (Epstein, 2017) or lens (Zabalza and Engineers, 2018) also cover feature-by-feature descriptive statistics and plots, bivariate visualizations of the relationships between predictors and target variable, contingency tables, basic feature engineering, and imputation.
Since EDA is both closely connected to feature engineering and based on visual insights, automated EDA can draw from existing tools for automated feature extraction (in autoML tools like TPOT(Olson et al., 2016)) and visualization recommendations. When it comes to aiding visual exploration of a dataset, standalone software carries possibilities beyond what we can expect from R packages or analogous libraries in other languages. A recent notable example is DIVE (Hu et al., 2018). It is an example of a growing number of tools for visual data exploration that aim to distinguish between relevant and irrelevant visualization and help the analyst find the most interesting plots. DIVE is one of the mixed-initiative visualization systems, meaning it uses both statistical properties of the dataset and user interactions to find the relevant plots. Building recommendation systems into autoEDA tools can help address the issue of dealing with high-dimensional data and multivariate dependencies by letting the ML-based system deal with the complexity of a large number of candidate visualizations. AI-assisted data exploration can be even faster and more efficient.
This work was financially supported by the NCN Opus grant 2016/21/B/ST6/02176.
- Biecek and Kosinski (2017) P. Biecek and M. Kosinski. archivist: An R package for managing, recording and restoring data analysis results. Journal of Statistical Software, 82(11):1–28, 2017. doi: 10.18637/jss.v082.i11.
- Bischl et al. (2017) B. Bischl, J. Richter, J. Bossek, D. Horn, J. Thomas, and M. Lang. mlrMBO: A Modular Framework for Model-Based Optimization of Expensive Black-Box Functions, 2017.
- Casas (2018) P. Casas. Data Science Live Book. https://livebook.datascienceheroes.com/, 2018. Retrieved on 14 March 2019.
- Casas (2019) P. Casas. funModeling: Exploratory Data Analysis and Data Preparation Tool-Box Book, 2019. URL https://CRAN.R-project.org/package=funModeling. R package version 1.7.
- Chang et al. (2018) W. Chang, J. Cheng, J. Allaire, Y. Xie, and J. McPherson. shiny: Web Application Framework for R, 2018. URL https://CRAN.R-project.org/package=shiny. R package version 1.2.0.
- Coates (2016) M. Coates. exploreR: Tools for Quickly Exploring Data, 2016. URL https://CRAN.R-project.org/package=exploreR. R package version 0.1.
- Comtois (2019) D. Comtois. summarytools: Tools to Quickly and Neatly Summarize Data, 2019. URL https://CRAN.R-project.org/package=summarytools. R package version 0.9.2.
- Cook (2016) D. Cook. O’Reilly Media, 2016.
- Csardi (2015) G. Csardi. cranlogs: Download Logs from the ’RStudio’ ’CRAN’ Mirror, 2015. URL https://CRAN.R-project.org/package=cranlogs. R package version 2.1.0.
- Cui (2018) B. Cui. DataExplorer: Data Explorer, 2018. URL https://CRAN.R-project.org/package=DataExplorer. R package version 0.7.0.
- Dahl et al. (2018) D. B. Dahl, D. Scott, C. Roosen, A. Magnusson, and J. Swinton. xtable: Export Tables to LaTeX or HTML, 2018. URL https://CRAN.R-project.org/package=xtable. R package version 1.8-3.
- Dowle and Srinivasan (2019) M. Dowle and A. Srinivasan. data.table: Extension of ‘data.frame‘, 2019. URL https://CRAN.R-project.org/package=data.table. R package version 1.12.0.
- Emerson and Green (2014) J. W. Emerson and W. A. Green. gpairs: gpairs: The Generalized Pairs Plot, 2014. URL https://CRAN.R-project.org/package=gpairs. R package version 1.2.
- Emerson et al. (2013) J. W. Emerson, W. A. Green, B. Schloerke, J. Crowley, D. Cook, H. Hofmann, and H. Wickham. The generalized pairs plot. Journal of Computational and Graphical Statistics, 22(1):79–91, 2013. doi: 10.1080/10618600.2012.694762.
- Epstein (2017) N. Epstein. Dora: Exploratory data analysis toolkit for python, 2017. URL https://github.com/NathanEpstein/Dora. Python library version 0.0.2.
- Feurer et al. (2015) M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter. Efficient and robust automated machine learning. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2962–2970. Curran Associates, Inc., 2015.
- Gassen (2018) J. Gassen. ExPanDaR: Explore Panel Data Interactively, 2018. URL https://CRAN.R-project.org/package=ExPanDaR. R package version 0.3.0.
- H2O.ai (2019) H2O.ai. H2O, march 2019. version 184.108.40.206.
- Harrell Jr et al. (2019) F. E. Harrell Jr, with contributions from Charles Dupont, and many others. Hmisc: Harrell Miscellaneous, 2019. URL https://CRAN.R-project.org/package=Hmisc. R package version 4.2-0.
- Heinzen et al. (2019) E. Heinzen, J. Sinnwell, E. Atkinson, T. Gunderson, and G. Dougherty. arsenal: An Arsenal of ’R’ Functions for Large-Scale Statistical Summaries, 2019. URL https://CRAN.R-project.org/package=arsenal. R package version 2.0.0.
- Hendricks (2015) P. Hendricks. describer: Describe Data in R Using Common Descriptive Statistics, 2015. URL https://CRAN.R-project.org/package=describer. R package version 0.2.0.
- Horn (2018a) X. Horn. autoEDA: Automated univariate and bivariate exploratory data analysis, 2018a. R package version 1.0.
- Horn (2018b) X. Horn. Automated exploratory data analysis in r. https://www.linkedin.com/pulse/automated-exploratory-data-analysis-r-xander-horn/, 2018b. Retrieved on 14 March 2019.
- Hu et al. (2018) K. Hu, D. Orghian, and C. Hidalgo. Dive: A mixed-initiative system supporting integrated data exploration workflows. In ACM SIGMOD Workshop on Human-In-the-Loop Data Analytics (HILDA). ACM, 2018. doi: 10.1145/3209900.3209910.
Jin et al. (2018)
H. Jin, Q. Song, and X. Hu.
Auto-Keras: Efficient Neural Architecture Search with Network Morphism.arXiv e-prints, Jun 2018.
- Kassambara and Kosinski (2018) A. Kassambara and M. Kosinski. survminer: Drawing Survival Curves using ’ggplot2’, 2018. URL https://CRAN.R-project.org/package=survminer. R package version 0.4.3.
- Kondapalli (2018) D. U. K. R. R. P. Kondapalli. SmartEDA: Summarize and Explore the Data, 2018. URL https://CRAN.R-project.org/package=SmartEDA. R package version 0.3.0.
Kotthoff et al. (2017)
L. Kotthoff, C. Thornton, H. H. Hoos, F. Hutter, and K. Leyton-Brown.
Auto-weka 2.0: Automatic model selection and hyperparameter optimization in weka.Journal of Machine Learning Research, 18(1):826–830, Jan. 2017.
- Kuhn and Vaughan (2019) M. Kuhn and D. Vaughan. parsnip: A Common API to Modeling and Analysis Functions, 2019. URL https://CRAN.R-project.org/package=parsnip. R package version 0.0.2.
- Lemon and Grosjean (2018) J. Lemon and P. Grosjean. prettyR: Pretty Descriptive Stats, 2018. URL https://CRAN.R-project.org/package=prettyR. R package version 2.2-2.
- Młynarczyk and Biecek (2017) M. Młynarczyk and P. Biecek. cr17: Testing Differences Between Competing Risks Models and Their Visualisations, 2017. URL https://CRAN.R-project.org/package=cr17. R package version 0.1.0.
- Nair (2018a) A. Nair. RtutoR: Shiny Apps for Plotting and Exploratory Analysis, 2018a. URL https://CRAN.R-project.org/package=RtutoR. R package version 1.2.
- Nair (2018b) A. Nair. Automating basic eda. https://www.r-bloggers.com/automating-basic-eda/, 2018b. Retrieved on 25 March 2019.
Olson et al. (2016)
R. S. Olson, R. J. Urbanowicz, P. C. Andrews, N. A. Lavender, L. C. Kidd, and
J. H. Moore.
Applications of Evolutionary Computation: 19th European Conference, EvoApplications 2016, Porto, Portugal, March 30 – April 1, 2016, Proceedings, Part I, chapter Automating Biomedical Data Science Through Tree-Based Pipeline Optimization, pages 123–137. Springer International Publishing, 2016. doi: 10.1007/978-3-319-31204-0_9.
- Petersen and Ekstrom (2018) A. H. Petersen and C. T. Ekstrom. dataMaid: A Suite of Checks for Identification of Potential Errors in a Data Frame as Part of the Data Screening Process, 2018. URL https://CRAN.R-project.org/package=dataMaid. R package version 1.2.0.
- Putatunda et al. (2019) S. Putatunda, K. Rama, D. Ubrangala, and R. Kondapalli. SmartEDA: An R Package for Automated Exploratory Data Analysis. arXiv e-prints, art. arXiv:1903.04754, Mar 2019.
- Ryu (2019) C. Ryu. dlookr: Tools for Data Diagnosis, Exploration, Transformation, 2019. URL https://CRAN.R-project.org/package=dlookr. R package version 0.3.8.
- Schloerke et al. (2018) B. Schloerke, J. Crowley, D. Cook, F. Briatte, M. Marbach, E. Thoen, A. Elberg, and J. Larmarange. GGally: Extension to ’ggplot2’, 2018. URL https://CRAN.R-project.org/package=GGally. R package version 1.4.0.
- Seibelt (2017) P. Seibelt. xray: X Ray Vision on your Datasets, 2017. URL https://CRAN.R-project.org/package=xray. R package version 0.2.
- Tang (2018) Y. Tang. autoplotly: An r package for automatic generation of interactive visualizations for statistical results. Journal of Open Source Software, 3, 2018. doi: 10.21105/joss.00657.
- Tang et al. (2016) Y. Tang, M. Horikoshi, and W. Li. ggfortify: Unified Interface to Visualize Statistical Results of Popular R Packages. The R Journal, 8(2):474–485, 2016. doi: 10.32614/RJ-2016-060.
- Tierney (2017) N. Tierney. visdat: Visualising whole data frames. JOSS, 2(16):355, 2017. doi: 10.21105/joss.00355.
- Wickham (2016) H. Wickham. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York, 2016. ISBN 978-3-319-24277-4. doi: 10.1007/978-0-387-98141-3. URL http://ggplot2.org.
- Wirth (2000) R. Wirth. Crisp-dm: Towards a standard process model for data mining. In Proceedings of the Fourth International Conference on the Practical Application of Knowledge Discovery and Data Mining, pages 29–39, 2000. doi: 10.1.1.198.5133.
- Xie (2015) Y. Xie. Dynamic Documents with R and knitr. Chapman and Hall/CRC, Boca Raton, Florida, 2nd edition, 2015. doi: 10.18637/jss.v056.b02.
- Yoshida and Bohn. (2018) K. Yoshida and J. Bohn. tableone: Create ’Table 1’ to Describe Baseline Characteristics, 2018. URL https://github.com/kaz-yos/tableone. R package version 0.9.3.
- Zabalza and Engineers (2018) V. Zabalza and F. Engineers. lens, Dec. 2018. Python library version 0.4.5.
- Zhang and Storey (2016) J. J. Zhang and K. B. Storey. Rbioplot: An easy-to-use r pipeline for automated statistical analysis and data visualization in molecular biology and biochemistry. PeerJ, 2016(9), Jan. 2016. doi: 10.7717/peerj.2436.