Quickstart Guide

Welcome to EDAmame, a first-pass hands-on appetizer to have a taste of your data. Like its sushi-train counterpart, EDAmame follows a smooth steady workflow that's easy to pick up.

The provided TestData1 is a reduced copy of the Wisconsin Diagnostic Breast Cancer (WDBC) public dataset. The provided TestData2 is a modified version of the Crohn’s Disease Adverse Events dataset. Look for boxes like these for instructions on analysing these datasets.

A more in-depth explanation of EDAmame can be found in the repository.

Ingesting Data

EDAmame starts on the In tab. Users can import tabular data from various sources (external files, copy-paste through the browser, an existing googlesheet, or via a remote URL).

To ingest data, please click on External file, then Browse... to upload a file. EDAmame can import various file types, but we recommend CSVs / TSVs. Press Import data to continue.

The grid_on Options Menu can be used to update the analysis in EDAmame in real-time. It can be opened by clicking on the grid_on symbol in the top-right.

Import the TestData2 dataset as described above. You will see the data has 117 rows and 9 columns.

Once the Import into EDAmame dialog box is closed, the Update & select variables screen is visible to edit columns before analysis.

Once columns are ready, click on Apply changes at the bottom. Look out for warnings below indicating potential issues with assigned classes - this can happen when strings such as "none" appear in numerical columns - you may need to update these columns to the correct class. For now, everything looks OK.

You'll notice the grid_on Options Menu on the right side is open on the Select tab. This data doesn't have a variable we can use as the ID column, so we'll check the checkbox for use row index instead, updating the ID selection.

Now we can continue to the Data tab, where more cleaning can be applied.

While in the Data tab, click on Detect Missing Value Strings. We can see that a few non-standard values for missing data are detected ("unknown", "NULL", "N/A"). Now click on Standardize Selected to NA to convert them all to true missing values.

In the Missing Data Filters section, we can remove columns or rows with too many missing values. The default thresholds are reasonable, so go ahead and click Apply Filters. We can see that only 2 rows were removed.

Data Quality

The green tabs provide a general overview of the input data. The Miss tab presents a barchart of missing data and allows you to perform imputation with adjustible parameters. The Cat and Num tabs present summary plots of categorical and numerical features, respectively. Finally, the Dim tab presents a dimensionality-reduced plot of the data as either PCA or t-SNE.

While in the Miss tab, we can see columns "Sex" and "Bmi" are missing around 7% and 15% of their values, respectively. Let's try to impute these by clicking on Run Imputation on the right. We can see the missing values have been replaced with best estimates.

Now let's move on to the Cat tab; here we are prompted to check that we have the correct columns selected for our ID and target variables - we don't want the target to be "Country" so we'll click on No, change and update this to "Treat" in grid_on Options Menu:Select tab.

Moving on to the Num tab, we can see from the histogram of "Time", all values are the same - this isn't useful to us so we can drop this variable by going to grid_on Options Menu:Filter tab and selecting "Time" in the Drop features menu.

The TestData2 is not very interesting from here on, but it was good practice for cleanup. Let's now go back to the In tab and import TestData1.

This dataset is already clean, so we can go ahead and click through to the Dim tab without changing anything. Just make sure ID is set to column "Id" and Target (prediction) is set to column "Diagnosis".

In the grid_on Options Menu:Filter tab, select "M" as the Target label and "B" as the Control/background label (M = Malignant, B = Benign).

While in the Dim tab, we can see a benign diagnosis and a malignant diagnosis clearly separate into two distinct clusters.

Linear Model Analysis

The blue tabs present relationships between numeric features. The Cor tab is a correlation matrix between all numeric features, while the Pair tab uses up to 8 numeric features to display richer information than in the first tab.

The Inter tab highlights any significant 2- or 3-way interactions in the dataset, including summary statistics of linear modelling.

In the Cor tab, when ordering by "AOE", the bottom right-most cells are highly correlated (Area1, Radius1, and Perimeter1).

In the Pair tab, Also notice that Radius1 and Radius3 are significantly correlated (update selection in grid_on Options Menu:Filter:Paired corr).

In the Inter tab, notice that several two-way interactions were found to be significant, including "Area1:Texture3" and "Area2:Symmetry3". If we switch to three-way interaction, we see that "Perimeter3:ConcavePoints3:Radius3" is found to be most significant.

Machine-learning Analysis

The pink tabs present ways to model and explain the dataset. The Sel tab uses the Boruta feature selection to show which features will be useful. Use the Drop features option in the grid_on Options Menu:Filter tab to drop features manually.

Training of binary classification models is performed in the Pred tab using the target and control/background labels of the target feature. A number of modelling methods can be selected in the grid_on Options Menu:Model tab, including hyperparameter tuning for advanced users. Multiple models can be run concurrently.

The Perf tab evaluates selected models using 5-fold cross-validation, displaying a range of performance metrics.

The Expl tab shows how each feature contributes to predictions of the chosen model selected in grid_on Options Menu:Model:Explain model.

Finally, the modelS tab presents an interactive dashboard where the selected model can be further interrogted with multiple explainers.

In the Sel, all but four features (yellow and red boxplots) are shown to be important using default settings.

In the Pred tab, all three default selected models appear to achieve decent discrimination of malignant and benign using default hyperparameters.

In the Perf tab, all three models produce good ROC curves (accuracies > 95%), with both Random Forest and XGBoost showing very similar performance across most metrics.

In the Expl tab, "Area3", "ConcavePoints3", "Radius3" and "Perimeter3" stand out as the most important features for the Random Forest model.

In the modelS tab, four panels are available to dislay from a list of interactive explainer plots. For example, selecting "Break Down" and "Shapley Values" for two panels, it can be seen that for a single test datapoint ("id: 99" in the dropdown menu), the Random Forest prediction of 1 for Malignant was influenced mainly by "Area3", "Perimeter3" and "Radius3" according to both methods.

While exploring these tabs, notice there are info buttons that provide overviews of analysis methods and interpretation. For reproducible workflows, an Export Workflow Log button next to grid_on Options Menu can be used at anytime to save a text file of all analysis steps and parameters used during the session.