Transform Basics

Your data transformation is complete when you have done the following:

The Transformer page experience is based on the design principles of predictive transformation. Predictive transformation allows you as the user to explore your data and guides you to possible next steps based on your explorations. Predictive transformation works in the following phases:

If all looks good, you add the transformation. The change is added to your recipe, which is the sequence of steps to transform your data.

Then, you repeat the above steps for the next data of interest. You can see this process in action below.

1 - Visualize

Below, you can see the Transformer page displaying a sample from the USDA-recognized Farmer's Markets in the United States as of the year 2014.

Figure: Visualize: Farmer's Market Data

The source of this data is an Avro file, which in its raw form does not look like this nice tabular representation. When a dataset is loaded into the Transformer page, the Trifacta Application attempts to represent it in tabular form, so that it is easy to navigate and analyze.

A few key things to notice:

1. Initial Data: This button indicates the current sample. Unless your dataset is small, the Transformer page displays in its data grid a sample taken from the first set of rows in the first file in your dataset. Be sure to read the section on Sampling below.

   1. The Initial Data button is part of the menu bar, which provide access to key application features and capabilities. The menu bar is described later.

   2. Below the menu bar is the toolbar, which contains a set of tools that you can select to apply to your data. The toolbar is described later.

2. The recipe is empty: This area is the recipe panel, where you can see the steps that you have added to your recipe. A recipe is the set of steps that you create to transform your data.

3. Primary_Website_or_URL: Above the data in each column you can see two bars.

   1. Histogram: The lower bar is a histogram of values in the column.

   2. Data quality bar: Above the histogram is the data quality bar, which displays the comparative percentage of valid (green), invalid (red), and missing (black) values in a column, as compared to the data type of the column.

2 - Interact

The data grid is not just for display; you can interact with it to select data elements of interest to you. For example, in the Farmer's Market dataset, you may have noticed that the Primary_Website_or_URL column has a large black bar, which indicates many missing values in the column. A reasonable interpretation of these missing values is that these Farmer's Markets do not have a known web site. To explore these markets, you click the black bar:

Figure: Interact: select missing values in a column

A couple of things to notice:

• The right panel now displays a list of suggestions. These suggestions correspond to transformations that you can apply to the current selection. In this case, the selected data are the missing values in the Primary_Website_or_URL column.

• The first suggestion is selected by default. In the data grid, you can see a preview of what would happen if this step was added to the recipe. The highlighted rows would be removed.

Things you can select:

• Column(s): Click a column header. Use SHIFT or CTRL/COMMAND to select multiple columns.

• Row(s): To select a row, click the dot on the left side of the data grid.

• Cell values: Click and drag to select part of a cell value. Double-click to select the whole cell value.

• Histogram values:You can select one or more values in the histogram. Based on your selection(s), a set of suggested changes is presented to you. These changes apply to the rows where the selected values occur.

• Data quality bar: You can select bars in the data quality bar.

3 - Predict

In the above example, predictive transformation predicted that based on your selection, you wanted to delete the rows where the website URL column was empty. This prediction is based on multiple factors, including:

• Platform algorithms that interpret the meaning of user selections

• Tracking of previous user interactions with your project or workspace of a similar nature.

So, the prediction makes sense; in most cases, if data is missing, you don't want the rest of the incomplete data in the row. However, suppose your interest in the data is different. Suppose that you are a website builder. You want to find the farmer's markets that don't have a website as potential customers. In this case, the second suggested transformation makes more sense: Keep these rows with missing values:

Figure: Predict: Multiple suggestions serve as guides to possible useful transformations

When the second suggestion is selected, the affected rows are highlighted in green, indicating that they will be retained.

Tip

To visualize assist in reviewing these rows, click the Show only affected checkbox at the bottom of the screen. Only the rows are displayed where the suggested transformation is to be applied.

4 - Present

If the selected suggestion looks good, click Add:

Figure: Present: Data grid is updated based on the transformation that has been applied

The transformation has been added to your recipe. Note at there is now a step in the recipe panel at right, which was previously empty. You've added a recipe step, and your sampled data is transformed.

Designer Cloud Powered by Trifacta Enterprise Edition supports the following methods for building recipes in the Transformer page. These methods are listed in order of ease of use:

Loading very large datasets in Designer Cloud Powered by Trifacta Enterprise Edition can overload your browser or otherwise impact performance, so the application is designed to work on a sample of data. After you have finished your recipe working on a sample, you execute the recipe across the entire dataset.

The default sample is the first set of rows of source data in the dataset, the number of which is determined by the platform. For smaller datasets, the entire dataset can be used as your sample. In the Transformer page, it's listed as Initial Data in the upper-left corner.

In some cases, the default sample might be inadequate or of the wrong type. To generate a new sample, click the name of the sample in the upper-left corner.

On the right side of the Transformer page, you can launch a new sampling job on your dataset from the Samples panel. You may have to open it first.

Data cleansing tasks address issues in data quality, which can be broadly categorized as follows:

When data is initially imported, it can contain multiple columns, rows, or specific values that you don't need for your final output. Specifically, this phase can involve the following basic activities:

Use a row to create headers:

In most cases, the names of your columns are inferred from the first row of the data in the dataset. If you need to specify a different row, please complete the following:

Generate metadata:

On the left side of the data grid, you might notice a set of black dots. If you hover over one of these, the original row number from the source data is listed. Since the data transformation process can change the number of rows or their order, you might want to retain the original order of the rows.

To retain the original row numbers in a column called, rowId, please complete the following:

Single row formula

$sourcerownumber

rowId

You can use a similar transformation to generate the full path and filename to file-based sources:

Single row formula

$filepath

filepath

Delete unused columns:

Your data might contain columns that are not of use to you, so it's in your interest to remove them to simplify the dataset. To delete a column, click the caret next to the column's title and select Delete.

Check column data types:

When a dataset is imported, Designer Cloud Powered by Trifacta Enterprise Edition attempts to identify the data type of the column from the first set of rows in the column. At times, however, type inference can be incorrect.

Display only columns of interest:

You can choose which columns you want to display in the data grid, which can be useful to narrow your focus to problematic areas.

In the Status bar at the bottom of the screen, click the Eye icon.

Review data quality:

After you have removed unused data, you can examine the quality of data within each column just below the column title.

Figure: Column header with data quality bar

The horizontal bar, known, as the data quality bar, identifies the quality of the data in the column by the following colors:

Example - Suggestion Cards

Based on your selections and its knowledge of common data patterns, Designer Cloud Powered by Trifacta Enterprise Edition prompts you with suggested transformations. You can then select pre-configured transformations in the right panel of the Transformer page to quickly add steps.

In the following example, the missing values in the SUBSCRIBER_AGE column have been selected, and a set of suggestion cards is displayed.

Figure: Selecting missing values

Depending on the nature of the data, you might want to keep, delete, or modify the values. Since the data is missing, the Delete card has been selected.

For more background information, see Overview of Predictive Transformation.

Change data types:

If a column contains a high concentration of mismatched data (red), the column might have been identified as the wrong data type. For example, your dataset includes internal identifiers that are primarily numeric data (e.g. 10000022) but have occasional alphabetical characters in some values (e.g. 1000002A). The column for this data might be typed for integer values, when it should be treated as string values.

Explore column details:

As needed, you can explore details about the column's data, including statistical information such as outliers. From the caret drop-down next to a column name, select Column Details.

You can review a histogram of the values found in the column. Just below a column's data quality bar, you can find the histogram of values. In the following example, the data histogram on the left applies to the ZIP column, and the right applies to the WEB_CHAT_ID column.

Figure: Column data histogram

When you mouse over the categories in the histogram, you can see the corresponding value, the count of instances in the sample's column, and the percentage of affected rows. In the left one, the bar with the greatest number of instances has been selected; the value 21202 occurs 506 times (21.28%) in the dataset. On the right, the darker shading indicates how rows with ZIP=21202 map to values in the WEB_CHAT_ID column.

Before you deliver your data to the target system, you might need to enhance or augment the dataset with new columns or values from other datasets.

Union datasets:

You can append a dataset of identical structure to your currently loaded one to expand the data volume. For example, you can string together daily log data to build weeks of log information using the Union page.

Join datasets:

You can also join together two or more datasets based on a common set of values. For example, you are using raw sales data to build a sales commission dataset:

This commission dataset is created by performing an inner join between the sales transaction dataset and the employee dataset. In the Search panel, enter join to join data.

Lookup values:

In some cases, you might need to include or replace values in your dataset with other columns from another dataset. For example, transactional data can reference product and customer by internal identifiers. You can create lookups into your master data set to retrieve user-friendly versions of customer and product IDs.

To perform a lookup for a column of values, click the caret drop-down next to the column title and select Lookup....

The data that you see in the Transformer page is a sample of your entire dataset.

For larger datasets, you must learn how to generate new samples, which can provide different perspectives on your data and, in complex flows, enhance performance.

As part of the transformation process, you can generate and review visual profiles of individual columns and your entire dataset. These interactive profiles can be very helpful in identifying anomalies, outliers, and other issues with your data.