Encode Transform

Use Encode to turn categorical values into numerical features. Standard execution (Pandas) keeps each source column and adds encoded columns. Several methods produce a different schema in Big Data/AWS Glue execution (PySpark).

Basic Usage

To encode categorical columns:

1. Select the Encode transform from the transform menu.
2. Select one or more fields under Target Columns.
3. Choose a Method.
4. Apply the transformation.

note

Target Columns lists categorical columns only.

If no categorical columns appear

If Target Columns says that no categorical columns were found, none of the columns in the current input schema are identified as categorical.

Add Infer Schema Type between the data source and Encode, then run or preview that upstream node. Reopen Encode to select the columns it identified as categorical under Target Columns.

Configuration Options

Basic Options

• Target Columns: Select one or more categorical columns to encode.
• Method: Choose one of these encodings:
  • One-Hot Encoding
  • Label Encoding
  • Ordinal Encoding
  • Binary Encoding
  • Frequency Encoding
  • Hashing Encoding

tip

Select the information icon beside Method to open the encoding method guide.

Standard Execution Output

In Standard execution, encoded columns are added alongside their source columns.

Method	Columns appended for each selected source column
One-Hot	One {source}_{category} column per observed non-null category
Label	One {source}_label column
Ordinal	One {source}_ordinal column
Binary	A data-dependent number of bit columns, normally named {source}_0, {source}_1, and so on
Frequency	One {source}_frequency column containing raw occurrence counts
Hashing	Eight columns named {source}_hash_0 through {source}_hash_7

The Encode panel cannot remove source columns, set the order of ordinal categories, or change the Standard hashing width.

Encoding Methods

One-Hot Encoding

Keeps the source column and adds one binary indicator for each observed category.

Use for: Unordered values such as red, blue, and green.

Label Encoding

Keeps the source column and assigns each category a unique integer. These codes may imply an order that does not exist.

Use for: Binary values or columns with many categories when the implied ordering is acceptable.

Ordinal Encoding

Keeps the source column and adds an integer-ranked column. Standard execution respects an existing ordered pandas categorical type. Otherwise, it infers an order that may not match the intended meaning.

Use for: Values with a defined order, such as shirt sizes.

caution

The Encode panel does not let you set category order. Do not assume S, M, L, and XL will receive that order unless it is already stored in the input column's categorical metadata.

Binary Encoding

Keeps the source column and adds binary-code bit columns. This usually creates fewer columns than one-hot encoding.

Use for: Columns with many categories where one-hot encoding would be too wide.

Frequency Encoding

Keeps the source column and adds each category's raw occurrence count.

Use for: Distinguishing rare and common values when frequency is meaningful.

Hashing Encoding

Maps categories through a hash function. Standard execution keeps the source and adds eight hash columns for each selected field.

Use for: High-cardinality fields where a fixed output width is useful.

The current Encode panel does not provide a hash-width control.

Execution Mode Differences

Method	Standard execution (Pandas)	Big Data/AWS Glue execution (PySpark)
One-Hot	Retains the source and appends one scalar column per category.	Replaces each selected source with one {source}_encoded vector column.
Label	Retains the source and appends {source}_label.	Replaces each selected source with {source}_indexed.
Ordinal	Retains the source and appends {source}_ordinal; respects an existing ordered categorical dtype.	Uses the same string-indexing behavior as Label and replaces the source with {source}_indexed.
Binary	Retains the source and appends bit columns.	Not supported.
Frequency	Retains the source and appends {source}_frequency.	Replaces each selected source with {source}_freq.
Hashing	Retains each source and appends eight scalar hash columns per source.	Replaces the selected sources with one combined hashed_features vector, using 262,144 features by default.

Execution mode affects the schema

The same configuration can produce different column names, column counts, and data types when you switch between Standard and Big Data/AWS Glue.

Examples

This example uses Standard execution.

Example: Encoding Product Categories

Input Dataset:

Product ID	Category	Price
1	Electronics	500
2	Clothing	50
3	Electronics	750
4	Home	200
5	Clothing	75

Configuration:

• Target Columns: Category
• Method: One-Hot Encoding

Result:

Product ID	Category	Price	Category_Clothing	Category_Electronics	Category_Home
1	Electronics	500	0	1	0
2	Clothing	50	1	0	0
3	Electronics	750	0	1	0
4	Home	200	0	0	1
5	Clothing	75	1	0	0

Best Practices

1. Choose a method based on whether the categories are ordered and how many unique values they contain.
2. Use Binary or Hashing Encoding if One-Hot would create too many columns.
3. Use Ordinal Encoding for a semantic rank only when the input already has the correct ordered categorical metadata.
4. In Standard execution, add Drop Columns later if downstream nodes should receive only the encoded fields.
5. Keep the encoding scheme consistent between training and test data.

Troubleshooting

• A field is missing: If it does not appear under Target Columns, check that it is identified as categorical.
• Ordinal Encoding produces an unexpected rank: Check whether the input column has ordered categorical metadata. You cannot change the order in the Encode panel.
• Hashing Encoding produces an unsuitable output shape: Standard execution always creates eight columns per selected source in the current UI. Choose another method if needed.