Prepare data for model building

Your machine learning dataset might require data preparation before you build your model. You might want to clean your data due to various issues, which might include missing values or outliers, and perform feature engineering to improve the accuracy of your model. Amazon SageMaker Canvas provides ML data transforms with which you can clean, transform, and prepare your data for model building. You can use these transforms on your datasets without any code. SageMaker Canvas adds the transforms you use to the Model recipe, which is a record of the data preparation done on your data before building the model. Any data transforms you use only modify the input data for model building and do not modify your original data source.

The preview of your dataset shows the first 100 rows of the dataset. If your dataset has more than 20,000 rows, Canvas takes a random sample of 20,000 rows and previews the first 100 rows from that sample. You can only search for and specify values from the previewed rows, and the filter functionality only filters the previewed rows and not the entire dataset.

The following transforms are available in SageMaker Canvas for you to prepare your data for building.

Drop columns

You can exclude a column from your model build by dropping it in the Build tab of the SageMaker Canvas application. Deselect the column you want to drop, and it isn't included when building the model.

Filter rows

The filter functionality filters the previewed rows (the first 100 rows of your dataset) according to conditions that you specify. Filtering rows creates a temporary preview of the data and does not impact the model building. You can filter to preview rows that have missing values, contain outliers, or meet custom conditions in a column you choose.

Filter rows by missing values

Missing values are a common occurrence in machine learning datasets. If you have rows with null or empty values in certain columns, you might want to filter for and preview those rows.

To filter missing values from your previewed data, do the following.

SageMaker Canvas filters for rows that contain missing values in the Column you selected and provides a preview of the filtered rows.

Filter rows by outliers

Outliers, or rare values in the distribution and range of your data, can negatively impact model accuracy and lead to longer building times. SageMaker Canvas enables you to detect and filter rows that contain outliers in numeric columns. You can choose to define outliers with either standard deviations or a custom range.

To filter for outliers in your data, do the following.

The Standard deviation option detects and filters for outliers in numeric columns using the mean and standard deviation. You specify the number of standard deviations a value must vary from the mean to be considered an outlier. For example, if you specify 3 for SD, a value must fall more than 3 standard deviations from the mean to be considered an outlier.

The Custom range option detects and filters for outliers in numeric columns using minimum and maximum values. Use this method if you know your threshold values that delimit outliers. You can set the Type of the range to either Percentile or Number. If you choose Percentile, the Min and Max values should be the minimum and maximum of the percentile range (0-100) that you want to allow. If you choose Number, the Min and Max values should be the minimum and maximum numeric values that you want to filter in the data.

Filter rows by custom values

You can filter for rows with values that meet custom conditions. For example, you might want to preview rows that have a price value greater than 100 before removing them. With this functionality, you can filter rows that exceed the threshold you set and preview the filtered data.

To use the custom filter functionality, do the following.

For the Operation, you can choose one of the following options. Note that the available operations depend on the data type of the column you choose. For example, you cannot create a is greater than operation for a column containing text values.

After you set the filter operation, SageMaker Canvas updates the preview of the dataset to show you the filtered data.

Functions and operators

You can use mathematical functions and operators to explore and distribute your data. You can use the SageMaker Canvas supported functions or create your own formula with your existing data and create a new column with the result of the formula. For example, you can add the corresponding values of two columns and save the result to a new column.

You can nest statements to create more complex functions. The following are some examples of nested functions that you might use.

You can specify functions in the data preparation stage before you build your model. To use a function, do the following.

SageMaker Canvas saves the result of your function to a new column using the name you specified in New Column Name. You can view or remove functions from the Model Recipe panel.

SageMaker Canvas supports the following operators for functions. You can use either the text format or the in-line format to specify your function.

SageMaker Canvas also supports aggregate operators, which can perform operations such as calculating the sum of all the values or finding the minimum value in a column. You can use aggregate operators in combination with standard operators in your functions. For example, to calculate the difference of values from the mean, you could use the function Abs(height – avg(height)). SageMaker Canvas supports the following aggregate operators.

Manage rows

With the Manage rows transform, you can perform sort, random shuffle, and remove rows of data from the dataset.

Sort rows

To sort the rows in a dataset by a given column, do the following.

Shuffle rows

To randomly shuffle the rows in a dataset, do the following.

Drop duplicate rows

To remove duplicate rows in a dataset, do the following.

Remove rows by missing values

Missing values are a common occurrence in machine learning datasets and can impact model accuracy. Use this transform if you want to drop rows with null or empty values in certain columns.

To remove rows that contain missing values in a specified column, do the following.

SageMaker Canvas drops rows that contain missing values in the Column you selected. After removing the rows from the dataset, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the rows return to your dataset.

Remove rows by outliers

Outliers, or rare values in the distribution and range of your data, can negatively impact model accuracy and lead to longer building times. With SageMaker Canvas, you can detect and remove rows that contain outliers in numeric columns. You can choose to define outliers with either standard deviations or a custom range.

To remove outliers from your data, do the following.

The Standard deviation option detects and removes outliers in numeric columns using the mean and standard deviation. You specify the number of standard deviations a value must vary from the mean to be considered an outlier. For example, if you specify 3 for Standard deviations, a value must fall more than 3 standard deviations from the mean to be considered an outlier.

The Custom numeric range and Custom quantile range options detect and remove outliers in numeric columns using minimum and maximum values. Use this method if you know your threshold values that delimit outliers. If you choose a numeric range, the Min and Max values should be the minimum and maximum numeric values that you want to allow in the data. If you choose a quantile range, the Min and Max values should be the minimum and maximum of the percentile range (0–100) that you want to allow.

After removing the rows from the dataset, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the rows return to your dataset.

Remove rows by custom values

You can remove rows with values that meet custom conditions. For example, you might want to exclude all of the rows with a price value greater than 100 when building your model. With this transform, you can create a rule that removes all rows that exceed the threshold you set.

To use the custom remove transform, do the following.

For the Operation, you can choose one of the following options. Note that the available operations depend on the data type of the column you choose. For example, you cannot create a is greater than operation for a column containing text values.

After removing the rows from the dataset, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the rows return to your dataset.

Rename columns

With the rename columns transform, you can rename columns in your data. When you rename a column, SageMaker Canvas changes the column name in the model input.

You can rename a column in your dataset by double-clicking on the column name in the Build tab of the SageMaker Canvas application and entering a new name. Pressing the Enter key submits the change, and clicking anywhere outside the input cancels the change. You can also rename a column by clicking the More options icon ( ), located at the end of the row in list view or at the end of the header cell in grid view, and choosing Rename.

Your column name can’t be longer than 32 characters or have double underscores (__), and you can’t rename a column to the same name as another column. You also can’t rename a dropped column.

The following screenshot shows how to rename a column by double-clicking the column name.

When you rename a column, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the column reverts to its original name.

Manage columns

With the following transforms, you can change the data type of columns and replace missing values or outliers for specific columns. SageMaker Canvas uses the updated data types or values when building your model but doesn’t change your original dataset. Note that if you've dropped a column from your dataset using the Drop columns transform, you can't replace values in that column.

Replace missing values

Missing values are a common occurrence in machine learning datasets and can impact model accuracy. You can choose to drop rows that have missing values, but your model is more accurate if you choose to replace the missing values instead. With this transform, you can replace missing values in numeric columns with the mean or median of the data in a column, or you can also specify a custom value with which to replace missing values. For non-numeric columns, you can replace missing values with the mode (most common value) of the column or a custom value.

Use this transform if you want to replace the null or empty values in certain columns. To replace missing values in a specified column, do the following.

After replacing the missing values in the dataset, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the missing values return to the dataset.

Replace outliers

Outliers, or rare values in the distribution and range of your data, can negatively impact model accuracy and lead to longer building times. SageMaker Canvas enables you to detect outliers in numeric columns and replace the outliers with values that lie within an accepted range in your data. You can choose to define outliers with either standard deviations or a custom range, and you can replace outliers with the minimum and maximum values in the accepted range.

To replace outliers in your data, do the following.

The Standard deviation option detects outliers in numeric columns using the mean and standard deviation. You specify the number of standard deviations a value must vary from the mean to be considered an outlier. For example, if you specify 3 for Standard deviations, a value must fall more than 3 standard deviations from the mean to be considered an outlier. SageMaker Canvas replaces outliers with the minimum value or maximum value in the accepted range. For example, if you configure the standard deviations to only include values from 200–300, then SageMaker Canvas changes a value of 198 to 200 (the minimum).

The Custom numeric range and Custom quantile range options detect outliers in numeric columns using minimum and maximum values. Use this method if you know your threshold values that delimit outliers. If you choose a numeric range, the Min and Max values should be the minimum and maximum numeric values that you want to allow. SageMaker Canvas replaces any values that fall outside of the minimum and maximum to the minimum and maximum values. For example, if your range only allows values from 1–100, then SageMaker Canvas changes a value of 102 to 100 (the maximum). If you choose a quantile range, the Min and Max values should be the minimum and maximum of the percentile range (0–100) that you want to allow.

After replacing the values in the dataset, SageMaker Canvas adds the transform in the Model recipe section. If you remove the transform from the Model recipe section, the original values return to the dataset.

Change data type

SageMaker Canvas provides you with the ability to change the data type of your columns between numeric, text, and datetime, while also displaying the associated feature type for that data type. A data type refers to the format of the data and how it is stored, while the feature type refers to the characteristic of the data used in machine learning algorithms, such as binary or categorical. This gives you the flexibility to manually change the type of data in your columns based on the features. The ability to choose the right data type ensures data integrity and accuracy prior to building models. These data types are used when building models.

The following table lists all of the supported data types in Canvas.

The following table lists all of the supported feature types in Canvas.

To modify data type of a column in a dataset, do the following.

The data type for your column should now be updated.

Prepare time series data

Use the following functionalities to prepare your time series data for building time series forecasting models.

Resample time series data

By resampling time-series data, you can establish regular intervals for the observations in your time series dataset. This is particularly useful when working with time series data containing irregularly spaced observations. For instance, you can use resampling to transform a dataset with observations recorded every one hour, two hour and three hour intervals into a regular one hour interval between observations. Forecasting algorithms require the observations to be taken at regular intervals.

To resample time series data, do the following.

The following table lists all of the Frequency types you can select when resampling time series data.

When applying the resampling transform, you can use the Advanced option to specify how the resulting values of the rest of the columns (other than the timestamp column) in your dataset are modified. This can be achieved by specifying the resampling methodology, which can either be downsampling or upsampling for both numeric and non-numeric columns.

Downsampling increases the interval between observations in the dataset. For example, if you downsample observations that are taken either every hour or every two hours, each observation in your dataset is taken every two hours. The values of other columns of the hourly observations are aggregated into a single value using a combination method. The following tables show an example of downsampling time series data by using mean as the combination method. The data is downsampled from every two hours to every hour.

The following table shows the hourly temperature readings over a day before downsampling.

The following table shows the temperature readings after downsampling to every two hours.

To downsample time series data, do the following:

If you don’t specify combination methods, the default values are Most Common for Non-numeric combination and Mean for Numeric combination. The following table lists the methods for numeric and non-numeric combination.

Upsampling reduces the interval between observations in the dataset. For example, if you upsample observations that are taken every two hours into hourly observations, the values of other columns of the hourly observations are interpolated from the ones that have been taken every two hours.

To upsample time series data, do the following:

If you don’t specify estimation methods, the default values are Forward Fill for Non-numeric estimation and Linear for Numeric estimation. The following table lists the methods for estimation.

The following screenshot shows the Advanced settings with the fields for downsampling and upsampling filled out.

Use datetime extraction

With the datetime extraction transform, you can extract values from a datetime column to a separate column. For example, if you have a column containing dates of purchases, you can extract the month value to a separate column and use the new column when building your model. You can also extract multiple values to separate columns with a single transform.

Your datetime column must use a supported timestamp format. For a list of the formats that SageMaker Canvas supports, see Time Series Forecasts in Amazon SageMaker Canvas. If your dataset does not use one of the supported formats, update your dataset to use a supported timestamp format and re-import it to Amazon SageMaker Canvas before building your model.

To perform a datetime extraction, do the following.

SageMaker Canvas creates a new column in the dataset for each of the values you extract. Except for Year values, SageMaker Canvas uses a 0-based encoding for the extracted values. For example, if you extract the Month value, January is extracted as 0, and February is extracted as 1.

You can see the transform listed in the Model recipe section. If you remove the transform from the Model recipe section, the new columns are removed from the dataset.