5. Set Default Cleaning Operations

Setup Cleaning Operations¶

Our data modeling is done, now we are ready to work with data itself.

A crucial step in every data science project is ensuring the data is clean and ready for feature engineering. Issues such as missing values, disguised missing values (missing values that are not explicitly encoded as missing values), or outliers can significantly impair the quality of features and eventually the quality of the final model.

FeatureByte offers an API to effectively address these concerns.

In [1]:

import featurebyte as fb

# Set your profile to the tutorial environment
fb.use_profile("tutorial")

catalog_name = "Grocery Dataset Tutorial"
catalog = fb.Catalog.activate(catalog_name)

import featurebyte as fb # Set your profile to the tutorial environment fb.use_profile("tutorial") catalog_name = "Grocery Dataset Tutorial" catalog = fb.Catalog.activate(catalog_name)

16:40:27 | INFO     | Using configuration file at: /Users/viktor/.featurebyte/config.yaml
16:40:27 | INFO     | Active profile: tutorial (https://tutorials.featurebyte.com/api/v1)
16:40:28 | INFO     | SDK version: 0.6.0.dev121
16:40:28 | INFO     | No catalog activated.
16:40:28 | INFO     | 10 feature lists, 59 features deployed
16:40:28 | INFO     | Using profile: tutorial
16:40:28 | INFO     | Using configuration file at: /Users/viktor/.featurebyte/config.yaml
16:40:28 | INFO     | Active profile: tutorial (https://tutorials.featurebyte.com/api/v1)
16:40:28 | INFO     | SDK version: 0.6.0.dev121
16:40:28 | INFO     | No catalog activated.
16:40:29 | INFO     | 10 feature lists, 59 features deployed
16:40:29 | INFO     | Catalog activated: Grocery Dataset Tutorial

Let's look into descriptive statistics of Invoice Amount

In [2]:

invoice_table = catalog.get_table("GROCERYINVOICE")

invoice_table = catalog.get_table("GROCERYINVOICE")

In [3]:

invoice_table.Amount.describe()

invoice_table.Amount.describe()

Out[3]:

	Amount
dtype	FLOAT
unique	6647
%missing	0.0
%empty	NaN
entropy	NaN
top	1
freq	798.0
mean	19.171153
std	23.717282
min	0.0
25%	4.29
50%	10.59
75%	24.5
max	360.84

Although the Amount column doesn't present any anomalies, we will set default cleaning operations to ensure that if any issues happen in the future, the data will remain clean.

We will here set the following cleaning operations:

• ignore disguised missing values equal to -99 and -98
• cap any amount less than 0 Euro
• cap any amount greater than 2000 Euros

Those operations will be applied by default when a view is created from the table. You can however overwrite those operations by creating a view in a manual mode, or access a view raw data

In [4]:

invoice_table["Amount"].update_critical_data_info(
    cleaning_operations=[
        fb.DisguisedValueImputation(disguised_values=[-99, -98], imputed_value=None),
        fb.ValueBeyondEndpointImputation(
            type="less_than", end_point=0, imputed_value=0
        ),
        fb.ValueBeyondEndpointImputation(
            type="greater_than", end_point=2000, imputed_value=2000
        ),
    ]
)

invoice_table["Amount"].update_critical_data_info( cleaning_operations=[ fb.DisguisedValueImputation(disguised_values=[-99, -98], imputed_value=None), fb.ValueBeyondEndpointImputation( type="less_than", end_point=0, imputed_value=0 ), fb.ValueBeyondEndpointImputation( type="greater_than", end_point=2000, imputed_value=2000 ), ] )

If we look at the columns_info, we'll see that critical_data_info for Amount column is populated with cleaning operations now.

In [5]:

invoice_table.info(verbose=True)

invoice_table.info(verbose=True)

Out[5]:

Event Table

name	GROCERYINVOICE
created_at	2023-11-27 15:38:20
updated_at	2023-11-27 15:40:33
description	Grocery invoice details, containing the timestamp and the total amount of the invoice
status	PUBLIC_DRAFT
catalog_name	Grocery Dataset Tutorial
record_creation_timestamp_column	record_available_at
table_details	database_name TUTORIAL_DATASETS schema_name GROCERY table_name GROCERYINVOICE
entities	name serving_names catalog_name 0 invoice [GROCERYINVOICEGUID] Grocery Dataset Tutorial 1 customer [GROCERYCUSTOMERGUID] Grocery Dataset Tutorial
semantics	['event_id', 'event_timestamp', 'record_creation_timestamp', 'time_zone']
column_count	6
columns_info	name dtype entity semantic critical_data_info description 0 GroceryInvoiceGuid VARCHAR invoice event_id None Unique identifier of each row in the table, in GUID format. Uniquely identifies each invoice. 1 GroceryCustomerGuid VARCHAR customer None None Unique identifier for each customer, in GUID format. 2 Timestamp TIMESTAMP None event_timestamp None The GMT timestamp of when this invoice transaction event occurred. 3 tz_offset VARCHAR None time_zone None The local timezone offset of the invoice event. 4 record_available_at TIMESTAMP None record_creation_timestamp None A timestamp for when this row was added to the database. 5 Amount FLOAT None None {'cleaning_operations': [{'imputed_value': None, 'type': 'disguised', 'disguised_values': [-99.0, -98.0]}, {'imputed_value': 0.0, 'type': 'less_than', 'end_point': 0.0}, {'imputed_value': 2000.0, 'type': 'greater_than', 'end_point': 2000.0}]} The total amount of the invoice, including all items and any discounts applied. Cannot be negative.
event_timestamp_column	Timestamp
event_id_column	GroceryInvoiceGuid
default_feature_job_setting	blind_spot 120s frequency 3600s time_modulo_frequency 120s

That's all! Now, every time we generate a new entry from the invoice table, we can be confident that no undesirable values will slip through.

To learn more, refer to following materials:¶

• Cleaning Operations
• Views

SDK reference for¶

• Table.get view()
• TableColumn.update_critical_data_info()