Feature EDA

See also

Concepts: Feature EDA | API Tutorial: Credit Default — Step 10b

Prerequisites

This page uses the wait_for_task helper defined in API Overview.

Run EDA on features to analyze their distributions and relationships with the target. You can provide features from different sources.

When the ideation pipeline runs EDA, it uses Residual EDA if a naive prediction is available (typical for forecast use cases). Instead of scoring features against the raw target, Residual EDA measures what each feature adds on top of the naive baseline — plots are computed on residuals (target minus naive) or ratios (target divided by naive), and scoring uses the Incremental Predictive Score. This ensures features are evaluated on their incremental contribution rather than on signal the naive prediction already captures. See Residual EDA for details.

You can trigger Residual EDA explicitly via POST /eda by including the naive_prediction field in the request body.

From Feature IDs

import featurebyte as fb

client = fb.Configurations().get_client()

response = client.post(
    "/eda",
    json={
        "feature_ids": [feature_id_1, feature_id_2],
        "use_case_id": use_case_id,
    },
)
task_id = response.json()["id"]
task = wait_for_task(client, task_id)

eda_id = task.get("payload", {}).get("output_document_id")

From a Feature Ideation

response = client.post(
    "/eda",
    json={
        "feature_ideation_id": feature_ideation_id,
        "use_case_id": use_case_id,
    },
)
task_id = response.json()["id"]
task = wait_for_task(client, task_id)

eda_id = task.get("payload", {}).get("output_document_id")

From a Feature List

response = client.post(
    "/eda",
    json={
        "feature_list_id": feature_list_id,
        "use_case_id": use_case_id,
    },
)
task_id = response.json()["id"]
task = wait_for_task(client, task_id)

eda_id = task.get("payload", {}).get("output_document_id")

Parameters:

Parameter	Type	Required	Description
use_case_id	string	Yes	ID of the use case providing context for EDA
feature_ids	array	One of	List of feature IDs to analyze
feature_ideation_id	string	One of	ID of a feature ideation to analyze all its features
feature_list_id	string	One of	ID of a feature list to analyze all its features
naive_prediction	object	No	Include to trigger Residual EDA (see Residual EDA)
naive_prediction.feature_id	string	—	ID of the naive prediction feature
naive_prediction.structure	string	—	"additive" (residuals: target - naive) or "multiplicative" (ratios: target / naive)
overwrite	boolean	No	Overwrite existing EDA results (default: false)

Get Feature EDA Details

Retrieve full EDA results for a specific feature, including power scores and analysis metadata:

response = client.get(f"/eda/{feature_eda_id}")
eda = response.json()

print(f"Feature: {eda.get('feature_id')}")
print(f"Predictive power score: {eda.get('predictive_power_score')}")
print(f"Feature type: {eda.get('feature_type')}")
print(f"Target type: {eda.get('target_type')}")
print(f"Target categories: {eda.get('target_categories')}")

Response fields:

Field	Type	Description
id	string	Feature EDA ID
feature_id	string	Associated feature ID
feature_type	string	Feature type: "numerical", "categorical", "text", "dict", "embedding"
target_type	string	Target type: "REGRESSION", "BINARY_CLASSIFICATION", etc.
predictive_power_score	float	Predictive power score (higher = more predictive)
null_power_score	float	Power score from null values alone
no_bucket_power_score	float	Power score without bucketing
target_categories	array	Available target categories (for classification)
feature_categories	array	Available feature categories (dictionary keys or embedding dimensions to filter by)
feature_source	string	Source of the feature (e.g., "CATALOG")
definition_hash	string	Hash of the feature definition
version	object	Feature version identifier
error_reason	string	Error description if the analysis failed
plots	array	Plot data objects, each containing an info field with summary statistics (see below)
use_case_id	string	Associated use case ID
context_id	string	Associated context ID
observation_table_id	string	Observation table used for EDA

Summary Statistics

Each item in plots contains an info object with distribution and target statistics:

eda = client.get(f"/eda/{feature_eda_id}").json()

for plot in eda.get("plots", []):
    info = plot.get("info", {})
    print(f"Count: {info.get('count')}, Unique: {info.get('unique')}")
    print(f"Mean: {info.get('mean')}, Std: {info.get('stddev')}")
    print(f"Min: {info.get('min_val')}, Max: {info.get('max_val')}")
    print(f"Missing: {info.get('num_missing')}, Zeros: {info.get('num_zeros')}")
    print(f"Target mean (non-missing): {info.get('target_mean_non_missing')}")

info fields (numeric features):

Field	Type	Description
count	integer	Total number of observations
unique	integer	Number of distinct values
mean	float	Mean value
stddev	float	Standard deviation
min_val	float	Minimum value
max_val	float	Maximum value
q01, q05, q10, q25, q50, q75, q90, q95, q99	float	Percentiles
num_missing	integer	Number of missing values
num_non_missing	integer	Number of non-missing values
num_zeros	integer	Number of zero values
num_non_zeros	integer	Number of non-zero values
pct_zeros	float	Percentage of zeros
num_lower_outliers	integer	Number of lower outliers
num_upper_outliers	integer	Number of upper outliers
target_mean_non_missing	float	Mean target value for non-missing feature values
target_mean_missing	float	Mean target value for missing feature values
target_mean_zeros	float	Mean target value for zero feature values
target_mean_non_zeros	float	Mean target value for non-zero feature values
target_mean_lower_outliers	float	Mean target value for lower outlier feature values
target_mean_upper_outliers	float	Mean target value for upper outlier feature values

Get Plot Options

Check what plot options are available for a feature (e.g., target categories for classification, or dictionary/embedding keys):

response = client.request(
    "OPTIONS",
    f"/eda/{feature_id}/plots",
    params={"use_case_id": use_case_id},
)
options = response.json()

Parameters:

Parameter	Type	Required	Description
use_case_id	string	No	ID of the use case
context_id	string	No	ID of the context (alternative to use_case_id)
feature_table_id	string	No	ID of a feature table

Response fields:

Field	Type	Description
target_categories	array	Available target categories (classification use cases)
feature_categories	array	Available feature categories (dictionary and embedding features — keys/dimensions to filter by)

Get EDA Plots

Retrieve rendered plots for a specific feature. The feature_id must be one of the features included in the EDA.

response = client.get(
    f"/eda/{feature_id}/plots",
    params={
        "use_case_id": use_case_id,
        "height": 500,
        "width": 1000,
        "font_size": 16,
        "output_format": "html",
    },
)
plots = response.json()

Parameters:

Parameter	Type	Required	Description
use_case_id	string	No	ID of the use case
context_id	string	No	ID of the context (alternative to use_case_id)
target_category	any	No	Filter by target category (from OPTIONS response)
feature_category	any	No	Filter by dictionary key or embedding dimension (from OPTIONS response)
height	integer	No	Plot height in pixels (default: 500)
width	integer	No	Plot width in pixels (default: 1000)
font_size	integer	No	Font size in pixels (default: 16)
output_format	string	No	"html" (default) or "json"

Response fields (each item):

Field	Type	Description
plot_type	string	Type of plot (e.g., distribution, feature vs target)
plots	array	List of rendered plot objects with content (HTML string)

The response is a plot list. See Displaying Plots for how to render plots in Jupyter, save as HTML, or embed in a web application.