Feature Selection

Feature selection ranks the columns of a dataset by how useful they are for predicting your targets, and can drop the weakest. It appears in three places:

Where	What it does	Changes data?
The Feature Eng tab	Ranks a completed run's columns without re-running anything.	No
Feature Selection, before imputation (Common tier)	A pipeline stage between type conversion and missing-data handling that keeps only the selected columns.	Yes
Column selection, after scaling (Enterprise, Experimental tier)	Target-aware selection on the fully prepared columns.	Yes

Targets and excluded columns are always kept, and if selection fails the data passes through unchanged.

Methods

Method	Id	Kind	Ranks or prunes by
Mutual Information (default)	mutual_information	supervised	Information each column carries about the target.
SHAP (LightGBM)	shap_importance	supervised	Contribution to a fast tree model's predictions.
Recursive Feature Elimination	rfe	supervised	Repeatedly removing the weakest column.
Variance Threshold	variance_threshold	unsupervised	Removes near-constant columns (threshold default 0.01).
Correlation Pruning	correlation_pruning	unsupervised	Removes one of each highly correlated pair (threshold default 0.95).
PCA	pca	unsupervised	Replaces the features with principal components.
Permutation importance (API)	permutation_importance	supervised	How much a model worsens when the column is shuffled.
Leakage guard (API)	leakage_guard	supervised	Flags columns that predict the target almost perfectly.

Supervised methods work on a sample of up to 50,000 rows; with several targets, a column useful for any of them survives.

The Feature Eng tab

1. Session: pick a completed run, or choose Enter a session ID manually and paste the full session ID (the shortened ID shown in Runs is rejected).
2. Target column(s): click the column chips (the run's own targets are pre-selected). The columns are those of the run's type-converted output. If they cannot be loaded, type them comma-separated.
3. Methods: toggle one or more; Top K 1 – 200, default 20.
4. Analyze: the Ranking table shows rank, feature, an importance bar and a score from 0 to 1. With several methods, each method's scores are scaled to 0 – 1 and merged. Warnings appear above the table.

If you like the ranking, re-run the pipeline with Feature Selection switched on.

API

Both /api/v1/feature-selection/... and /api/v1/feature-engineering/... work (also without /v1), with a session cookie or an API key.

GET /api/v1/feature-selection/columns?session_id=<full id>
-> {"success": true, "columns": [...], "row_count": 5000}

POST /api/v1/feature-selection/recommend
{"session_id": "<full id>", "target_columns": ["churn"],
 "methods": ["mutual_information", "shap_importance"], "top_k": 15}
-> {"success": true, "ranking": [{"column": "tenure", "score": 1.0, "rank": 1}],
    "per_method": {...}, "leakage": {...}, "warnings": [...],
    "methods": [...], "top_k": 15, "row_count": 5000}

Field	Type	Default	Notes
session_id	string	required	Full ID of a completed run you own or that is shared with you.
target_columns	list	required	Must be columns of the ranked output.
methods	list	["mutual_information"]	Any ids above; unknown ids are ignored with a warning.
top_k	integer	20	Length of the merged ranking.

Errors: 400 for missing fields or unknown targets, 403 for another user's run, 404 for an unknown session or a run with no stored output.

Feature Selection as a pipeline stage

Tick Feature Selection, before imputation in the Common tier and open Advanced parameters for Method and Keep top N (1 – 1000, default 20; Variance and Correlation prune on their threshold instead). The stage adds a progress step and writes feature_selection_output.csv and feature_selection_report.json.

Setting	JSON key	Multipart field	Default
On / off	enable_feature_selection	enableFeatureSelection	false
Method	feature_selection_config.method	featureSelectionConfig	mutual_information from the dashboard; variance_threshold when omitted
Keep top N	feature_selection_config.top_k	(same object)	20
Threshold (API)	feature_selection_config.threshold	(same object)	variance 0.01, correlation 0.95

In JSON, method may also be a list (each runs on the previous result) or auto (leakage guard, variance, correlation, then mutual information). On /api/v1/process switch the stage on with tools.feature_selection and pass feature_selection_config.

Column selection, after scaling (Enterprise)

Enterprise accounts can add target-aware selection on the prepared columns: leakage guard, variance, correlation, mutual information, permutation importance, recursive elimination and SHAP, run in the order chosen, with an optional top N, and suspected leakage flagged or removed. Key selection_config (multipart selectionConfig). For other accounts the Dimensionality & Similarity Manager removes only constant and near-duplicate columns.