ML Python 400 - Impute KNN-Based Mixed-Type

Download the whole workflow group "Missing Value Imputation":

https://hub.knime.com/s/smId4VcHVVVL5D_t

Explanation of the KNN-based Imputation Method

This imputation pipeline uses K-Nearest Neighbors (KNN) to fill missing values in both numeric and categorical (string) variables.

1️⃣ Numeric variables

• All numeric features are scaled to the 0–1 range using MinMaxScaler so that variables with large magnitudes (e.g., income vs. age) don’t dominate the distance calculation.

• The KNNImputer then replaces each missing numeric value with the average of the corresponding values from its k nearest neighbors (default k = 5).

• “Nearest” means “most similar” rows, based on Euclidean distance across all available features (both numeric and encoded categorical).

• After imputation, the numeric values are rescaled back to their original scale.

2️⃣ Categorical (string) variables

• Categorical columns are first converted to numeric codes using OrdinalEncoder. Each unique category becomes a distinct integer code, and missing values remain as NaN.

• During KNN imputation, missing codes are replaced with the mean code of the nearest neighbors, which is then rounded to the nearest integer before converting back to the original category labels.

• This effectively assigns the most typical category among similar rows to each missing entry.

3️⃣ Datetime variables

• Datetimes (if present) are temporarily converted to seconds since epoch (numeric form) so that KNN can treat them as numeric features.

✅ In summary:
KNN imputation fills gaps by “borrowing” information from similar rows rather than using a global statistic like mean or mode. It works for mixed data types (numeric + categorical) and preserves relationships between features, producing more context-aware imputations than simple mean/mode filling.

KNN-Based Mixed-Type Imputation Workflow

Imputes missing values in numeric and categorical data using a K-Nearest Neighbors model with scaling and reversible encoding.

Apply KNN Missing Value Imputation to new data

../nmodel/ folder will contain

py_400_knn_imputer.pkl

py_400_knn_scaler.pkl

py_400_knn_encoder.pkl