Feature engineering - Sourcetable

Feature engineering turns raw data into inputs that improve model performance. Sourcetable automates common transformations through the AI assistant.

Missing value handling

"Handle missing values in the dataset — use median for numeric and mode for categorical"

Strategy	When to use
Mean/median imputation	Numeric columns, few missing values
Mode imputation	Categorical columns
Forward/backward fill	Time series data
KNN imputation	Values depend on other features
Drop rows	Small percentage missing, large dataset
Indicator column	Missingness itself is informative

"Encode all categorical columns for the ML model"

Method	When to use
One-hot encoding	Nominal categories (color, city) with low cardinality
Label encoding	Ordinal categories (low/medium/high)
Target encoding	High-cardinality categories (zip code, product ID)
Binary encoding	High-cardinality with many categories
Frequency encoding	When category frequency matters

"Normalize all numeric features to 0-1 range"

Method	When to use
StandardScaler (Z-score)	Most ML algorithms, normally distributed data
MinMaxScaler (0-1)	Neural networks, distance-based algorithms
RobustScaler	Data with outliers
Log transform	Right-skewed distributions (income, prices)
Box-Cox transform	Make data more normal, various skew levels

"Extract year, month, day of week, and hour from the timestamp column"

Creates: year, quarter, month, week, day of week, day of month, hour, minute, is_weekend, is_holiday, days_since_start.

"Create customer-level features: total purchases, average order value, days since last order"

Group-by aggregations: sum, mean, median, count, min, max, std, first, last.

"Create interaction features between price and quantity"

Products, ratios, and polynomial features between numeric columns.

"Extract features from the product description column"

Creates: word count, character count, average word length, sentiment score, TF-IDF vectors.

"Rank features by importance using a random forest"

Trains a model and ranks features by their contribution to predictions.

"Check for multicollinearity and drop highly correlated features"

Calculates VIF for each feature. Values above 5-10 indicate problematic multicollinearity.

"Use RFE to find the optimal set of features for the model"

Iteratively removes the least important feature until performance stops improving.