While the Log and Box-Cox transformations are effective for handling positive skewed data, they run into limitations when your data includes zero or negative values. Log transformation is undefined for non-positive numbers, and the standard Box-Cox transformation requires strictly positive input. This is where the Yeo-Johnson transformation comes into play.

Developed by I.K. Yeo and R.A. Johnson in 2000, this transformation is part of the power transformation family, similar to Box-Cox, but it extends its applicability to data containing non-positive values. The goal remains the same: to stabilize variance, reduce skewness, and make the data distribution more closely resemble a normal (Gaussian) distribution, which can be beneficial for certain modeling algorithms.

How Yeo-Johnson Works

Like Box-Cox, the Yeo-Johnson transformation seeks an optimal parameter, often denoted as $\lambda$ , to apply a power transformation. However, it uses slightly different formulas depending on whether the data point $x$ is non-negative or negative:

For $x \ge 0$ : $y = \begin{cases} \frac{(x + 1)^\lambda - 1}{\lambda} & \text{if } \lambda \neq 0 \\ \log(x + 1) & \text{if } \lambda = 0 \end{cases}$
For $x < 0$ : $y = \begin{cases} -\frac{(-x + 1)^{2 - \lambda} - 1}{2 - \lambda} & \text{if } \lambda \neq 2 \\ -\log(-x + 1) & \text{if } \lambda = 2 \end{cases}$

You don't typically need to memorize these formulas. The important concept is that the transformation provides a continuous function across zero and handles positive, zero, and negative values consistently to achieve symmetry. The optimal $\lambda$ is usually determined computationally, often using maximum likelihood estimation, to find the transformation that makes the resulting data distribution as close to Gaussian as possible.

Applying Yeo-Johnson with Scikit-learn

Scikit-learn provides a convenient implementation through the PowerTransformer class. You simply specify method='yeo-johnson'.

Let's generate some skewed data that includes non-positive values and apply the transformation:

import numpy as np
import pandas as pd
from sklearn.preprocessing import PowerTransformer
import matplotlib.pyplot as plt
import seaborn as sns

# Generate skewed data including zero and negative values
np.random.seed(42)
data_positive = np.random.exponential(scale=2, size=70)
data_nonpositive = -np.random.exponential(scale=2, size=30) + 1 # Shift to include 0 and negatives
skewed_data = np.concatenate([data_positive, data_nonpositive])
skewed_data = skewed_data.reshape(-1, 1) # Reshape for Scikit-learn transformer

# Initialize and apply Yeo-Johnson transformation
yj_transformer = PowerTransformer(method='yeo-johnson', standardize=True)
# standardize=True applies zero-mean, unit-variance scaling *after* the transformation
# Set to False if you only want the power transformation

transformed_data_yj = yj_transformer.fit_transform(skewed_data)

# Print the optimal lambda found
print(f"Optimal lambda found: {yj_transformer.lambdas_[0]:.4f}")

# Prepare data for plotting
df = pd.DataFrame({
    'Original': skewed_data.flatten(),
    'Yeo-Johnson Transformed': transformed_data_yj.flatten()
})

# Visualize distributions before and after
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
sns.histplot(df['Original'], kde=True, ax=axes[0], color='#4263eb')
axes[0].set_title('Original Data Distribution')
sns.histplot(df['Yeo-Johnson Transformed'], kde=True, ax=axes[1], color='#37b24d')
axes[1].set_title('Yeo-Johnson Transformed Data Distribution')
plt.tight_layout()
plt.show()

Running this code will likely show that the original data has significant skew, while the histogram of the transformed data appears much more symmetric and bell-shaped. The standardize=True argument (default) ensures the output also has zero mean and unit variance, combining the transformation and scaling steps.

Here's a Plotly visualization comparing the distributions:

Comparison of data distributions. The left plot shows the original skewed data containing negative values, while the right plot shows the distribution after applying the Yeo-Johnson transformation, appearing much closer to a normal distribution. Sample data is used for illustration.

When to Choose Yeo-Johnson

Handles Non-Positive Data: Its primary advantage over Box-Cox is the ability to handle features with zero or negative values naturally.
Reduces Skewness: Like other power transformations, it's effective at making skewed distributions more symmetric.
Meets Model Assumptions: Helps satisfy normality assumptions required or beneficial for certain models (e.g., linear regression residual normality).

Important Considerations

Interpretability: Applying Yeo-Johnson changes the scale and interpretation of the feature. A one-unit increase in the transformed feature does not correspond to a one-unit increase in the original feature. Model coefficients will relate to the transformed scale.
Fit on Training Data: As with all preprocessing steps, fit the PowerTransformer only on your training dataset. Use the same fitted transformer (with the learned $\lambda$ ) to apply the transformation to your validation and test sets to avoid data leakage and ensure consistency.
Standardization: Be mindful of the standardize parameter in PowerTransformer. If True (default), it applies Z-score scaling after the Yeo-Johnson transformation. If you need the raw power-transformed values without standardization, set standardize=False.

Yeo-Johnson provides a valuable and flexible tool for normalizing the distribution of numerical features, especially when those features contain the full range of real numbers, overcoming a key limitation of the Box-Cox method. It's a common technique used to prepare data for algorithms sensitive to feature distribution.