Mean Encoding

Mean Encoding


  • also called target encoding and likelihood encoding

It is a way to encode a categorical feature. Uses the fraction of times the feature is 1 out of all the times the feature is in the data set (for binary classification).

Why does it work?

Unlike regular encoding, which has no real meaning to the labels, mean encoding imposes an ordering. This allows you to reduce your loss while using shorter trees.

How do you calculate it?

There are multiple ways.

  • Likelihood = \(\frac{count of ones}/{total count}\) = mean(target)
  • Weight of evidence = \(\ln\left(\frac{count of ones}{count of zeros}\right)\)
  • Count = sum(target) = count of ones
  • Diff = count of ones - count of zeros

When does it fail?

If you have lots of feature instances with few cases it will tend to overfit.


Four Types

  • Cross-validation loop inside the training data
  • Smoothing
  • Adding random noise
  • Sorting and calculating the expanding mean

Cross Validation

  • Usually 4 or 5 folds are enough
  • Need to watch out for extreme cases like leave-out-one (LOO)

Here's an example of this method using sklearn.

y_train = training["target"].values
folds = StratifiedKFold(y_train, 5, shuffle=True)

for training_index, validation_index in folds:
    x_train = training.iloc[training_index]
    x_validation = training.iloc[validation_index]
    # 'columns' is a list of columns to encode
    for column in columns:
	means = x_validation[column].map(x_train.groupby(column).target.mean())
	x_validation[coulmn + "_mean_target"] = means
    # train_new is a dataframe copy we made of the training data
    train_new.iloc[value_index] = x_validation

global_mean = training["target"].mean()

# replace nans with the global mean
train_new.fillna(global_mean, inplace=True)


Use a value \(\alpha\) to control the amount of regularization. This isn't a regularization method in and of itself, you use it with other methods.

\[ \frac{mean(targte) \times n_{rows} + \textit{global mean} \times \alpha}{n_{rows} + \alpha} \]


Adding noise degrades the quality of the encoding in the training data. This is usually used with leave-one-out encoding to prevent overfitting. You have to figure out how much noise to add through experimentation.

Expanding Mean

This introduces the least amount of leakage from the target variable and doesn't require hyper-parameters for you to tune. The downside is that the encoding quality is irregular. There is a built-in implementation in the CatBoost library.

Here's a pandas implementation.

cumulative_sum = training.groupby(column)["target"].cumsum() - training["target"]
cumulative_count = training.groupby(column).cumcount()
train_new[column + "_mean_target"] = cumulative_sum/cumulative_count

Which one should you use?

Cross Validation Loops and Expanding Means are the most practical to use.

Generalizations and Extensions

Regression and Multiclass



  • Compact transformation of categorical variables
  • Powerful basis for feature engineering


  • Needs careful validation, it's easy to overfit
  • Only certain data sets will show a significant improvement from using it