The book Hands-On Machine Learning has a section on Out-of-Bag Evaluation related to Decision Trees, where it's stated that,

By default a BaggingClassifier samples m training instances with replacement (bootstrap=True), where m is the size of the training set. This means that only about 63% of the training instances are sampled on average for each predictor.

I am curious how they arrived at 63%. Here's what I have so far,

Let $X$ be a random variable representing the fraction of the training set sampled with replacement with size m. Hence, $X$ takes values $\frac{1}{m}, \frac{2}{m}, ..., 1$ Then, the PMF is as follows,

$$ P(X=\frac{i}{m}) = {m\choose i} * \frac{1}{m^m} $$

Then, the average number of training samples with replacement will be,

$$ E[X] = \sum_{i=1}^{m} {m\choose i} * \frac{1}{m^m} * \frac{i}{m} $$

Is this the right way to derive the 63%? I am not fully convinced the PMF is correct because the PMF doesn't seem to sum to 1.

$$ \sum_{i=1}^{m}P(X=\frac{i}{m}) = (\frac{2}{m})^m - \frac{1}{m^m} = \frac{2^m-1}{m^m} \text{(using Binomial Theorem)} $$


1 Answer 1


Here the author explains:

If you randomly draw one instance from a dataset of size $m$, each instance in the dataset obviously has probability $\frac{1}{m}$ of getting picked, and therefore it has a probability $1 – \frac{1}{m}$ of not getting picked.

If you draw $m$ instances with replacement, all draws are independent and therefore each instance has a probability $(1 – \frac{1}{m})^m$ of not getting picked. Now let's use the fact that $e^x$ is equal to the limit of $(1 + \frac{x}{m})^m$ as $m$ approaches infinity.

So if $m$ is large, the ratio of out-of-bag instances will be about $e^{–1}\approx0.37$. So roughly $63\%$ $(1 – 0.37)$ will be sampled.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.