Errata

In Chapter 4, Acquire an initial Dataset. There is an error on the code example.

********************************************************************************************
questions_with_accepted_answers = df[
df["is_question"] & ~(df["AcceptedAnswerId"].isna())
]
q_and_a = questions_with_accepted_answers.join(
df[["Text"]], on="AcceptedAnswerId", how="left", rsuffix="_answer"
)
*******************************************************************************************

df[['Text']] shall have been df[['body_text']]. There is no 'Text' in the df.info().

Note from the Author or Editor:
df[["Text"]] should be changed to df[["body_text"]]

"five" in five most popular ones should be "seven"

Because we have more than three hundred tags in our dataset, here we chose to only create a column for the five most popular ones

The text on last paragraph of page 86 describes crossing features as multiplying them, this means on the table in page 87, i expect DoW x DoM = Cross column. However DoW in the table is 7,7,...,1 down the rows. It should be 6,6,...7. Only then when multiplied by DoM of 29,29,...,30 gives 174,174,...210.

Note from the Author or Editor:
This is correct, there is a slight inaccuracy in the figure. It currently reads

7 | 29 | 174
7 | 29 | 174
...
1 | 30 | 210

It should be

6 | 29 | 174
6 | 29 | 174
...
7 | 30 | 210

Text explains calibration curve as "gives a probability of being classified as positive that is higher than 80%".

This makes students think the upper bound is 100%, whereas in reality, the intuitive understanding is a bucket of size 10% and the y-axis is calculated from observations within a bucket of predicted probability of 80-90% and not the 80-100% as the language (higher than 80%) in the book implies.

The above error would be easier to identify if a 10% example was given, then it would be obviously wrong that only 10% of the observations with predicted probability of 10-100% are actually correct. Anyway, this was how i reasoned that something is wrong before actually researching more to find a better definition. I'm actually not sure what the upper bound or bucket size should be, maybe 80-90% is still not narrow enough for a bucket.

Note from the Author or Editor:
This is an imprecise use of language here. The paragraph currently reads:

"For example, out of all the data points our classifier gives a probability of being classified as positive that is higher than 80%, how many of those data points are actually positive?"

It should read:

"For example, out of all the data points our classifier gives a probability of being classified as positive that is close to 80%, how many of those data points are actually positive?"