Cohorts

A cohort is a group of users who share a common starting point. Usually it's the day, week, or month they first had an event. Say we want to see how many users from each cohort are still active in later months. We can create a monthly cohort label with dt.to_period(), which converts each timestamp into a pandas Period value like 2024-01:

first_event = events.groupby("user_id")["timestamp"].min().dt.to_period("M")
first_event.name = "cohort"

Then merge it back into the events and calculate how many months have passed since each user's first event:

# Add cohort Series to the events DataFrame
events_c = events.merge(first_event.reset_index(), on="user_id")

# Calculate months since the user's first event
events_c["event_month"] = events_c["timestamp"].dt.to_period("M")
events_c["months_since"] = (
    # Subtracting two Period objects returns the time difference
    events_c["event_month"] - events_c["cohort"]
).apply(lambda x: x.n)  # .n is the number of months in the difference

cohort_table = (
    events_c.groupby(["cohort", "months_since"])["user_id"].nunique().unstack()
)
print(cohort_table)

Which prints something like this:

months_since  0  1  2
cohort
2024-01       3  2  2
2024-02       2  1  2
2024-03       1  1  NaN

Which means:

• The January cohort had 3 users. After 1 month, 2 of them were still active. After 2 months, 2 were still active.
• The February cohort had 2 users. After 1 month, 1 was still active. After 2 months, 2 were still active.
• The March cohort had 1 user. After 1 month, 1 was still active. Month 2 hasn't happened yet, so it is NaN.