Using Featuretools to aggregate per time time of day - feature-extraction

Question

I'm wondering if there's any way to calculate all the same variables I already am using deep feature synthesis (ie counts, sums, mean, etc) for different time segments within a day?
I.e. count of morning events (hours 0-12) as a separate variable from evening events (13-24).
Also, within the same vein, what would be the easiest to get counts by day of week, day of month, day of year, etc. Custom aggregate primitives?

Answer 1

Yes, this is possible. First, let's generate some random data and then I'll walkthrough how
import featuretools as ft
import pandas as pd
import numpy as np
# make some random data
n = 100
events_df = pd.DataFrame({
"id" : range(n),
"customer_id": np.random.choice(["a", "b", "c"], n),
"timestamp": pd.date_range("Jan 1, 2019", freq="1h", periods=n),
"amount": np.random.rand(n) * 100
})
def to_part_of_day(x):
if x < 12:
return "morning"
elif x < 18:
return "afternoon"
else:
return "evening"
events_df["time_of_day"] = events_df["timestamp"].dt.hour.apply(to_part_of_day)
events_df
the first thing we want to do is add a new column for the segment we want to calculate features for
def to_part_of_day(x):
if x < 12:
return "morning"
elif x < 18:
return "afternoon"
else:
return "evening"
events_df["time_of_day"] = events_df["timestamp"].dt.hour.apply(to_part_of_day)
now we have a dataframe like this
id customer_id timestamp amount time_of_day
0 0 a 2019-01-01 00:00:00 44.713802 morning
1 1 c 2019-01-01 01:00:00 58.776476 morning
2 2 a 2019-01-01 02:00:00 94.671566 morning
3 3 a 2019-01-01 03:00:00 39.271852 morning
4 4 a 2019-01-01 04:00:00 40.773290 morning
5 5 c 2019-01-01 05:00:00 19.815855 morning
6 6 a 2019-01-01 06:00:00 62.457129 morning
7 7 b 2019-01-01 07:00:00 95.114636 morning
8 8 b 2019-01-01 08:00:00 37.824668 morning
9 9 a 2019-01-01 09:00:00 46.502904 morning
Next, let's load it into our entityset
es = ft.EntitySet()
es.entity_from_dataframe(entity_id="events",
time_index="timestamp",
dataframe=events_df)
es.normalize_entity(new_entity_id="customers", index="customer_id", base_entity_id="events")
es.plot()
Now, we are ready to set the segments we want to create aggregations for by using interesting_values
es["events"]["time_of_day"].interesting_values = ["morning", "afternoon", "evening"]
Then we can run DFS and place the aggregation primitives we want to do on a per segment basis in the where_primitives parameter
fm, fl = ft.dfs(target_entity="customers",
entityset=es,
agg_primitives=["count", "mean", "sum"],
trans_primitives=[],
where_primitives=["count", "mean", "sum"])
fm
In the resulting feature matrix, you can now see we have aggregations per morning, afternoon, and evening
COUNT(events) MEAN(events.amount) SUM(events.amount) COUNT(events WHERE time_of_day = afternoon) COUNT(events WHERE time_of_day = evening) COUNT(events WHERE time_of_day = morning) MEAN(events.amount WHERE time_of_day = afternoon) MEAN(events.amount WHERE time_of_day = evening) MEAN(events.amount WHERE time_of_day = morning) SUM(events.amount WHERE time_of_day = afternoon) SUM(events.amount WHERE time_of_day = evening) SUM(events.amount WHERE time_of_day = morning)
customer_id
a 37 49.753630 1840.884300 12 7 18 35.098923 45.861881 61.036892 421.187073 321.033164 1098.664063
b 30 51.241484 1537.244522 3 10 17 45.140800 46.170996 55.300715 135.422399 461.709963 940.112160
c 33 39.563222 1305.586314 9 7 17 50.129136 34.593936 36.015679 451.162220 242.157549 612.266545