jstark 0.2.0

No project description provided

jstark

Built with

---

A PySpark library for generating time-based features for machine learning. All features are calculated relative to an as at date, enabling point-in-time feature engineering over configurable time periods.

Feature period mnemonics

Feature names end with a mnemonic describing the time window. The format is {start}{unit}{end} where the unit is one of d (days), w (weeks), m (months), q (quarters) or y (years).

For example, BasketCount_3m1 is the distinct count of baskets from 3 months before to 1 month before the as at date.

Multiple periods can be calculated in a single Spark job:

from datetime import date
from jstark.grocery import GroceryFeatures

gf = GroceryFeatures(as_at=date(2022, 1, 1), feature_periods=["3m1", "6m4"])
output_df = input_df.groupBy("Store").agg(*gf.features)

This produces BasketCount_3m1, BasketCount_6m4, and every other feature for both periods. See the Features reference for a list of all available features.

Quick start

Prerequisites: Java runtime required for PySpark. On macOS: brew install openjdk@11.

pip install jstark[faker]

The faker extra installs Faker, which is needed for the sample data generator used below. If you don't need sample data, pip install jstark is sufficient.

from datetime import date
from jstark.sample.transactions import FakeGroceryTransactions
from jstark.grocery import GroceryFeatures

input_df = FakeGroceryTransactions().df
gf = GroceryFeatures(date(2022, 1, 1), ["4q4", "3q3", "2q2", "1q1"])
output_df = input_df.groupBy("Store").agg(*gf.features)
output_df.select(
    "Store", "BasketCount_4q4", "BasketCount_3q3", "BasketCount_2q2", "BasketCount_1q1"
).show()

+-----------+---------------+---------------+---------------+---------------+
|      Store|BasketCount_4q4|BasketCount_3q3|BasketCount_2q2|BasketCount_1q1|
+-----------+---------------+---------------+---------------+---------------+
|    Staines|             47|             46|             48|             51|
| Twickenham|             55|             57|             48|             49|
|     Ealing|             52|             51|             50|             54|
|Hammersmith|             47|             40|             43|             51|
|   Richmond|             54|             40|             64|             53|
+-----------+---------------+---------------+---------------+---------------+

Feature descriptions and references

Every feature carries a description in its column metadata:

from pprint import pprint
pprint([(c.name, c.metadata["description"]) for c in output_df.schema if c.name.endswith("1q1")])

[('BasketCount_1q1',
  'Distinct count of Baskets between 2021-10-01 and 2021-12-31'),
 ...]

You can also inspect what input columns each feature requires:

gf.references["BasketCount_1q1"]                   # ['Basket', 'Timestamp']
gf.references["CustomerCount_1q1"]                 # ['Customer', 'Timestamp']
gf.references["AvgGrossSpendPerBasket_1q1"]        # ['Basket', 'GrossSpend', 'Timestamp']

All features require a Timestamp column (TimestampType). Most require additional columns depending on what they measure.

Features reference

Grocery features

A list of all Grocery features available if one were to call:

GroceryFeatures(date(2022, 1, 1), ["3m1"])

Feature	Description
ApproxBasketCount_3m1	Approximate distinct count of Baskets between 2021-10-01 and 2021-12-31
ApproxCustomerCount_3m1	Approximate distinct count of Customers between 2021-10-01 and 2021-12-31
AverageBasketsPerMonth_3m1	Average number of baskets per month between 2021-10-01 and 2021-12-31
AvgDiscountPerBasket_3m1	Average Discount per Basket between 2021-10-01 and 2021-12-31
AvgGrossSpendPerBasket_3m1	Average GrossSpend per Basket between 2021-10-01 and 2021-12-31
AvgPurchaseCycle_3m1	Average purchase cycle between 2021-10-01 and 2021-12-31
AvgQuantityPerBasket_3m1	Average Quantity per Basket between 2021-10-01 and 2021-12-31
BasketCount_3m1	Distinct count of Baskets between 2021-10-01 and 2021-12-31
BasketMonths_3m1	Number of months in which at least one basket was purchased between 2021-10-01 and 2021-12-31
ChannelCount_3m1	Distinct count of Channels between 2021-10-01 and 2021-12-31
Count_3m1	Count of rows between 2021-10-01 and 2021-12-31
CustomerCount_3m1	Distinct count of Customers between 2021-10-01 and 2021-12-31
CyclesSinceLastPurchase_3m1	Cycles since last purchase between 2021-10-01 and 2021-12-31
Discount_3m1	Sum of Discount between 2021-10-01 and 2021-12-31
EarliestPurchaseDate_3m1	Earliest purchase date between 2021-10-01 and 2021-12-31
GrossSpend_3m1	Sum of GrossSpend between 2021-10-01 and 2021-12-31
MaxGrossPrice_3m1	Maximum of (GrossSpend / Quantity) between 2021-10-01 and 2021-12-31
MaxGrossSpend_3m1	Maximum GrossSpend value between 2021-10-01 and 2021-12-31
MaxNetPrice_3m1	Maximum of (NetSpend / Quantity) between 2021-10-01 and 2021-12-31
MaxNetSpend_3m1	Maximum of NetSpend value between 2021-10-01 and 2021-12-31
MinGrossPrice_3m1	Minimum of (GrossSpend / Quantity) between 2021-10-01 and 2021-12-31
MinGrossSpend_3m1	Minimum GrossSpend value between 2021-10-01 and 2021-12-31
MinNetPrice_3m1	Minimum of (NetSpend / Quantity) between 2021-10-01 and 2021-12-31
MinNetSpend_3m1	Minimum of NetSpend value between 2021-10-01 and 2021-12-31
MostRecentPurchaseDate_3m1	Most recent purchase date between 2021-10-01 and 2021-12-31
NetSpend_3m1	Sum of NetSpend between 2021-10-01 and 2021-12-31
ProductCount_3m1	Distinct count of Products between 2021-10-01 and 2021-12-31
Quantity_3m1	Sum of Quantity between 2021-10-01 and 2021-12-31
RecencyDays_3m1	Minimum number of days since occurrence between 2021-10-01 and 2021-12-31
RecencyWeightedApproxBasketMonths90_3m1	Exponentially weighted moving average, with smoothing factor of 0.9, of the approximate number of baskets per month between 2021-10-01 and 2021-12-31
RecencyWeightedApproxBasketMonths95_3m1	Exponentially weighted moving average, with smoothing factor of 0.95, of the approximate number of baskets per month between 2021-10-01 and 2021-12-31
RecencyWeightedApproxBasketMonths99_3m1	Exponentially weighted moving average, with smoothing factor of 0.99, of the approximate number of baskets per month between 2021-10-01 and 2021-12-31
RecencyWeightedBasketMonths90_3m1	Exponentially weighted moving average, with smoothing factor of 0.9, of the number of baskets per month between 2021-10-01 and 2021-12-31
RecencyWeightedBasketMonths95_3m1	Exponentially weighted moving average, with smoothing factor of 0.95, of the number of baskets per month between 2021-10-01 and 2021-12-31
RecencyWeightedBasketMonths99_3m1	Exponentially weighted moving average, with smoothing factor of 0.99, of the number of baskets per month between 2021-10-01 and 2021-12-31
StoreCount_3m1	Distinct count of Stores between 2021-10-01 and 2021-12-31

License

jstark is distributed under the terms of the MIT license.

Why "jstark"?

The name is phonetically similar to PySpark, is a homage to comic book character Jon Stark, and contains the initials of the original contributor (j, k & t).