Deep Feature Synthesis

Deep Feature Synthesis (DFS) is an automated method for performing feature engineering on relational and temporal data.

Input Data

Deep Feature Synthesis requires structured datasets in order to perform feature engineering. To demonstrate the capabilities of DFS, we will use a mock customer transactions dataset.

Note

Before using DFS, it is recommended that you prepare your data as an EntitySet. See Representing Data with EntitySets to learn how.

[1]:
import featuretools as ft
es = ft.demo.load_mock_customer(return_entityset=True)
es
[1]:
Entityset: transactions
  DataFrames:
    transactions [Rows: 500, Columns: 6]
    products [Rows: 5, Columns: 3]
    sessions [Rows: 35, Columns: 5]
    customers [Rows: 5, Columns: 5]
  Relationships:
    transactions.product_id -> products.product_id
    transactions.session_id -> sessions.session_id
    sessions.customer_id -> customers.customer_id

Once data is prepared as an .EntitySet, we are ready to automatically generate features for a target dataframe - e.g. customers.

Running DFS

Typically, without automated feature engineering, a data scientist would write code to aggregate data for a customer, and apply different statistical functions resulting in features quantifying the customer’s behavior. In this example, an expert might be interested in features such as: total number of sessions or month the customer signed up.

These features can be generated by DFS when we specify the target_dataframe as customers and "count" and "month" as primitives.

[2]:
feature_matrix, feature_defs = ft.dfs(entityset=es,
                                      target_dataframe_name="customers",
                                      agg_primitives=["count"],
                                      trans_primitives=["month"],
                                      max_depth=1)
feature_matrix
[2]:
zip_code COUNT(sessions) MONTH(birthday) MONTH(join_date)
customer_id
5 60091 6 7 7
4 60091 8 8 4
1 60091 8 7 4
3 13244 6 11 8
2 13244 7 8 4

In the example above, "count" is an aggregation primitive because it computes a single value based on many sessions related to one customer. "month" is called a transform primitive because it takes one value for a customer transforms it to another.

Note

Feature primitives are a fundamental component to Featuretools. To learn more read Feature primitives.

Creating “Deep Features”

The name Deep Feature Synthesis comes from the algorithm’s ability to stack primitives to generate more complex features. Each time we stack a primitive we increase the “depth” of a feature. The max_depth parameter controls the maximum depth of the features returned by DFS. Let us try running DFS with max_depth=2

[3]:
feature_matrix, feature_defs = ft.dfs(entityset=es,
                                      target_dataframe_name="customers",
                                      agg_primitives=["mean", "sum", "mode"],
                                      trans_primitives=["month", "hour"],
                                      max_depth=2)
feature_matrix
[3]:
zip_code MODE(sessions.device) MEAN(transactions.amount) MODE(transactions.product_id) SUM(transactions.amount) HOUR(birthday) HOUR(join_date) MONTH(birthday) MONTH(join_date) MEAN(sessions.MEAN(transactions.amount)) MEAN(sessions.SUM(transactions.amount)) MODE(sessions.HOUR(session_start)) MODE(sessions.MODE(transactions.product_id)) MODE(sessions.MONTH(session_start)) SUM(sessions.MEAN(transactions.amount)) MODE(transactions.sessions.device)
customer_id
5 60091 mobile 80.375443 5 6349.66 0 5 7 7 78.705187 1058.276667 0 3 1 472.231119 mobile
4 60091 mobile 80.070459 2 8727.68 0 20 8 4 81.207189 1090.960000 1 1 1 649.657515 mobile
1 60091 mobile 71.631905 4 9025.62 0 10 7 4 72.774140 1128.202500 6 4 1 582.193117 mobile
3 13244 desktop 67.060430 1 6236.62 0 15 11 8 67.539577 1039.436667 5 1 1 405.237462 desktop
2 13244 desktop 77.422366 4 7200.28 0 23 8 4 78.415122 1028.611429 3 3 1 548.905851 desktop

With a depth of 2, a number of features are generated using the supplied primitives. The algorithm to synthesize these definitions is described in this paper. In the returned feature matrix, let us understand one of the depth 2 features

[4]:
feature_matrix[['MEAN(sessions.SUM(transactions.amount))']]
[4]:
MEAN(sessions.SUM(transactions.amount))
customer_id
5 1058.276667
4 1090.960000
1 1128.202500
3 1039.436667
2 1028.611429

For each customer this feature

  1. calculates the sum of all transaction amounts per session to get total amount per session,

  2. then applies the mean to the total amounts across multiple sessions to identify the average amount spent per session

We call this feature a “deep feature” with a depth of 2.

Let’s look at another depth 2 feature that calculates for every customer the most common hour of the day when they start a session

[5]:
feature_matrix[['MODE(sessions.HOUR(session_start))']]
[5]:
MODE(sessions.HOUR(session_start))
customer_id
5 0
4 1
1 6
3 5
2 3

For each customer this feature calculates

  1. The hour of the day each of his or her sessions started, then

  2. uses the statistical function mode to identify the most common hour he or she started a session

Stacking results in features that are more expressive than individual primitives themselves. This enables the automatic creation of complex patterns for machine learning.

Note

You can graphically visualize the lineage of a feature by calling featuretools.graph_feature() on it. You can also generate an English description of the feature with featuretools.describe_feature(). See Generating Feature Descriptions for more details.

Changing Target DataFrame

DFS is powerful because we can create a feature matrix for any dataframe in our dataset. If we switch our target dataframe to “sessions”, we can synthesize features for each session instead of each customer. Now, we can use these features to predict the outcome of a session.

[6]:
feature_matrix, feature_defs = ft.dfs(entityset=es,
                                      target_dataframe_name="sessions",
                                      agg_primitives=["mean", "sum", "mode"],
                                      trans_primitives=["month", "hour"],
                                      max_depth=2)
feature_matrix.head(5)
[6]:
customer_id device MEAN(transactions.amount) MODE(transactions.product_id) SUM(transactions.amount) HOUR(session_start) MONTH(session_start) customers.zip_code MODE(transactions.HOUR(transaction_time)) MODE(transactions.MONTH(transaction_time)) customers.MODE(sessions.device) customers.MEAN(transactions.amount) customers.MODE(transactions.product_id) customers.SUM(transactions.amount) customers.HOUR(birthday) customers.HOUR(join_date) customers.MONTH(birthday) customers.MONTH(join_date)
session_id
1 2 desktop 76.813125 3 1229.01 0 1 13244 0 1 desktop 77.422366 4 7200.28 0 23 8 4
2 5 mobile 74.696000 5 746.96 0 1 60091 0 1 mobile 80.375443 5 6349.66 0 5 7 7
3 4 mobile 88.600000 1 1329.00 0 1 60091 0 1 mobile 80.070459 2 8727.68 0 20 8 4
4 1 mobile 64.557200 5 1613.93 0 1 60091 0 1 mobile 71.631905 4 9025.62 0 10 7 4
5 4 mobile 70.638182 5 777.02 1 1 60091 1 1 mobile 80.070459 2 8727.68 0 20 8 4

As we can see, DFS will also build deep features based on a parent dataframe, in this case the customer of a particular session. For example, the feature below calculates the mean transaction amount of the customer of the session.

[7]:
feature_matrix[['customers.MEAN(transactions.amount)']].head(5)
[7]:
customers.MEAN(transactions.amount)
session_id
1 77.422366
2 80.375443
3 80.070459
4 71.631905
5 80.070459

Improve feature output

To learn about the parameters to change in DFS read Tuning Deep Feature Synthesis.