Tuning Deep Feature Synthesis¶
There are several parameters that can be tuned to change the output of DFS. We’ll explore these parameters using the following transactions
EntitySet.
[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
Using “Seed Features”¶
Seed features are manually defined and problem specific features that a user provides to DFS. Deep Feature Synthesis will then automatically stack new features on top of these features when it can.
By using seed features, we can include domain specific knowledge in feature engineering automation. For the seed feature below, the domain knowlege may be that, for a specific retailer, a transaction above $125 would be considered an expensive purchase.
[2]:
expensive_purchase = ft.Feature(es["transactions"].ww["amount"]) > 125
feature_matrix, feature_defs = ft.dfs(entityset=es,
target_dataframe_name="customers",
agg_primitives=["percent_true"],
seed_features=[expensive_purchase])
feature_matrix[['PERCENT_TRUE(transactions.amount > 125)']]
[2]:
PERCENT_TRUE(transactions.amount > 125) | |
---|---|
customer_id | |
5 | 0.227848 |
4 | 0.220183 |
1 | 0.119048 |
3 | 0.182796 |
2 | 0.129032 |
We can now see that the PERCENT_TRUE
primitive was automatically applied to the boolean expensive_purchase
feature from the transactions
table. The feature produced as a result can be understood as the percentage of transactions for a customer that are considered expensive.
Add “interesting” values to columns¶
Sometimes we want to create features that are conditioned on a second value before calculations are performed. We call this extra filter a “where clause”. Where clauses are used in Deep Feature Synthesis by including primitives in the where_primitives
parameter to DFS.
By default, where clauses are built using the interesting_values
of a column.
Interesting values can be automatically determined and added for each DataFrame in a pandas EntitySet by calling es.add_interesting_values()
.
Note that Dask and Spark EntitySets cannot have interesting values determined automatically for their DataFrames. For those EntitySets, or when interesting values are already known for columns, the dataframe_name
and values
parameters can be used to set interesting values for individual columns in a DataFrame in an EntitySet.
[3]:
values_dict = {'device': ["desktop", "mobile", "tablet"]}
es.add_interesting_values(dataframe_name='sessions', values=values_dict)
Interesting values are stored in the DataFrame’s Woodwork typing information.
[4]:
es['sessions'].ww.columns['device'].metadata
[4]:
{'dataframe_name': 'sessions',
'entityset_id': 'transactions',
'interesting_values': ['desktop', 'mobile', 'tablet']}
Now that interesting values are set for the device
column in the sessions
table, we can specify the aggregation primitives for which we want where clauses using the where_primitives
parameter to DFS.
[5]:
feature_matrix, feature_defs = ft.dfs(entityset=es,
target_dataframe_name="customers",
agg_primitives=["count", "avg_time_between"],
where_primitives=["count", "avg_time_between"],
trans_primitives=[])
feature_matrix
[5]:
zip_code | AVG_TIME_BETWEEN(sessions.session_start) | COUNT(sessions) | AVG_TIME_BETWEEN(transactions.transaction_time) | COUNT(transactions) | AVG_TIME_BETWEEN(sessions.session_start WHERE device = desktop) | AVG_TIME_BETWEEN(sessions.session_start WHERE device = mobile) | AVG_TIME_BETWEEN(sessions.session_start WHERE device = tablet) | COUNT(sessions WHERE device = desktop) | COUNT(sessions WHERE device = mobile) | ... | AVG_TIME_BETWEEN(transactions.sessions.session_start) | AVG_TIME_BETWEEN(transactions.sessions.session_start WHERE sessions.device = tablet) | AVG_TIME_BETWEEN(transactions.sessions.session_start WHERE sessions.device = desktop) | AVG_TIME_BETWEEN(transactions.sessions.session_start WHERE sessions.device = mobile) | AVG_TIME_BETWEEN(transactions.transaction_time WHERE sessions.device = tablet) | AVG_TIME_BETWEEN(transactions.transaction_time WHERE sessions.device = desktop) | AVG_TIME_BETWEEN(transactions.transaction_time WHERE sessions.device = mobile) | COUNT(transactions WHERE sessions.device = tablet) | COUNT(transactions WHERE sessions.device = desktop) | COUNT(transactions WHERE sessions.device = mobile) | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
customer_id | |||||||||||||||||||||
5 | 60091 | 5577.000000 | 6 | 363.333333 | 79 | 9685.0 | 13942.500000 | NaN | 2 | 3 | ... | 357.500000 | 0.000000 | 345.892857 | 796.714286 | 65.000000 | 376.071429 | 809.714286 | 14 | 29 | 36 |
4 | 60091 | 2516.428571 | 8 | 168.518519 | 109 | 4127.5 | 3336.666667 | NaN | 3 | 4 | ... | 163.101852 | 0.000000 | 223.108108 | 192.500000 | 65.000000 | 238.918919 | 206.250000 | 18 | 38 | 53 |
1 | 60091 | 3305.714286 | 8 | 192.920000 | 126 | 7150.0 | 11570.000000 | 8807.5 | 2 | 3 | ... | 185.120000 | 419.404762 | 275.000000 | 420.727273 | 442.619048 | 302.500000 | 438.454545 | 43 | 27 | 56 |
3 | 13244 | 5096.000000 | 6 | 287.554348 | 93 | 4745.0 | NaN | NaN | 4 | 1 | ... | 276.956522 | 0.000000 | 233.360656 | 0.000000 | 65.000000 | 251.475410 | 65.000000 | 15 | 62 | 16 |
2 | 13244 | 4907.500000 | 7 | 328.532609 | 93 | 6890.0 | 1690.000000 | 5330.0 | 3 | 2 | ... | 320.054348 | 197.407407 | 417.575758 | 56.333333 | 226.296296 | 435.303030 | 82.333333 | 28 | 34 | 31 |
5 rows × 21 columns
Now, we have several new potentially useful features. Here are two of them that are built off of the where clause “where the device used was a tablet”:
[6]:
feature_matrix[["COUNT(sessions WHERE device = tablet)", "AVG_TIME_BETWEEN(sessions.session_start WHERE device = tablet)"]]
[6]:
COUNT(sessions WHERE device = tablet) | AVG_TIME_BETWEEN(sessions.session_start WHERE device = tablet) | |
---|---|---|
customer_id | ||
5 | 1 | NaN |
4 | 1 | NaN |
1 | 3 | 8807.5 |
3 | 1 | NaN |
2 | 2 | 5330.0 |
The first geature, COUNT(sessions WHERE device = tablet)
, can be understood as indicating how many sessions a customer completed on a tablet.
The second feature, AVG_TIME_BETWEEN(sessions.session_start WHERE device = tablet)
, calculates the time between those sessions.
We can see that customer who only had 0 or 1 sessions on a tablet had NaN
values for average time between such sessions.
Encoding categorical features¶
Machine learning algorithms typically expect all numeric data or data that has defined numeric representations, like boolean values corresponding to 0
and 1
. When Deep Feature Synthesis generates categorical features, we can encode them using Featureools.
[7]:
feature_matrix, feature_defs = ft.dfs(entityset=es,
target_dataframe_name="customers",
agg_primitives=["mode"],
trans_primitives=['time_since'],
max_depth=1)
feature_matrix
[7]:
zip_code | MODE(sessions.device) | TIME_SINCE(birthday) | TIME_SINCE(join_date) | |
---|---|---|---|---|
customer_id | ||||
5 | 60091 | mobile | 1.196810e+09 | 3.772861e+08 |
4 | 60091 | mobile | 5.010305e+08 | 3.543373e+08 |
1 | 60091 | mobile | 8.821409e+08 | 3.535932e+08 |
3 | 13244 | desktop | 5.872577e+08 | 3.433804e+08 |
2 | 13244 | desktop | 1.131923e+09 | 3.220979e+08 |
This feature matrix contains 2 columns that are categorical in nature, zip_code
and MODE(sessions.device)
. We can use the feature matrix and feature definitions to encode these categorical values into boolean values. Featuretools offers functionality to apply one hot encoding to the output of DFS.
[8]:
feature_matrix_enc, features_enc = ft.encode_features(feature_matrix, feature_defs)
feature_matrix_enc
[8]:
TIME_SINCE(birthday) | TIME_SINCE(join_date) | zip_code = 60091 | zip_code = 13244 | zip_code is unknown | MODE(sessions.device) = mobile | MODE(sessions.device) = desktop | MODE(sessions.device) is unknown | |
---|---|---|---|---|---|---|---|---|
customer_id | ||||||||
5 | 1.196810e+09 | 3.772861e+08 | True | False | False | True | False | False |
4 | 5.010305e+08 | 3.543373e+08 | True | False | False | True | False | False |
1 | 8.821409e+08 | 3.535932e+08 | True | False | False | True | False | False |
3 | 5.872577e+08 | 3.433804e+08 | False | True | False | False | True | False |
2 | 1.131923e+09 | 3.220979e+08 | False | True | False | False | True | False |
The returned feature matrix is now encoded in a way that is interpretable to machine learning algorithms. Notice how the columns that did not need encoding are still included. Additionally, we get a new set of feature definitions that contain the encoded values.
[9]:
features_enc
[9]:
[<Feature: zip_code = 60091>,
<Feature: zip_code = 13244>,
<Feature: zip_code is unknown>,
<Feature: MODE(sessions.device) = mobile>,
<Feature: MODE(sessions.device) = desktop>,
<Feature: MODE(sessions.device) is unknown>,
<Feature: TIME_SINCE(birthday)>,
<Feature: TIME_SINCE(join_date)>]
These features can be used to calculate the same encoded values on new data. For more information on feature engineering in production, read the Deployment guide.