FeatureTools：time_index参数是否支持指定多列时间字段？

Great question! First, a key point to clarify: in Featuretools, each entity (dataframe within an EntitySet) can only have one time_index assigned at a time. So you can't directly set both start_date and end_date as the time index for the same entity. But don't worry—there are two straightforward workarounds to get the exact features you want without merging the date columns.

Option 1: Create two separate entities in the same EntitySet #

You can add two copies of your dataframe to the EntitySet, each with a different time index. Then generate features for each entity and merge the results.

Here's a concrete code example (assuming your dataframe has a primary key column named id):

import featuretools as ft
import pandas as pd

# Sample dataframe (replace with your actual data)
df = pd.DataFrame({
    "id": [1, 1, 2, 2],
    "start_date": pd.to_datetime(["2023-01-01", "2023-02-01", "2023-01-15", "2023-03-01"]),
    "end_date": pd.to_datetime(["2023-01-10", "2023-02-10", "2023-01-25", "2023-03-10"]),
    "value": [10, 20, 15, 25]
})

# Initialize EntitySet
es = ft.EntitySet(id="my_time_data")

# Add first entity with start_date as time_index
es = es.add_dataframe(
    dataframe_name="data_with_start",
    dataframe=df,
    index="id",
    time_index="start_date"
)

# Add second entity with end_date as time_index (unique name is important)
es = es.add_dataframe(
    dataframe_name="data_with_end",
    dataframe=df,
    index="id",
    time_index="end_date"
)

# Add self-relationship if you want to calculate features per ID (e.g., time since first start_date for the same ID)
# For the start-date entity
es = es.add_relationship(
    ft.Relationship(es["data_with_start"]["id"], es["data_with_start"]["id"])
)
# For the end-date entity
es = es.add_relationship(
    ft.Relationship(es["data_with_end"]["id"], es["data_with_end"]["id"])
)

# Generate features based on start_date
features_start, _ = ft.dfs(
    entityset=es,
    target_dataframe_name="data_with_start",
    agg_primitives=["time_since_first", "time_since_last", "avg_time_between", "trend"],
    trans_primitives=[]  # Add any transformation primitives you need
)

# Generate features based on end_date
features_end, _ = ft.dfs(
    entityset=es,
    target_dataframe_name="data_with_end",
    agg_primitives=["time_since_first", "time_since_last", "avg_time_between", "trend"],
    trans_primitives=[]
)

# Merge the two feature dataframes into one
final_features = features_start.merge(features_end, on="id")

Option 2: Reassign the time index and run DFS twice #

If you don't want to duplicate entities, you can create a single entity, run DFS with one time index, then update the time index and run DFS again.

import featuretools as ft
import pandas as pd

# Using the same sample dataframe as above
df = pd.DataFrame({
    "id": [1, 1, 2, 2],
    "start_date": pd.to_datetime(["2023-01-01", "2023-02-01", "2023-01-15", "2023-03-01"]),
    "end_date": pd.to_datetime(["2023-01-10", "2023-02-10", "2023-01-25", "2023-03-10"]),
    "value": [10, 20, 15, 25]
})

# Initialize EntitySet with start_date as time_index
es = ft.EntitySet(id="my_time_data")
es = es.add_dataframe(
    dataframe_name="data",
    dataframe=df,
    index="id",
    time_index="start_date"
)

# Add self-relationship for per-ID calculations
es = es.add_relationship(
    ft.Relationship(es["data"]["id"], es["data"]["id"])
)

# Generate start-date features
features_start, _ = ft.dfs(
    entityset=es,
    target_dataframe_name="data",
    agg_primitives=["time_since_first", "time_since_last", "avg_time_between", "trend"]
)

# Update the time index to end_date
es["data"].time_index = "end_date"

# Generate end-date features
features_end, _ = ft.dfs(
    entityset=es,
    target_dataframe_name="data",
    agg_primitives=["time_since_first", "time_since_last", "avg_time_between", "trend"]
)

# Merge results
final_features = features_start.merge(features_end, on="id")

Key Notes: #

• Self-Relationship: Adding a self-relationship (linking the entity to itself via the primary key) is crucial if you want to calculate features like time_since_first_start_date for each unique id—it tells Featuretools to group data by the ID when computing time-based aggregates.
• Feature Naming: Featuretools will automatically name features to reflect which time index they use (e.g., TIME_SINCE_FIRST(data_with_start.start_date)), so you can easily distinguish between start-date and end-date derived features.
• Cutoff Times: If you're working with temporal validation (using cutoff_times), make sure to adjust the cutoff times to align with whichever time index you're using for each DFS run.

内容的提问来源于stack exchange，提问作者Ganesh Bhat