Skip to content

Verify Point-in-Time Correctness

Time leakage occurs when a feature uses information that would not have been available at prediction time. GraphReduce prevents this by applying separate historical and future windows around a cutoff.

Window contract

For cutoff T:

features: T - compute period < timestamp < T
labels:   T                  < timestamp < T + label period

For non-null timestamps, the comparisons are strict. Rows on any boundary are excluded.

Inspect one node

Run this example from the repository root:

import datetime as dt
from pathlib import Path

import pandas as pd

from graphreduce.enum import ComputeLayerEnum, PeriodUnit
from graphreduce.node import DynamicNode

orders = DynamicNode(
    fpath=str(Path("tests/data/cust_data/orders.csv")),
    fmt="csv",
    pk="id",
    prefix="ord",
    date_key="ts",
    compute_layer=ComputeLayerEnum.pandas,
    cut_date=dt.datetime(2023, 7, 1),
    compute_period_val=365,
    compute_period_unit=PeriodUnit.day,
    label_period_val=60,
    label_period_unit=PeriodUnit.day,
)

orders.do_data()

feature_rows = orders.prep_for_features()
label_rows = orders.prep_for_labels()

print(feature_rows[["ord_id", "ord_ts"]])
print(label_rows[["ord_id", "ord_ts"]])

feature_times = pd.to_datetime(feature_rows["ord_ts"])
label_times = pd.to_datetime(label_rows["ord_ts"])
cutoff = pd.Timestamp("2023-07-01")

assert (feature_times < cutoff).all()
assert (label_times > cutoff).all()

do_data() loads the CSV and adds the ord_ prefix. The two helper methods return filtered frames; they do not mix feature and label rows. The explicit pd.to_datetime() calls above normalize the result before the verification assertions.

Configure the whole graph

Set shared time controls on GraphReduce:

graph = GraphReduce(
    parent_node=customers,
    compute_layer=ComputeLayerEnum.pandas,
    cut_date=dt.datetime(2023, 7, 1),
    compute_period_val=365,
    compute_period_unit=PeriodUnit.day,
    label_node=orders,
    label_field="id",
    label_operation="count",
    label_period_val=60,
    label_period_unit=PeriodUnit.day,
    auto_features=True,
)

During graph hydration, these values are pushed to the nodes so one execution uses a consistent reference time.

Checklist before training

  • Give every time-varying node the correct date_key.
  • Confirm timestamps represent availability time, not merely when an event was later recorded or updated.
  • Treat static dimensions without a date_key carefully if their attributes can change over time.
  • Decide whether rows with null timestamps should be retained. The dataframe feature helper currently retains them, while SQL comparisons normally do not.
  • Test both helper windows on a small known sample.
  • Assert the output grain after all joins.
  • Reserve validation and test cutoffs that are later than training cutoffs.

GraphReduce can enforce configured timestamps; it cannot infer whether a source column itself contains future-derived information. Declare or remove such columns before feature generation. See Feature schema profiles for explicit unsafe-column handling.