Time Aggregation Features for XGBoost Models

TL;DR

The paper tackles time-aware feature engineering for CTR prediction by evaluating no-lookahead time aggregation features within XGBoost on the Avazu dataset. It compares a strong time-aware target encoding baseline to models augmented with entity-history windows across various window designs, finding trailing windows to be a robust default and event-count windows offering small but consistent gains. Under no-lookahead, gap and bucketized shapes underperform, while calendar-aligned windows yield mixed results; target encoding provides a large uplift, with time aggregation supplying additional, albeit modest, improvements. These findings inform practical feature engineering choices for temporally drifting, high-cardinality CTR tasks and emphasize careful evaluation to avoid leakage and overestimation of gains.

Abstract

This paper studies time aggregation features for XGBoost models in click-through rate prediction. The setting is the Avazu click-through rate prediction dataset with strict out-of-time splits and a no-lookahead feature constraint. Features for hour H use only impressions from hours strictly before H. This paper compares a strong time-aware target encoding baseline to models augmented with entity history time aggregation under several window designs. Across two rolling-tail folds on a deterministic ten percent sample, a trailing window specification improves ROC AUC by about 0.0066 to 0.0082 and PR AUC by about 0.0084 to 0.0094 relative to target encoding alone. Within the time aggregation design grid, event count windows provide the only consistent improvement over trailing windows, and the gain is small. Gap windows and bucketized windows underperform simple trailing windows in this dataset and protocol. These results support a practical default of trailing windows, with an optional event count window when marginal ROC AUC gains matter.

Figures (10)

• Figure 1: Evaluation protocol and no-lookahead feature constraint. For an impression at hour $H$, history-based features use only events from hours $< H$, and they exclude same-hour information.
• Figure 2: No target encoding sensitivity view. Panel A reports test ROC AUC by window length tuple, averaged over shapes and folds. Panel B reports test ROC AUC by window shape, averaged over length tuples and folds.
• Figure 3: Traffic light summary of shape effects relative to trailing windows, split by whether target encoding is used.
• Figure 4: League table of best and worst specifications by paired deltas in test ROC AUC relative to trailing $(1, 6, 24, 48, 168)$.
• Figure 5: Uplift from adding target encoding on top of time aggregation features.