Abstract
Predicting user conversions after an ad interaction is a critical task in online advertising. However, advertisers often utilize varying optimization windows, ranging from 1-day to 30-day periods. Training dedicated models for every possible window is computationally prohibitive. This paper introduces Personalized Interpolation, a resource-efficient framework that leverages two fixed-window models to accurately estimate conversion rates for any intermediate window. By treating conversion probability as a Cumulative Distribution Function (CDF), the method achieves performance parity with dedicated models while reducing training and infrastructure costs by approximately 83%.
Introduction: The Challenge of Flexible Optimization Windows
In digital advertising, a "conversion"—such as a purchase, app installation, or lead generation—rarely occurs immediately after a click. There is typically a variable time delay between the initial interaction and the final action. Advertisers optimize their campaigns based on a specific optimization window, which defines the maximum allowable time for a conversion to be attributed to a specific click.
As shown in Figure 1, these windows vary significantly by industry and objective. A mobile game developer might prioritize installs within 24 hours, whereas a high-consideration retailer (e.g., furniture or electronics) might track conversions over 14 or 30 days. Supporting these "Flexible Optimization Windows" (FOW) traditionally requires training a separate model for each timeframe, leading to a "combinatorial explosion" in system complexity and resource consumption.
Methodology: Personalized Interpolation
The proposed solution treats the conversion probability as a Cumulative Distribution Function (CDF) of the delay time $\tau$. Because a CDF is naturally monotonic and generally concave, intermediate values can be estimated through interpolation between two anchor points.
The system trains a single model that outputs two primary values:
- Short Optimization Window (SOW): e.g., $\mathbb{P}(\tau \leq T_s|e)$ for a 1-day window.
- Long Optimization Window (LOW): e.g., $\mathbb{P}(\tau \leq T_l|e)$ for a 7-day window.
To estimate the probability for any intermediate window $T_f$ (where $T_s < T_f < T_l$), the framework applies an interpolation factor ($\alpha$). As illustrated in Figure 2, the model estimates the value at $T_f$ by interpolating between the known probabilities at $T_s$ and $T_l$. The authors explored three mathematical variants for this interpolation:
- Linear: Assumes a constant rate of conversion growth between the two anchors.
- Rational: Uses a hyperparameter to model curvature, providing a smoother transition.
- Exponential: Models conversion likelihood as a decaying process, which often aligns most closely with observed user behavior.
This interpolation layer is applied during the inference stage, requiring no structural changes to the underlying production neural networks.
Experimental Setup and Results
The method was evaluated using production data from Meta’s ads delivery system, covering a 40-day period from late 2023 to early 2024. The researchers compared Personalized Interpolation against "Dedicated Models" (the gold standard of training a specific model for each window) and standard multi-task baselines.
Key Findings
- Efficiency: The interpolation approach achieved Normalized Entropy (NE) scores nearly identical to dedicated models but required only 1/6th of the training and infrastructure resources.
- Accuracy: Compared to a standard 7-TaskHead baseline model, the interpolation methods provided a consistent 0.2% to 0.3% improvement in NE.
- Calibration: All interpolation variants maintained high calibration (scores near 1.0), ensuring that predicted probabilities accurately reflected real-world conversion rates.
- Robustness: The exponential decay model proved to be the most resilient to hyperparameter variations, effectively capturing the natural "decay" of user intent over time.
Conclusion
Personalized Interpolation offers a scalable solution for managing flexible optimization windows in large-scale advertising systems. By anchoring predictions at two fixed points and interpolating intermediate values, the method eliminates the need for redundant model training without sacrificing predictive accuracy. Future work will explore making the interpolation factor dynamic, potentially tailoring the curvature to specific user or advertiser features.
Summary updated