Distillation-Accelerated Uncertainty Modeling for Multi-Objective RTA Interception

TL;DR

This paper addresses efficient and reliable traffic interception in Real-Time Auctions by jointly modeling multiple downstream objectives and prediction uncertainty. The authors introduce DAUM, a two-stage framework that combines multi-objective learning with uncertainty estimation, and they further apply knowledge distillation to produce a lightweight model that preserves uncertainty signals. Their experiments on the JD advertisement dataset show DAUM improves predictive performance across metrics while enabling a roughly tenfold increase in inference speed with the distilled model. The approach also demonstrates effective uncertainty transfer across correlated metrics, particularly useful under imbalanced label conditions, making uncertainty-aware interception practical for real-time deployment.

Abstract

Real-Time Auction (RTA) Interception aims to filter out invalid or irrelevant traffic to enhance the integrity and reliability of downstream data. However, two key challenges remain: (i) the need for accurate estimation of traffic quality together with sufficiently high confidence in the model's predictions, typically addressed through uncertainty modeling, and (ii) the efficiency bottlenecks that such uncertainty modeling introduces in real-time applications due to repeated inference. To address these challenges, we propose DAUM, a joint modeling framework that integrates multi-objective learning with uncertainty modeling, yielding both traffic quality predictions and reliable confidence estimates. Building on DAUM, we further apply knowledge distillation to reduce the computational overhead of uncertainty modeling, while largely preserving predictive accuracy and retaining the benefits of uncertainty estimation. Experiments on the JD advertisement dataset demonstrate that DAUM consistently improves predictive performance, with the distilled model delivering a tenfold increase in inference speed.

Figures (8)

• Figure 1: The main network of DAUM model. We first train the model using the PLE structure and collect the model weights from the last k epochs. For each neuron, we construct a Gaussian distribution based on these k sets of weights. We then perform multiple samplings from the weight distributions to compute the mean and variance of the outputs for each target. Finally, according to interception requirements, we apply a threshold on the predicted uncertainty: samples with high uncertainty are directly passed, while low-uncertainty samples are processed through the standard classification procedure.
• Figure 2: The number of deal samples in directly passed samples with respect to passing samples.
• Figure 3: The number of deal samples in remaining samples with respect to passing samples.
• Figure 4: The number of deal samples with respect to passing samples.
• Figure 5: The number of deal samples with respect to passed deal samples.