Representative Negative Instance Generation for Online Ad Targeting (RNIG)

Based on support from the CCF‑Tencent Rhino Bird Fund, Tencent Ads and Prof. Li Yong’s team at Tsinghua University collaborated on large‑scale distributed recommendation algorithms, resulting in a paper accepted at ACM CIKM 2020 that addresses data imbalance in targeted advertising.

Targeted advertising aims to rank potential audiences for each ad based on relevance scores, but the overwhelming majority of impressions are negative instances, leading to severe class imbalance.

Existing solutions such as random undersampling, advanced negative sampling, or GAN‑based generation have limitations in the ad domain because exposure data already contains abundant observed negatives, and the challenge is to generate representative negative instances without selection bias.

The authors propose a new Generative Adversarial framework called Representative Negative Instance Generator (RNIG). RNIG jointly utilizes observed and unobserved exposure data to produce high‑quality representative negatives while avoiding selection bias, and is paired with a discriminator for ad‑user relevance ranking.

The generator consists of an embedding layer, separate MLP encoders for ads and users, a higher‑order interaction MLP, and outputs a representative‑negative score. A novel feature‑matching scheme aligns generated negatives with the feature distribution of observed negatives, using maximum mean discrepancy (MMD) and Jensen‑Shannon divergence (JSD) to select important features.

Experiments on a real Tencent Ads dataset show that RNIG outperforms previous methods on GAUC and RelaImpr metrics, demonstrating the effectiveness of the proposed approach.

Future work will explore the generality of RNIG across other targeted‑ad scenarios.