CULC-Net: A Recipe for Tailored Creative Selection in Online Advertising

Baosheng Zhang, Liufang Sang, Haoran Wang, Wei Wang, Wenlong Chen, Changping Peng, Zhangang Lin, Jingping Shao, Jie He, Haoqian Wang, Yuchen Guo

Abstract: Online advertising is a major application of recommendation systems. The primary process first involves recommending appropriate items to users, followed by selecting suitable creatives, such as ad posters. While much research has focused on optimizing item recommendations to increase user clicks, creative selection has often been overlooked. Properly chosen creatives can significantly enhance purchasing intent by aligning with the diverse preferences, ages, and genders of users. Current state-of-the-art methods typically rely on historical Click-Through Rates (CTR), which may exhibit biases during initial exposures due to limited data. In this paper, we introduce CULC-Net, which builds detailed profiles to uncover hidden connections between users and creatives, utilizing a creative relevance score for soft-decision making. This approach improves recommendation effectiveness and reduces reliance on sparse CTR data. Furthermore, we advance beyond the traditional CTR-based “only top for training” strategy by introducing FlexiRank. Creatives are sorted based on the relative strength of their CTRs, effectively managing noise and outliers. We test CULC-Net in a real-world search ad system, demonstrating a 3.43% increase in online and a 4.01% increase in offline. Further validation on a public benchmark confirms the effectiveness of our approach.

External IDs:dblp:conf/pkdd/ZhangSWWCPLSHWG25