GenLI model cuts CTR prediction complexity to O(1) with generative interests
New generative approach reduces retrieval time to constant while capturing diverse latent user interests.
Traditional click-through rate (CTR) prediction relies on a two-stage framework: a general search unit retrieves the top-k relevant past behaviors, then an exact search unit applies attention. This approach suffers from two key limitations: it ignores latent user interests not directly related to the target item, and the pairwise similarity matching between the target and each historical behavior grows time-consuming as user history accumulates.
To solve these problems, the authors introduce GenLI, which consists of three modules. The Interest Generation Module (IGM) generates multiple interest distributions independent of the target item, capturing diverse aspects of real-time user interests while incorporating interaction information among behaviors. The Behavior Retrieval Module (BRM) then selects related behaviors with a simple O(1) lookup, avoiding expensive pairwise scoring. Finally, the Interest Fusion Module (IFM) uses adaptive gating to combine generated interest features. This design enables GenLI to balance accuracy and efficiency, making it well-suited for large-scale online advertising and recommendation systems.
- GenLI reduces behavior retrieval time complexity from O(n) to O(1) using a simple lookup operation.
- The target-independent Interest Generation Module captures multiple latent interest distributions, improving diversity over traditional target-centered approaches.
- A gating-based Interest Fusion Module combines generated features for enhanced CTR prediction accuracy.
Why It Matters
Faster, more diverse interest modeling could significantly improve ad revenue and user engagement at scale.