Progressive learning unifies content relevance and personalization for better ad clicks
Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Information RetrievalMultimedia
Summary
Predicting which ads people will click on is tricky because it depends on the ad's details, how well it matches the search query, and the person’s individual preferences. The paper presents FineSID, a step-by-step method that breaks down ad content into fine parts, trains the system to match ads with queries better, and then tailors predictions based on user behavior. This approach improves click predictions and works better than older methods that rely mainly on broad identifiers. The authors tested their method extensively and showed it also helps distinguish creative ad choices more effectively.
What this means in practice
- •For online advertising engineers: Improve click prediction accuracy by integrating detailed content tokens, semantic query-item matching, and personalized user preferences in ad ranking systems.
- •For e-commerce recommendation teams: Enhance product ranking by jointly optimizing item content details, search relevance, and customer click behavior in recommender models.
Authors
Zhiguang Yang, Yulin Yang, Fuyu Lv, Dan Ou
Abstract
Multimodal click-through rate (CTR) prediction is central to search systems, where ranking depends on item content, query relevance, and user preference. Existing approaches incorporate content through dense embeddings or coarse discrete IDs, yet effectively integrating fine-grained heterogeneous content with ID-based predictors remains challenging. We propose FineSID, a progressive learning framework that unifies content tokenization, query--item semantic pretraining, and personalized post-training via a shared fine-grained discrete representation. Each stage optimizes an explicit objective while preserving token-level access to content. First, a pretrained multi-scale visual tokenizer and text tokenizers convert heterogeneous content into fine-grained Semantic IDs (SIDs), explicitly aligning them with the ID features. Second, full-parameter semantic pretraining under query--item relevance supervision establishes a robust matching prior, jointly optimizing the interactions of multimodal SIDs and conventional fields within a shared predictor. Third, this initialized predictor is directly fine-tuned with user click supervision to capture implicit personalized preferences. Extensive offline and online evaluations demonstrate substantial gains in both standard CTR prediction and core business metrics. Furthermore, FineSID improves within-product creative discrimination, demonstrating its semantic generalization beyond conventional ID memorization.