Abstract
Modern app store recommender systems struggle with multiple-category apps, astraditional taxonomies fail to capture overlapping semantics, leading tosuboptimal personalization. We propose PCR-CA (Parallel CodebookRepresentations with Contrastive Alignment), an end-to-end framework forimproved CTR prediction. PCR-CA first extracts compact multimodal embeddingsfrom app text, then introduces a Parallel Codebook VQ-AE module that learnsdiscrete semantic representations across multiple codebooks in parallel --unlike hierarchical residual quantization (RQ-VAE). This design enablesindependent encoding of diverse aspects (e.g., gameplay, art style), bettermodeling multiple-category semantics. To bridge semantic and collaborativesignals, we employ a contrastive alignment loss at both the user and itemlevels, enhancing representation learning for long-tail items. Additionally, adual-attention fusion mechanism combines ID-based and semantic features tocapture user interests, especially for long-tail apps. Experiments on alarge-scale dataset show PCR-CA achieves a +0.76% AUC improvement over strongbaselines, with +2.15% AUC gains for long-tail apps. Online A/B testing furthervalidates our approach, showing a +10.52% lift in CTR and a +16.30% improvementin CVR, demonstrating PCR-CA's effectiveness in real-world deployment. The newframework has now been fully deployed on the Microsoft Store.