A Business Intelligence Innovative Approach to Ad Recall: Cross-Attention Multi-Task Learning for Digital Advertising

This paper introduces AdRecallNet, a new frame-work for ad recall modeling aimed at improving click-through rate (CTR) prediction, ad ranking, and search term generation. The model combines a Cross-Attention Mechanism with a Multi-Task Learning Framework and uses the Unimo-text-large generative model to optimize ad ranking and search term generation. It extracts query and ad feature vectors using a query-ad encoding module and computes semantic relevance through a cross-attention layer. The ad recall ranking module ranks ads by cosine similarity, and the search term generation module predicts search terms based on ad content. The model also includes a Feedback Loop Mechanism that uses real-time user data to improve continuously and adapt to changing market conditions. Experimental results show that AdRecallNet outperforms baseline models in metrics like Precision @ K, Recall @ K, and NDCG. Ablation studies highlight the significance of the query-ad interaction and search term generation modules. This work advances ad recall optimization and provides a solid technical foundation for future research in online advertising systems.