Learning Compositional, Visual and Relational Representations for CTR Prediction in Sponsored Search

As the main revenue source for search engines, sponsored search system retrieves and allocates ads to display on the search result pages. Click-through rate (CTR) prediction is a crucial task for search ads, due to it plays a key role in ranking and pricing of candidate ads. Commercial search engines typically model it as a classification problem and use machine learning models for CTR prediction. Recently deep learning based models have been proposed to learn latent representations of query-ad relevance and historical information to improve the accuracy, which follow the Embedding&MLP paradigm. As the learning feasible embeddings requires sufficient samples and these models rely mainly and heavily on textual features and sparse ID features, representations learning for new ads is inadequate and models are confronted with the cold-start problem. Meanwhile, as search ad offerings become increasingly complex, rich ads with various sizes, decorations and formats are growing rapidly. Due to the diverse ad extensions and layouts, rich ads pose new challenges for CTR prediction. To tackle these problems, in this paper, we propose an approach to improve the accuracy of CTR prediction by learning supplementary representations from three new aspects: the compositional components, the visual appearance and the relational structure of ads. This method can utilize straightforward and auxiliary information for new and rich ads, which can improve the expressive ability and the generalization of model greatly. We demonstrate the performance of this method on datasets obtained from a real sponsored search system in the offline environment. Experimental results show that our approach can improve the accuracy of CTR prediction and achieve superior performance compared to the baseline method, especially for the rich and new ads.