Answers are grounded in this paper and cite their sources.
Abstract Predicting click-through rates involves estimating how likely it is that users will click on a website’s recommendations, such as advertisements or product listings. With the growth in data complexity and volume, it has become economically vital for online platforms to recommend appropriate items to users. Yet, this task remains difficult due to the sparse and high-dimensional categorical nature of the data inputs used in prediction models. Thanks to the superior performance of deep learning models compared to traditional methods, various deep learning strategies have been introduced to understand both simple and complex interactions from the mentioned data inputs. Nevertheless, feed-forward networks struggle to effectively capture common feature interactions and are limited in their capacity to efficiently model functions involving complex, high-order interactions. Moreover, as the complexity of the model layers increases, it becomes more challenging to reflect simple, low-order interactions in the output. In this paper, we introduce Low-Order Cross Attention Networks (LOCAN) designed to explicitly capture high-order interactions while also reflecting low-order interactions in the final prediction. We incorporate low-order interactions, identified through Factorization Machines, into the attention layer in a continuous manner, ensuring these interactions are carried through to the final prediction. Our experiments on three real-world datasets show that our model surpasses current leading methods, confirming the effectiveness of the proposed model.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: