Improved DeepCross AttentionNet for Recommendation Systems via Semantic Feature Transformation

The Deep & Cross Network (DCN) and its derivative models have become important paradigms for click-through rate (CTR) prediction due to their efficient and superior performance. CTR prediction is one of the core tasks in recommendation systems, as it directly impacts users' interest and engagement with recommended content. Existing DCN models rely on explicit feature interaction methods, which heavily depend on feature quality. However, in prediction tasks with low-quality features, especially those with large amounts of sparse input, these interactions often introduce significant noise and may not be efficient in learning all types of feature crosses.This paper proposes an improved Deep Cross Network, named DeepCross AttentionNet (DCAnet), which retains the advantages of the DCN model while introducing a novel attention-based cross network designed to more effectively learn specific feature interactions. In particular, DeepCross AttentionNet (DCAnet) enhances the representation of sparse features through semantic feature [3] transformation. By converting sparse features into more informative semantic features, this approach not only enables the model to more accurately capture complex feature relationships but also significantly improves the accuracy of CTR prediction. We evaluated the CTR model in practical applications for recommendation systems, and the experimental results demonstrate that our approach outperforms many competitive models in our real-world recommendation system.