<p>Modeling feature interactions is one of the most crucial aspects of Click-Through Rate (CTR) prediction tasks. The application of Deep Neural Networks (DNN) for automatically learning implicit nonlinear interactions from raw, sparse features has gained widespread popularity in CTR prediction tasks. However, there are two issues with the cross-features modeling of DNN-based recommendation models: 1) Applying a fixed feature interaction approach for embedding features from different feature domains, which neglects the differences in spatial distribution across various feature domains; 2) Implicit feature interactions cannot explicitly model user-user, user-item, and item-item cross-features while modeling explicit feature interactions will introduce parameter redundancy. To address these issues, this paper proposes a novel DaDc model, which includes two crucial modules: Dual Attention Net (Da) and Dual Co-action Unit (Dc), that the parameter redundancy can be reduced by this clever combination. The Da module reweights features based on the semantic relevance of cross-domain features, alleviating the heterogeneity among cross-domain features. The Dc module implements explicit feature interactions between cross-domain and same-domain features in a parameter-efficient manner. Through experiments, the Area Under Curve (AUC) of DaDc was evaluated for performance on two real datasets, achieving state-of-the-art results. Additionally, the number of parameters introduced by DaDc is lower than that of most existing DNN-based CTR prediction models. Ablation experiments demonstrate the effectiveness of the proposed Da and Dc modules, and when the dataset has fewer available features, Da and Dc modules exhibit stronger combined effects.</p>