The purpose of the click-through rate is to predict the probability that a user is most likely to click on a recommended item, garnering extensive attention in both academia and industry. In recent studies, it has been shown that high-quality user representation and feature interaction contribute significantly to improving accuracy in prediction tasks. However, the current methods still face two challenging problems. First, the behavior sequences contain complex interest features, and it is difficult to effectively capture the latent dominant interests. Besides, most models neglect the latent synergy between fine-grained features (i.e., they pay less attention to key feature interaction). To address these problems, in this paper, a target item-enhanced parallel fusion network (PFN) is proposed. First, the user’s historical interests are enhanced through a transformer. Then, the Pearson function is employed to gauge the strength of the relationship between the enhanced interest features and the target item, thus accentuating the user’s latent dominant interests. Third, feature interaction is learned via an equal interaction network, and then a soft-attention network is used to filter out unnecessary noise and retain fine-grained feature interaction to enhance the expressive ability of latent feature synergies. In addition, a multi-layer perceptron network is used to model the above features and learn the high-order representation of users’ more relevant interests. We have conducted extensive experiments on four public datasets and the PFN shows excellent performance.