<p>Cross-domain aspect-based sentiment analysis (ABSA) aims to transfer knowledge from labeled source domain data to perform fine-grained sentiment analysis tasks in an unlabeled target domain. To address the issues of model performance degradation due to domain discrepancy and the insufficient quality of generated augmented data, this paper proposes a unified feature refinement approach that dynamically integrates contrastive learning and domain alignment strategies for cross-domain ABSA. Firstly, we design a strategy that dynamically combines contrastive learning modules with a domain adversarial neural network (DANN). Contrastive learning extends the distribution of features from different domains in the reproducing kernel Hilbert space (RKHS), enhancing inter-domain separability and intra-domain compactness. Subsequently, adversarial training is employed to narrow the distribution gap between source and target domain features. Additionally, our framework incorporates Kullback–Leibler (KL) divergence as a regularization measure to optimize the model’s adjustment to different domain probability distributions. After training on source domain data, the model assigns pseudo-labels to unlabeled data in the target domain. Based on these pseudo-labels, we apply graph optimal transport (GOT) theory, framing cross-domain alignment as a graph matching problem to further improve domain alignment efficiency and generate enhanced text data for the target domain. Extensive experiments on several public ABSA datasets demonstrate that our model significantly outperforms competing approaches across four cross-domain ABSA sub-tasks, thereby confirming its effectiveness.</p>