<![CDATA[Background: Transfer learning has become increasingly prominent in data science due to the challenges posed by limited labeled data and distribution shifts between training and deployment environments. However, the success of transfer learning depends significantly on the structural compatibility between source and target domains.Aims: This study aims to investigate the relationship between domain similarity and transfer learning performance using an experimental framework termed Similarity-Aware Transfer Evaluation (SATE).Methods: Twelve pairs of benchmark datasets were selected to simulate varying levels of domain similarity and were made publicly available. Domain similarity was computed using Maximum Mean Discrepancy (MMD) in the learned representation space. Transfer performance was measured using a predefined Transfer Gain metric under bounded fine-tuning strategies. Correlation analysis and statistical testing were conducted to examine the relationship between similarity scores and transfer effectiveness, while fine-tuning depth was analyzed in relation to similarity magnitude.Result: The results demonstrate a strong positive correlation between domain similarity and transfer gain (r = 0.83, p < 0.01), indicating that approximately 69% of performance variability can be explained by similarity-based transfer effects. Negative transfer was observed when similarity scores were S ≤ 0.41. Furthermore, higher similarity levels were associated with deeper and more stable fine-tuning, whereas lower similarity resulted in increased instability during adaptation. These findings establish similarity as a structural compatibility constraint in transfer learning.Conclusion: The study confirms that domain similarity plays a fundamental role in determining transfer learning success. By operationalizing similarity measurement and linking it to performance thresholds, the proposed SATE framework provides a structured method for evaluating transfer feasibility in real-world data science applications.]]>
