Scholarlink

<p>Transfer learning (TL) algorithms enhance classification accuracy on a particular domain by transferring information from another related domain. These algorithms are particularly effective when the source and target domains are not equally distributed. They use labeled source domain samples to reinforce learning in the new target domain. A particularly innovative model of TL, Heterogeneous Domain Adaptation (HDA), is described by two domains having unlike feature spaces, unlike labels and data distributions. While several techniques have been used to approximate the feature space of the two domains and data distribution, the effect of the Normalization process on bringing the distributions of the domains and feature space together has yet to be thoroughly investigated. This study investigates the effect of thirteen normalization techniques including Min-Max, Max, Power Transformation, Mean Centered, Z-Score, Pareto Scaling, Variable Stability Scaling, Median and Median Absolute Deviation, Decimal Scaling, Logistic Sigmoid, Hyperbolic Tangent, Tanh, Variant of Tanh, to make feature space and distribution of source and target domain closer together to solve Domain Adaptation (DA) problems. In addition, a Data Topology Keeping (DTK) method during DA is presented. The proposed DTK method transforms data without requiring complex computations, such as those involving optimization techniques and deep neural networks. Various experiments on the Office-Caltech dataset with SURF features show the influence of feature normalization on improving semi-supervised learning. The experimental results revealed that the Z-Score and Pareto Scaling normalization techniques achieved the highest classification accuracy on SURF features in 3 out of 12 domain shifts (25%). Similarly, for DeCAF6 features, the Z-Score and Mean Centered normalization methods each recorded the highest accuracy across 3 out of 12 domain shifts (25%). In comparison, other normalization techniques achieved the best performance in at most 16.7% of the domain shifts. Consequently, the aforementioned normalization methods outperformed the others by at least 8.3%. Among all evaluated techniques, the Z-Score Normalization (ZSN) method consistently demonstrated superior performance across both sets of experiments, highlighting its strong potential for enhancing domain adaptation (DA) effectiveness in machine learning applications. The use of ZSN normalization in the proposed FN-DTK method achieved the highest classification accuracy in 6 out of 12 domain shift cases (50%), outperforming other state-of-the-art methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Influence of feature normalization methods on transfer Learning - A comparison study

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Mohammad Amin Pirbonyeh,
  • Mohammad Amin Shayegan

摘要

Transfer learning (TL) algorithms enhance classification accuracy on a particular domain by transferring information from another related domain. These algorithms are particularly effective when the source and target domains are not equally distributed. They use labeled source domain samples to reinforce learning in the new target domain. A particularly innovative model of TL, Heterogeneous Domain Adaptation (HDA), is described by two domains having unlike feature spaces, unlike labels and data distributions. While several techniques have been used to approximate the feature space of the two domains and data distribution, the effect of the Normalization process on bringing the distributions of the domains and feature space together has yet to be thoroughly investigated. This study investigates the effect of thirteen normalization techniques including Min-Max, Max, Power Transformation, Mean Centered, Z-Score, Pareto Scaling, Variable Stability Scaling, Median and Median Absolute Deviation, Decimal Scaling, Logistic Sigmoid, Hyperbolic Tangent, Tanh, Variant of Tanh, to make feature space and distribution of source and target domain closer together to solve Domain Adaptation (DA) problems. In addition, a Data Topology Keeping (DTK) method during DA is presented. The proposed DTK method transforms data without requiring complex computations, such as those involving optimization techniques and deep neural networks. Various experiments on the Office-Caltech dataset with SURF features show the influence of feature normalization on improving semi-supervised learning. The experimental results revealed that the Z-Score and Pareto Scaling normalization techniques achieved the highest classification accuracy on SURF features in 3 out of 12 domain shifts (25%). Similarly, for DeCAF6 features, the Z-Score and Mean Centered normalization methods each recorded the highest accuracy across 3 out of 12 domain shifts (25%). In comparison, other normalization techniques achieved the best performance in at most 16.7% of the domain shifts. Consequently, the aforementioned normalization methods outperformed the others by at least 8.3%. Among all evaluated techniques, the Z-Score Normalization (ZSN) method consistently demonstrated superior performance across both sets of experiments, highlighting its strong potential for enhancing domain adaptation (DA) effectiveness in machine learning applications. The use of ZSN normalization in the proposed FN-DTK method achieved the highest classification accuracy in 6 out of 12 domain shift cases (50%), outperforming other state-of-the-art methods.