Scholarlink

<p>The generalization and transferability are important factors to evaluate the performance of a machine learning algorithm. For a supervised learning algorithm, it can only evaluate the generalization by drawing a validation dataset from training dataset. This evaluation method has to face a reality that if the distributions of training and test datasets are similar, the prediction results are usually good. But if the data distributions are not similar, the bias of the model can be increased, which leads to a large prediction error. Thus, how to quantitatively evaluate the similarity between different datasets is crucial, especially in the seismic interpretation based on the supervised deep learning algorithm, which needs more training data and is more inclined to result in over-fitting than other machine learning algorithms. From the perspective of data distribution measurement, a general method for evaluating the similarity between seismic datasets is proposed. Based on a deep learning discriminative network, K-fold cross-validation is used to calculate the average AUC (Area under Curve) between the training dataset and test dataset, which can be regarded as a kind of similarity evaluation index for different datasets. In the experiments, two groups of experiments are designed to discuss the rationality of the proposed framework. Additionally, the problem related with imbalance and generalization are also considered. To compare with the proposed method, other two methods based on MK-MMD (Multiple Kernel variant of MMD) and KL (Kullback-Leibler) divergence respectively are used to measure the distributions for the same seismic datasets. Finally, it is shown that the similarity index calculated by the proposed method is consistent with the similarity representation of MK-MMD, and the average AUC calculated is inversely proportional with the FW-IoU (Frequency Weighted Intersection over Union), which is the metric of prediction accuracy. The average AUC closing to the minimum value means the prediction result is reliable and favorable. Another usage of this proposed framework is that it can output the identification probability of a test sample, which will be helpful for the data enhancement in seismic interpretation..</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

Similarity evaluation of seismic datasets by an adversarial validation framework based on deep learning

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Shu-na Chen,
  • Zhe-ge Liu,
  • Shu-ying Ma,
  • Jing-Bo Zhai,
  • Ya-juan Xue

摘要

The generalization and transferability are important factors to evaluate the performance of a machine learning algorithm. For a supervised learning algorithm, it can only evaluate the generalization by drawing a validation dataset from training dataset. This evaluation method has to face a reality that if the distributions of training and test datasets are similar, the prediction results are usually good. But if the data distributions are not similar, the bias of the model can be increased, which leads to a large prediction error. Thus, how to quantitatively evaluate the similarity between different datasets is crucial, especially in the seismic interpretation based on the supervised deep learning algorithm, which needs more training data and is more inclined to result in over-fitting than other machine learning algorithms. From the perspective of data distribution measurement, a general method for evaluating the similarity between seismic datasets is proposed. Based on a deep learning discriminative network, K-fold cross-validation is used to calculate the average AUC (Area under Curve) between the training dataset and test dataset, which can be regarded as a kind of similarity evaluation index for different datasets. In the experiments, two groups of experiments are designed to discuss the rationality of the proposed framework. Additionally, the problem related with imbalance and generalization are also considered. To compare with the proposed method, other two methods based on MK-MMD (Multiple Kernel variant of MMD) and KL (Kullback-Leibler) divergence respectively are used to measure the distributions for the same seismic datasets. Finally, it is shown that the similarity index calculated by the proposed method is consistent with the similarity representation of MK-MMD, and the average AUC calculated is inversely proportional with the FW-IoU (Frequency Weighted Intersection over Union), which is the metric of prediction accuracy. The average AUC closing to the minimum value means the prediction result is reliable and favorable. Another usage of this proposed framework is that it can output the identification probability of a test sample, which will be helpful for the data enhancement in seismic interpretation..