LPGL: A Model Applicable to Data Augmentation Algorithms for Depression Detection

• Published in: 2023 3rd International Conference on Consumer Electronics and Computer Engineering (ICCECE) pp. 129 - 139
• Main Authors: Wang, Shaoqiang, Wang, Shudong, Wang, Yifan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.01.2023
• Subjects: Classification algorithms; computer-aided diagnosis; Depression; detection; graph structure learning; label propagation; Neural networks; semi-supervised learning; Semisupervised learning; Sociology; Training; Training data
• DOI: 10.1109/ICCECE58074.2023.10135201

The clinical detection of depression generally judges depression and the degree of depression by scoring according to the depression scale. For doctors, it is time-consuming to perform the same scale screening for each patient, and the program based on computer-aided diagnosis has practical significance. At present, large-scale neural networks often fall into the dilemma of insufficient training data labeling in practical applications. We propose an algorithm based on label propagation that combines semi-supervised learning with graph structure learning (LPGL). The LPGL algorithm can simultaneously utilize three types of data: labeled data, unlabeled data, and data enhancement, so that the trained classifier is better than the training data and outperforms than other few-shot learning methods when labels are insufficient. This article uses the AVEC 2013 dataset to verify the superiority of the LPGL algorithm in terms of accuracy. Early diagnosis of depression is also conducive to early medical intervention, which will be of great help in alleviating depression.