融合樣本選擇的深度圖半監(jiān)督分類

打開文本圖片集

A semi-supervised classification model for fusion sample selection based on depth map

LI Shun-yong1'2，WEN Nan1， ZHAO Xing-wang3 （1.School of Mathematics and Statistics，Shanxi University，Taiyuan O3oo06，China； 2. Key Laboratory of Complex Systems and Data Science of Ministry of Education， Shanxi University，Taiyuan O3oo06，China; 3.School of Computer and Information Technology，Key Laboratory of Computational Intellgence and Chinese Information Processing of Ministry of Education，Shanxi University， Taiyuan O3ooo6，China）

Abstract： Traditional supervised learning requires a large number of labeled samples for model training，which makes it difficult to apply traditional supervised models to tasks lacking labeled samples.To address this issue，a semi-supervised classification model for fusion sample selection based on depth map（SSC_ FSSDM） is proposed. The model is divided into two parts： graph structure clustering and semi-supervised classification. In graph structure clustering，unlabeled samples are represented as high-quality graph structures using Laplace rank constraints，and the class information of labeled data is used as prior information to cluster the graph structures to obtain pseudo labels of unlabeled samples.A sample selection mechanism is used to select reliable samples from the pseudo labels，reducing the impact of noisy samples on model performance. In semi-supervised classification，reliable samples and their pseudo labels are used as inputs for deep learning to predict the labels of unlabeled samples in the original data. The performance of the SSC-FSSDM model was tested on three datasets， and various indicators showed that the SSC-FSSDM model outperformed other semi-supervised classification models.

Key words：sample selection; diagram structure; Laplace; clustering; semi-supervise

0 引言

在實際應(yīng)用中，如文本分類、語音識別、電子郵件分類和計算機(jī)輔助醫(yī)療診斷，存在大量的未標(biāo)記數(shù)據(jù)需要手工標(biāo)記或者通過實驗獲取，此過程費(fèi)時費(fèi)力.充分利用這些未標(biāo)記數(shù)據(jù)來完成最終的標(biāo)簽預(yù)測是非常重要的，因此，半監(jiān)督學(xué)習(xí)得到了越來越多的關(guān)注.

半監(jiān)督學(xué)習(xí)介于無監(jiān)督學(xué)習(xí)和監(jiān)督學(xué)習(xí)之間，利用同時包含標(biāo)簽和無標(biāo)簽的數(shù)據(jù)來構(gòu)建一個模型對未標(biāo)記樣例進(jìn)行標(biāo)記，使得模型能在預(yù)測階段更好地泛化到“新"數(shù)據(jù)。（剩余8773字）