A Semi-Supervised Stacked Autoencoder Using the Pseudo Label for Classification Tasks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Manual sample labeling is inefficient, leading to numerous unlabeled training samples in practical scenarios.
• Semi-supervised learning aims to utilize both labeled and unlabeled data, which is crucial for overcoming labeling limitations.
• Traditional stacked autoencoders (SAE) are supervised and struggle with semi-supervised tasks due to their reliance solely on labeled data.

Purpose of the Study:

• To propose a novel pseudo label-based semi-supervised stacked autoencoder (PL-SSAE) for effective semi-supervised classification.
• To enhance the utilization of unlabeled data in stacked autoencoder models.
• To improve classification performance in scenarios with limited labeled data.

Main Methods:

• Introduced a pseudo-labeling method into the stacked autoencoder framework, creating the PL-SSAE.
• Employed unsupervised pre-training using autoencoders (AE) on all available samples to initialize network parameters.
• Implemented iterative fine-tuning using labeled samples to generate pseudo labels for unlabeled samples, which were then used for regularization.

Main Results:

• The PL-SSAE effectively utilizes both labeled and unlabeled samples by incorporating pseudo-labeled data during fine-tuning.
• Empirical evaluations demonstrated that PL-SSAE significantly outperforms traditional SAE, SSAE, Semi-SAE, and Semi-SSAE on benchmark datasets.
• The proposed method shows competitive semi-supervised classification performance.

Conclusions:

• The PL-SSAE successfully addresses the challenge of limited labeled data by effectively exploiting unlabeled samples.
• The integration of pseudo-labeling enhances the capability of stacked autoencoders for semi-supervised learning tasks.
• PL-SSAE offers a more competitive and effective approach for semi-supervised classification compared to existing methods.