Rank Flow Embedding for Unsupervised and Semi-Supervised Manifold Learning.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Advances in multimedia technology have led to vast collections, but labeled data for supervised learning remains scarce and costly.
• The need for effective retrieval and classification methods is growing, yet supervised approaches are limited by data availability.

Purpose of the Study:

• To introduce a novel manifold learning algorithm, Rank Flow Embedding (RFE), designed for unsupervised and semi-supervised learning scenarios.
• To develop a method that can operate effectively with minimal or no labeled data, addressing the limitations of supervised approaches.

Main Methods:

• The proposed Rank Flow Embedding (RFE) algorithm utilizes hypergraphs, Cartesian products, and connected components from manifold learning.
• It computes context-sensitive embeddings refined through a rank-based processing flow, incorporating complementary contextual information.
• Embeddings are generated for unsupervised retrieval and semi-supervised classification using Graph Convolutional Networks.

Main Results:

• Experiments were conducted on 10 diverse collections, utilizing features from Convolutional Neural Networks (CNN) and Vision Transformer (ViT) models.
• The Rank Flow Embedding (RFE) method demonstrated high effectiveness in unsupervised image retrieval, semi-supervised classification, and person re-identification (Re-ID).
• RFE proved competitive or superior to state-of-the-art methods across various evaluated scenarios.

Conclusions:

• The Rank Flow Embedding (RFE) algorithm presents a powerful solution for machine learning tasks with limited labeled data.
• Its effectiveness in unsupervised and semi-supervised settings, including image retrieval and classification, highlights its versatility and performance.
• RFE offers a significant advancement in handling multimedia data challenges where labeled datasets are constrained.