Weakly-supervised locally linear embedding model for discriminant feature learning

doi:10.1016/j.knosys.2025.113966

Weakly-supervised locally linear embedding model for discriminant feature learning

2025-06-25

0

PRE

AI

DOI：10.1016/j.knosys.2025.113966

Original

Original request for help

Save

Abstract

En 中文

Discriminant feature learning is an important and popular research topic in machine learning, because it allows the exploration of explore high-dimensional data for discriminant feature information. Locally linear embedding (LLE) is widely used topological structure projection, but it ignores the discriminative information between clusters, which is critical in subsequent machine learning tasks such as data clustering and classification. To strengthen the LLE’s discriminative power, this study proposes a weakly-supervised discriminant feature learning (WSLLE) model. First, the WSLLE objective function is carefully designed to theoretically enhance the discriminative properties of data features, while preserving the data’s manifold structure. Second, the objective function is rigorously derived using the gradient descent method, which guarantees the existence of a solution and its convergence. Then, based on the inference of the objective function, an algorithm is proposed that utilises clustering methods to calculate the dissimilarity between samples. Finally, comparative experiments were conducted using clustering and classification methods to measure the discriminant properties of the WSLLE model. The experimental results demonstrate that the WSLLE model outperforms other feature learning approaches in terms of discriminative performance.

Weakly-supervised locally linear embedding model for discriminant feature learning

Abstract

Journal

Organization