Linear and Deep Order-Preserving Wasserstein Discriminant Analysis.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Supervised dimensionality reduction for sequence data is challenging due to complex temporal structures.
• Existing methods often struggle to effectively measure sequence separability.

Purpose of the Study:

• To propose novel linear and non-linear methods for discriminative subspace learning in sequence data.
• To enhance the separability of sequences from different classes by considering temporal dynamics.

Main Methods:

• Developed order-preserving Wasserstein discriminant analysis (OWDA) for linear subspace learning.
• Introduced DeepOWDA, a deep extension for non-linear subspace learning.
• Utilized order-preserving Wasserstein (OPW) distance to measure inter-class and intra-class distances based on temporal structures.

Main Results:

• OWDA and DeepOWDA effectively learn discriminative linear and non-linear subspaces.
• The methods demonstrated superior performance on four 3D action recognition datasets.
• Novel separability measures based on OPW distance capture essential temporal differences.

Conclusions:

• OWDA and DeepOWDA offer effective solutions for dimensionality reduction in sequence data.
• These methods improve classification by focusing on distinctive temporal features.
• The approach shows significant potential for applications like action recognition.