Semi-supervised dimension reduction using trace ratio criterion.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Semi-supervised dimension reduction aims to learn low-dimensional representations from partially labeled data.
• Existing methods like Semi-Supervised Discriminant Analysis (SDA) often impose rigid constraints on data representation.
• The trace ratio (TR) optimization problem is a key challenge in dimension reduction.

Purpose of the Study:

• To address the trace ratio problem in semi-supervised dimension reduction.
• To relax the linear subspace constraint in SDA by introducing a flexible regularizer.
• To develop an efficient algorithm for optimizing the proposed method.

Main Methods:

• Reformulated the objective function of SDA into a trace ratio form.
• Introduced a flexible regularizer ||F - X(T) W||(2) to relax linear constraints.
• Developed an iterative algorithm for simultaneous optimization of data representation (F) and projection matrix (W).

Main Results:

• The proposed TR-FSDA method demonstrated improved ability to handle data from nonlinear manifolds close to linear subspaces.
• The iterative algorithm was theoretically proven to converge to the optimum using the Newton-Raphson method.
• Experimental results showed TR-FSDA outperformed existing semi-supervised dimension reduction methods on face, shape, and webpage datasets.

Conclusions:

• TR-FSDA offers a more flexible and effective approach to semi-supervised dimension reduction compared to traditional methods.
• The developed iterative optimization algorithm is robust and guarantees convergence.
• The method shows significant potential for applications in pattern recognition and data analysis.