Semisupervised learning from dissimilarity data.

Area of Science:

• Machine Learning
• Data Science
• Dimensionality Reduction

Background:

• Learning from dissimilarity data presents challenges in representation and classification.
• Traditional methods often rely solely on labeled data, limiting performance when unlabeled data is abundant.

Purpose of the Study:

• To present a novel two-stage approach for learning from dissimilarity data.
• To investigate the impact of including unlabeled objects in the embedding phase on classifier performance.
• To emphasize the synergy between classical multidimensional scaling and linear discriminant analysis.

Main Methods:

• A two-stage approach: (1) embedding labeled and unlabeled objects in Euclidean space, and (2) training a classifier on labeled objects.
• Utilizing classical multidimensional scaling for embedding and linear discriminant analysis for classification.
• Investigating the effect of embedding dimension selection and the inclusion of unlabeled data.

Main Results:

• The choice of embedding dimension is critical for classifier performance.
• Including unlabeled objects in the embedding stage demonstrably improves classifier performance, particularly when data exhibits spherical covariances.
• The proposed method shows effectiveness across several presented examples.

Conclusions:

• The described two-stage learning framework offers an effective way to leverage both labeled and unlabeled dissimilarity data.
• Unlabeled data inclusion is a valuable strategy for enhancing machine learning models in dissimilarity-based tasks.
• The method provides a robust approach for representation learning and subsequent classification.