Predicting User Preferences of Dimensionality Reduction Embedding Quality.

Area of Science:

• Data Science
• Computer Vision
• Machine Learning

Background:

• Dimensionality reduction techniques offer numerous choices for data reduction.
• Current evaluation of 2D embeddings is often qualitative, relying on human judgment.
• Users frequently treat dimensionality reduction as a black box, ignoring method-specific properties.

Purpose of the Study:

• To develop a quantitative method for evaluating 2D embeddings that incorporates human perception.
• To build a machine learning model that mimics human judgment in selecting high-quality embeddings.
• To identify and quantify factors influencing human preference for specific embeddings.

Main Methods:

• A comparative study was conducted where participants identified "good" and "misleading" low-dimensional image embeddings.
• A supervised machine learning model was trained using participant data as labels.
• The trained model was used as a proxy for human judgment to rank and explain embedding relevance.

Main Results:

• A quantitative framework for embedding evaluation was established, grounded in human perception.
• The developed model can effectively rank embeddings and provide explanations for their relevance.
• The degree of subjectivity in human selection of preferred embeddings can be quantified.

Conclusions:

• This work provides a novel, quantitative approach to assessing dimensionality reduction embeddings.
• The method bridges the gap between human perception and automated embedding evaluation.
• It enables more objective and interpretable selection of dimensionality reduction techniques.