Visual Clustering Factors in Scatterplots.

Area of Science:

• Data Science
• Computer Vision
• Human-Computer Interaction

Background:

• Cluster analysis is vital in data analysis, but evaluating clustering visually on scatterplots lacks a robust theoretical framework.
• Human visual perception is the benchmark for cluster quality, yet its application to large datasets is challenging due to the need for extensive human subject studies.

Purpose of the Study:

• To develop an empirical, data-driven approach for modeling human perception of visual clustering in large scatterplot datasets.
• To create a computational model that can objectively evaluate clustering quality based on human visual judgment.

Main Methods:

• Systematic construction and labeling of a large, publicly available scatterplot dataset.
• Qualitative analysis to identify visual factors influencing clustering perception.
• Training a deep neural network using the labeled dataset to model human visual clustering perception.

Main Results:

• The developed deep neural network model successfully replicates human visual perception of clusters.
• The data-driven model demonstrates superior performance compared to conventional clustering algorithms on both synthetic and real-world datasets.
• Identification of key visual factors influencing human perception of clusters in scatterplots.

Conclusions:

• A novel, data-driven deep learning model effectively captures human visual perception for scatterplot cluster analysis.
• This approach overcomes the limitations of traditional methods and subjective human evaluations for large-scale data.
• The findings offer a new paradigm for evaluating and improving automated cluster analysis techniques.