Charting EDA: Characterizing Interactive Visualization Use in Computational Notebooks with a Mixed-Methods Formalism.

Area of Science:

• Data Science
• Human-Computer Interaction
• Information Visualization

Background:

• Interactive visualizations are key tools for Exploratory Data Analysis (EDA).
• Understanding their impact on analyst observations is crucial for tool design.
• Previous research often focuses on task completion rather than the observation process.

Purpose of the Study:

• To investigate how interactive visualizations influence the observations data scientists make during EDA.
• To develop a formalism for describing the states and transitions within EDA processes.
• To identify design implications for interactive data analysis tools.

Main Methods:

• A qualitative experiment involving 13 professional data scientists using Jupyter notebooks.
• Collection of interaction traces and think-aloud utterances during data analysis.
• Qualitative coding of utterances to develop a formalism for EDA states (representations and observations).

Main Results:

• Interactive visualizations led to earlier and more complex insights into attribute relationships compared to static ones.
• Analysis revealed that some representations function as "planning aids" rather than solely for hypothesis testing.
• Metrics like revisit count and representational diversity highlighted patterns such as the "80-20 rule" in EDA behavior.

Conclusions:

• Interactive visualizations enhance the depth and timeliness of insights in EDA.
• The proposed formalism provides a framework for analyzing EDA processes and tool effectiveness.
• Findings inform the design of more effective interactive tools for data exploration and analysis.