Insight-centric Visualization Recommendation

Camille Harris, Ryan A. Rossi, Sana Malik, Jane Hoffswell, Fan Du, Tak Yeon Lee, Eunyee Koh, Handong ZhaoarXiv 2021 · OpenReview

TL;DR

This paper introduces "Insight-centric Visualization Recommendation," a new paradigm that goes beyond simply recommending charts. Instead, it automatically discovers, ranks, and visualizes groups of related "insights" (meaningful patterns in data) and the most important insights within those groups. This approach helps users quickly find valuable information in unfamiliar datasets by prioritizing what's important, not just what's possible.

What Problem Does This Paper Address?

Exploratory Data Analysis (EDA) is crucial for understanding data, but it can be challenging, especially for users unfamiliar with a dataset. Existing visualization recommendation systems often present a long list of potential visualizations, making it difficult and time-consuming for users to sift through and find the truly interesting or relevant insights. This "global ranking" approach doesn't effectively guide users to the most valuable discoveries. The authors aim to solve this by shifting the focus from recommending individual visualizations to recommending groups of related insights and then highlighting the most significant insights within those groups.

Methods

The core of the proposed system, named "SpotLight," is an "insight-centric" approach. This involves several key steps:

  • Auto-Insight Discovery: SpotLight automatically discovers insights from a dataset. It employs a diverse set of machine learning methods (information theoretic, statistical, supervised, and unsupervised) to identify various "insight-types" (e.g., time-series outliers, multivariate outliers, nonlinear correlations). A key innovation here is using multiple methods for each insight-type to ensure robustness across different data characteristics, addressing the "no-free lunch" theorem.
  • Insight Ranking: Once insights are discovered within an insight-type, they are scored and ranked. The system uses a novel formulation that combines scores from multiple learning-based methods, potentially using ensemble techniques or rank-based averaging to derive a final score for each insight. This allows for a more nuanced understanding of insight importance.
  • Insight-Type Ranking: The system then scores and ranks the "insight-types" themselves. This is done by aggregating the scores of the insights within each type, effectively prioritizing entire categories of insights that are most prevalent or important in the data.
  • Visualization Inference: For each recommended insight, SpotLight infers an appropriate visualization and provides an annotation to explain the insight.

Datasets

The system was evaluated using two datasets:

  • Weather Dataset: Consisted of weather patterns for two US cities (Seattle and New York) over four years (2014-2017), including measurements like precipitation, wind, temperature, and categorical weather types.
  • Wholesale Customer Dataset: Contained annual spending data for 440 businesses, including customer type (hospitality/retail) and spending across various departments.

Both datasets were sourced from the UCI ML Repository.

Key Insights and Ideas

  • Shifting from Visualizations to Insights: The fundamental idea is that users are looking for knowledge (insights), not just charts. Recommending insights directly, grouped by type, is more efficient than presenting a vast array of visualizations.
  • The Power of Insight-Type Grouping: Ranking and presenting entire categories of insights (e.g., "all the interesting correlations") first helps users quickly grasp the main themes in their data and then dive into specific examples.
  • Leveraging Multiple Discovery Methods: The "no-free lunch" theorem is a key motivator for using a diverse set of insight discovery algorithms. This ensures that the system can find relevant insights regardless of the specific data distribution.
  • Generalizability: The approach is designed to be generalizable across various attribute types (numerical, categorical, temporal) and complex combinations thereof, making it applicable to a wide range of datasets.

Key Results

A user study with 12 participants was conducted to evaluate SpotLight. Key findings include:

  • Insight Ranking Effectiveness: 68-61% of bookmarked insights (weather/wholesale) came from the top-5 recommended insights within their respective insight-type rows. When considering the top-10, this rose to 91-86%. This indicates users found the ranked insights highly relevant.
  • Insight-Type Ranking Effectiveness: The ranking of insight-types by SpotLight was significantly correlated with user preferences (Kendall's τ rank correlation coefficient of 0.87 for the wholesale dataset). This suggests the system effectively prioritizes the most relevant categories of insights.
  • User Perception: Participants generally found SpotLight useful for understanding data (mean score 6.3) and generating hypotheses (mean score 6.25). They appreciated the automated and comprehensive nature of the recommendations.
  • Insight-Type Usefulness: Users found most of the supported insight-types useful and diverse, with only one insight-type (heavy tails) not being bookmarked by any participant.

Limitations and Future Work

  • Limitations:
    • The user study used datasets with a relatively small number of columns (<10). The authors suggest that the importance of insight-type and insight ranking would be even more pronounced with datasets having significantly more columns.
  • Hypothesis-driven exploration could be enhanced with features like filtering by values/ranges, customizing chart axes, and enabling more interactive visualizations.
  • Future Work:
    • Investigate personalizing the insight-type and insight rankings based on user behavior logs (e.g., bookmarked insights, clicked insights).
  • Explore adding more descriptive statistics and explanations for insights to make them more accessible to users with less statistical background.
  • Enhance visualization interactivity and consistency.

Summary updated

Abstract

Visualization recommendation systems simplify exploratory data analysis (EDA) and make understanding data more accessible to users of all skill levels by automatically generating visualizations for users to explore. However, most existing visualization recommendation systems focus on ranking all visualizations into a single list or set of groups based on particular attributes or encodings. This global ranking makes it difficult and time-consuming for users to find the most interesting or relevant insights. To address these limitations, we introduce a novel class of visualization recommendation systems that automatically rank and recommend both groups of related insights as well as the most important insights within each group. Our proposed approach combines results from many different learning-based methods to discover insights automatically. A key advantage is that this approach generalizes to a wide variety of attribute types such as categorical, numerical, and temporal, as well as complex non-trivial combinations of these different attribute types. To evaluate the effectiveness of our approach, we implemented a new insight-centric visualization recommendation system, SpotLight, which generates and ranks annotated visualizations to explain each insight. We conducted a user study with 12 participants and two datasets which showed that users are able to quickly understand and find relevant insights in unfamiliar data.
BibTeX
@misc{harris2021insight,
  author = {Camille Harris and Ryan A. Rossi and Sana Malik and Jane Hoffswell and Fan Du and Tak Yeon Lee and Eunyee Koh and Handong Zhao},
  title = {{Insight-centric Visualization Recommendation}},
  year = {2021},
  eprint = {2103.11297},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.11297},
}
Figure