Current Research Projects

Enhancing the Reliability of Analyses in Graphical Data Interfaces

Graphical data analysis systems, including Tableau, R, and Jupyter Notebooks have become vital instruments in data science. These tools empower analysts to explore, visualize, and detect patterns in data. While valuable, these tools can also exacerbate the presence of noise that is inherent in many datasets. Spurious patterns are frequently misinterpreted by analysts, leading to false discoveries and conclusions that do not replicate or generalize. This phenomenon has contributed to a crisis of confidence in many science domains. One contributing factor is the ease with which analysts can wield interactive analysis systems, providing great flexibility in result interpretation but lacking sufficient guardrails to ensure the soundness of inferences and conclusions.

My lab has been developing new techniques to ensure the reliability of analyses in interactive systems. These techniques manifest as unobtrusive interventions and interactions that can be integrated into existing systems. They are designed to naturally promote and support more robust analytical practices without burdening the analyst with excessive constraints.

For instance, in our CHI '19 paper, we introduced an approach in which analysts can incorporate their hypotheses into the visual analysis process. Analysts articulate their intuitions, hypotheses, and expectations in natural language or via direct visual interaction before reviewing any data. Subsequently, the system responds by generating data plots that explicitly emphasize disparities between their assumptions and the actual data. We have developed fully functional visualization systems based on this concept (see our CHI '22 and IV papers). Evaluation involving data scientists have demonstrated that these minimally invasive interventions effectively promote a more methodical and thoughtful analytical approach. Furthermore, our experiments have revealed tangible benefits in terms of analysis reliability, showcasing reductions of up to 20% in false inferences (see our CHI '23 Honorable Mention paper). These findings have significant implications for enhancing the reliability of data analyses and the ensuing conclusions, potentially impacting a wide range of applications in data-intensive science.

This project has been supported by NSF CAREER and CRII awards.

Decoding Visualizations: Uncovering Perceptual and Cognitive Influences in Visual Data Interpretation

Information graphics, dashboards, and various visualizations have evolved into indispensable tools for conveying data to expert and lay audiences. The choice of the 'right' visual representation can often make the crucial difference between accurately comprehending the data or potentially misacting upon it. One of the primary objectives of my research lab is to elucidate the perceptual and cognitive factors that play a pivotal role in the creation and interpretation of effective data displays. My research employs a diverse array of empirical and design-based approaches, including controlled experiments, qualitative methods, crowdsourcing, and participatory design. Based on the insights gleaned, I seek to develop models and new design tools that make it easy for designers to create visualizations while adhering to the best practices.

For example, in a series of experiments, we explored the influence of color choice on people's ability to read maps, such as those used for public communication during weather emergencies. We found a strong correlation between the nameability of colors used in the display and viewers' ability to interpret these maps. Specifically, visualizations incorporating distinctively nameable colors appeared to be more easily interpretable (VIS '21 paper). In subsequent work, we also demonstrated how seemingly inconspicuous color selections can bias viewers, causing them to focus on specific data features within the visualization while potentially overlooking others (TVCG '23 paper).

These findings carry significant implications for data visualization and communication, especially in critical contexts such as hurricane forecasts and hazard maps. Moreover, these insights offer valuable perspectives that explain gaps between perceptual theory and visualization practice, demonstrating that the interpretation of data displays is frequently tied to linguistic attributes, rather than solely relying on perceptual characteristics (see our EuroVis '21 Best Paper).

This work has been supported by awards from the DOE Office of Science and Argonne National Laboratory.

Improving Drug Overdose Prevention with Real-Time Data Dashboards

The US continues to confront an escalating drug overdose crisis, which has intensified in the wake of the pandemic. Overdose Fatality Reviews (OFR) have emerged as a vital tool in the fight against accidental drug fatalities. OFR teams are typically tasked with reviewing overdose cases and recommending community-tailored strategies to reduce harm and prevent future overdoses. However, teams have been hampered by a lack of actionable data, limiting their reviews to only a handful of cases that may not represent population-level overdose patterns.

In collaboration with colleagues from the IU School of Medicine and RTI International, we are developing an enhanced data-driven OFR model. The key innovation is to shift teams away from an individual case-review model and towards using up-to-date population-level data, which is provided via real-time visualizations and data displays. Our central hypothesis is that this new model will scaffold effective data-driven decision-making among review teams, resulting in enhanced overdose response and prevention strategies, ultimately leading to a decrease in avoidable fatalities.

This interdisciplinary project is supported by a 5-year R61/R33 NIH award from the Heal: Data2Action program. The implementation phase of this project will involve the deployment of data dashboards to OFR teams across 19 Indiana counties. This intervention will in turn be evaluated in a clinical trial over the next three years.