Source-linked AI summary
Talking datasets: Understanding data sensemaking behaviours
Laura Koesten, Kathleen Gregory, Paul Groth, Elena Simperl
TL;DR
Researchers’ data reuse requires making sense of data, but strategies for data-centric sensemaking are less understood than those for textual information. This study examines researchers’ interactions with familiar and unfamiliar data and identifies three clusters of sensemaking activity. The findings organize these activities around inspecting data, engaging with content, and placing data in broader contexts.
Problem
Strategies for making sense of data are relatively underexplored, although consumers must understand data to reuse it.
Source: §1 · p. 1, §2.2 · p. 2Method
The study used verbal summarization to examine researchers’ sensemaking with familiar data they brought and unfamiliar data supplied by the researchers.
Source: §1 · p. 1, §3.1 · p. 4Results
The findings identify three activity clusters: inspecting data, engaging with its content, and placing it in broader contexts.
Source: §5 · pp. 18–19Takeaways & Limitations
The synthesis relates sensemaking activity clusters to information structures needed for data reuse.
Source: §5 · p. 18The observed patterns may depend on the datasets used, and summarizing familiar data first may have primed participants for the unfamiliar-data task.
Source: §6 · p. 23
Abstract
from arXiv · showhide
The sharing and reuse of data are seen as critical to solving the most complex problems of today. Despite this potential, relatively little is known about a key step in data reuse: people's behaviours involved in data-centric sensemaking. We aim to address this gap by presenting a mixed-methods study combining in-depth interviews, a think-aloud task and a screen recording analysis with 31 researchers as they summarised and interacted with both familiar and unfamiliar data. We use our findings to identify and detail common activity patterns and necessary data attributes across three clusters of sensemaking activities: inspecting data, engaging with content, and placing data within broader contexts. We conclude by proposing design recommendations for tools and documentation practices which can be used to facilitate sensemaking and subsequent data reuse.
1. Introduction
Data reuse requires consumers to make sense of data, yet the behaviours involved in data-centric sensemaking remain insufficiently understood. The study examines those behaviours to identify patterns and inform tools and documentation practices for sensemaking and reuse.
1. Introduction
- Data reuse depends on consumers making sense of others’ data, a difficult and time-intensive task even within researchers’ own disciplines.Sensemaking also requires social and technical structures to convey meaning.Source: §1 · p. 1
- The study uses verbal summarization to examine sensemaking with familiar and unfamiliar data, treating summarizing as a way to uncover cognitive processes.The research questions ask about activity patterns in initial sensemaking and how those patterns afford potential data reuse.Source: §1 · pp. 1–2
- The study combines interviews, a think-aloud task, and screen-recording analysis to identify activity patterns and data attributes across three sensemaking clusters.Its contributions include a framework, user needs for data reuse, and design recommendations for tools and documentation practices.Source: §1 · p. 2
2. Background
Prior work has studied sensemaking and data reuse, but strategies for making sense of data remain less understood than those for textual information. This section frames data-centric sensemaking as shaped by data characteristics, task, context, and the consumer’s perspective.
2. Background
- In information science and HCI, sensemaking constructs meaning from information through an iterative process of linking pieces into a conceptual representation.Related models describe evolving steps, influences such as skills and preconceptions, and information-processing components.Source: §2 · p. 2, §2.1 · p. 2
2.2. Data-centric sensemaking
- Research on strategies for making sense of data is relatively limited, while data consumers’ perspectives shape how data are perceived, interpreted, and acquired.HCI work includes visualization, tailored tools, and exploratory analysis, but task purpose affects the importance of quality and uncertainty indicators.Source: §2.2 · pp. 2–3
2.3. Evaluating data for reuse
- Data reuse decisions draw on contextual information about production, repositories, and use, as well as collection conditions, processing, relevance, quality, and usability.Knowledge needed to make sense of data may be tacit and absent from documentation.Source: §2.3 · p. 3
2.4. Summarisation as a way to understand sensemaking
- Verbal summarization can reveal spontaneous cognitive processes involved in comprehending data, which written dataset summaries may not capture.The study builds on selection, condensation, and transformation as cognitive activities in summarization.Source: §2.4 · pp. 3–4
2.5. Summary of key points
- The study treats sensemaking as involving data-specific activities and social, disciplinary, and communication contexts that shape consumer engagement.It focuses on understanding data rather than reviewing operational, machine-readability, and interoperability problems in detail.Source: §2.5 · p. 4
3. Methodology
The researchers used semi-structured interviews and verbal summarization to examine sensemaking with familiar and unfamiliar data. Participants brought their own data and also explored a prepared dataset.
3. Methodology
- A semi-structured interview design examined how researchers verbally summarized and made sense of familiar and unfamiliar data.The design drew on earlier work on textual data summarization and research-data reuse.Source: §3 · p. 4
3.1. Study design
- Participants chose familiar data to bring; most brought data they had created, and the researchers did not request supporting documentation or metadata.Most brought spreadsheets, while others brought text, images, videos, or other artefacts.Source: §3.1 · p. 4
- Researchers prepared an unfamiliar, domain-general dataset containing numerical and textual data, missing values, formatting inconsistencies, and ambiguous variables.It was a modified spreadsheet from the Guardian Data Blog, United Nations dataset.Source: §3.1 · pp. 4–5, Figure 1 · p. 5
3.2. Data collection
- Interviews asked participants to summarize data for a potential reuser, explore it, and discuss data creation, documentation, and reuse.Screen recordings captured interactions in 26 interviews; five were not created because of technical problems.Source: §3.2 · p. 5
Section EXAMPLES
The study recruited researchers with varied disciplinary backgrounds and analyzed interview transcripts and screen recordings to develop a coding framework and identify common interactions.
Section EXAMPLES
- The interview schedule included a summarization task and example topics for different interview sections; two pilot interviews in October 2018 informed interview duration and refined the questions and task setup.Source: Table 1 · p. 5, Section EXAMPLES · p. 5
- The 31 participants included 27 selected from survey respondents and four recruited through convenience and purposive sampling.Recruitment sought a range of disciplines and nationalities, and participants had published research papers.Source: Section EXAMPLES · pp. 5–6
3.3. Data analysis
- Thematic analysis combined deductive and inductive coding, with authors comparing and revising codebooks before producing a unified coding tree.The resulting tree contained three primary codes and 30 child codes, and senior researchers checked a sample.Source: §3.3 · pp. 6–7
- Screen-recording analysis identified the first occurrence of 15 common interactions with the unfamiliar dataset, without measuring their duration.The analysis covered recordings from 26 interviews and accounted for how much spreadsheet content was visible without scrolling.Source: §3.3 · p. 7
4. Findings
The findings describe data sensemaking through inspecting data, engaging with its content, and placing it in broader contexts, with actions and interpretations shaped by data attributes and relevant context. They also address how topic familiarity, information structures tailored to data, consumers, and research methods, and documentation relate to understanding and reuse.
4.1.1. Inspecting
- For unknown data, 23 participants gave linear summaries and 8 interwove attributes with analysis; interwoven summaries involved more searching for column relationships and expressing uncertainty.Headers, quality or uncertainty, topic or title, and analysis or dependencies were among the most frequently mentioned attributes.Source: §4.1.1 · pp. 8–9
- Most descriptions of participants’ own data followed a linear pattern (n = 24), while interwoven summaries (n = 7) mixed methodology with headers, variables, and format; descriptions also covered field, access, and supporting documents.Participants’ spreadsheet actions commonly began with exploring dataset dimensions and topic, then examining cells and moving between columns; 10 reformatted spreadsheets, while analysis features appeared in only four instances.Source: §4.1.1 · pp. 9–10
4.1.2. Engaging with the content
- Participants encoded data and sought patterns using plots, relationships, scales, and formulas, while analysis software shaped how they structured data and which encodings they used.They also grouped semantically related columns and looked for how null values were represented; few used spreadsheet analysis tools during the task.Source: §4.1.2 · pp. 11–12
- Participants examined encodings, missing values, anomalies, and unexpected values to assess data quality, and these “strange things” often prompted deeper engagement and reconsideration of assumptions.They also worried that producer-created categories or processing decisions could be misunderstood by reusers if not explained.Source: §4.1.2 · p. 12
- Collaborative sensemaking helped participants clarify data meanings, study decisions, and errors, with important details sometimes emerging only when people could inspect and discuss data together.Participants described consulting data generators about unclear columns and tracking sample responsibilities so errors could be traced to contributors.Source: §4.1.2 · p. 13
4.1.3. Placing
- Participants placed data in immediate creation contexts, disciplinary practices, and broader social or real-world contexts, recognizing that such context can shape meaning and judgments of representativeness.They questioned whether country data were complete and whether counts represented entire countries, samples, or estimates.Source: §4.1.3 · pp. 14–15
- Participants often needed to know the data’s original purpose to place it, and disciplinary and data expertise helped; most found their own data easier to describe than unfamiliar data.Unknown data were harder to contextualize when their original study objective was not available.Source: §4.1.3 · p. 15
- Researchers found familiar data easier to make sense of because they already knew its quality and content; unfamiliar data outside their field was much harder.This contrast came from a participant comparing their own data with a database they had never seen and that was outside their discipline.Source: §4.1.3 · p. 15
4.2. RQ2: How do patterns of data-centric sensemaking afford potential data reuse?
- Participants often imagined reuse by people close to the topic, such as reanalysis under another model or confirmation of an existing interpretation.Some believed meaningful reuse required same-discipline expertise, specific programs, or knowledge of evolving research practices; many struggled to envision uses outside their field.Source: §4.2 · p. 15
4.2.1. Structures needed for sensemaking and reuse
- Some participants said shared domain structures, terminology, and methods let data “speak for themselves” to others with similar expertise.One participant expected such colleagues would understand the data without a description.Source: §4.2.1 · p. 15
- Missing column names and data split across sheets could obscure meaning, so documents and codebooks support reuse across domains and by original producers later.Participants also described common domain structures, terminology, and methods as allowing data to “speak for themselves” to similarly expert users.Source: §4.2.1 · p. 15
- Producers may need codebooks and records of variable decisions and research instruments when returning to data years later.One participant said that after ten years they would need to consult the codebook despite having once known the data by heart.Source: §4.2.1 · p. 15
- Documentation formats were shaped by journal, schema, repository, and usability constraints; information could be separate or embedded, and different consumers might need different structures.Examples included README files, feature and programming documentation, tutorials, papers, and community support.Source: §4.2.1 · p. 16
- Information needs varied with familiarity: close consumers often wanted granular creation conditions, prior work, and variable uses, while descriptions could also emphasize common attributes or basic fields.Descriptions for data scientists could be more granular about benefits and problems; epidemiologists might need study design, participant origins, and dates. The authors frame these needs by distance and method, reflecting participant scenarios rather than a comprehensive list.Source: §4.2.1 · pp. 16–17
- Method-specific information mattered: quantitative accounts needed experimental setup and real-world differences, while qualitative accounts needed categories, representativeness, study setup, and social context.For qualitative work, social context could affect whether participants felt safe enough to answer truthfully.Source: §4.2.1 · pp. 17–18
5. Discussion
The discussion synthesizes data sensemaking into three clusters—inspecting, engaging with content, and placing data in context—and relates these activities to the information structures needed for reuse. It uses the patterns to motivate flexible, workflow-compatible design recommendations.
5. Discussion
- Sensemaking clustered into inspecting data, engaging deeply with its content, and placing it in broader contexts; C1 was mostly done alone, while C2 and C3 were often collaborative.Participants also referenced missing contextual information, such as purpose and collection methods, as a stumbling block when understanding unfamiliar data.Source: §5 · pp. 18–19
5.1. Understanding shape
- First-time inspection could be linear, treating attributes one by one, or interwoven, mixing attribute descriptions with analyses and questions.Interwoven inspection involved more frequent shifts between a dataset-wide view and specific values or parts.Source: §5.1 · p. 20
5.1.1. Recommendations
- Support whole-dataset and column-level views with summaries, interactive visualizations, fast scanning and zooming, and horizontal scrolling for wide spreadsheets.A README can support zoomed-out inspection, while an interactive map may support zoomed-in exploration and C2 activities.Source: §5.1.1 · pp. 20–21
5.2. “Strange things” as an entrypoint, not an obstacle
- Participants treated outliers, errors, missing data, and formatting inconsistencies as prompts for C2 activities such as expressing uncertainty, seeking relationships, and analyzing data.They emphasized communicating errors and uncertainties, but some information was embedded in data or not sustainably linked, risking machine readability or long-term preservation of meaning.Source: §5.2 · p. 21
5.2.1. Recommendations
- Rather than simply cleaning data, tools could flag anomalies, link column summaries to explanations and supporting information, and connect flags to sustainable documentation.Interactive visualizations could highlight relationships or trends, while producer-identified anchor variables could focus summarization.Source: §5.2.1 · p. 21
5.3. Perspectives in placing
- Placing data draws on related contexts, including disciplinary norms and study methods; consumers need narratives explaining why methods were chosen and local setup constraints.Narrative granularity relates to consumer expertise or distance from the data; study objectives, usage restrictions, and explanations of categories and acronyms were identified as useful regardless of expertise.Source: §5.3 · pp. 21–22
5.3.1. Recommendations
- Flexible tools could let consumers drill down to their desired level of detail and link disciplinary terms, geographic information, standards, and provenance to relevant context.Focused documentation across media and adaptable representations could support the producer-consumer conversation implicit in data reuse while fitting researchers’ workflows.Source: §5.3.1 · p. 22
6. Study Limitations
The study’s sample was limited to researchers with Scopus-indexed publications, and the authors identify limits on applying its findings beyond that population. The data selected and the order of the tasks may also have shaped the observed patterns.
6. Study Limitations
- The sample consisted of researchers with Scopus-indexed publications, and Scopus underrepresents some fields, including Arts & Humanities.The selection criteria also ensured participants had data experience.Source: §6 · pp. 22–23
- The findings may not directly apply to people outside the researcher population; applying them more broadly requires repeating the study with other populations.The authors suggest that general patterns may transfer across skill levels, while how people achieve sensemaking goals may differ with data literacy.Source: §6 · p. 23
- Data, research, and sensemaking practices vary with social, legal, and economic contexts specific to countries and disciplines.Participants came from varied countries and disciplines, but not all were native English speakers; the sample was selected for high English fluency.Source: §6 · p. 23
- Observed patterns may depend on the data used, and describing familiar data first may have primed participants’ summaries of the unfamiliar data.Including participants’ own data alongside an unfamiliar dataset was intended to balance the potential data-related bias.Source: §6 · p. 23
- Participants were not asked to bring metadata for familiar data, allowing the study to capture what information they chose to bring and what they felt was missing from unfamiliar data.The authors note that data may lack complete, accurate, or meaningful metadata or enough information for reuse.Source: §6 · p. 23
7. Conclusions
The study examined initial encounters with data in relation to potential reuse and identified three clusters of sensemaking activities, with associated activities and data attributes. The authors connect these findings to design efforts and identify further work to validate and apply them across contexts.
7. Conclusions
- Data-centric sensemaking involves inspecting data, engaging with its content, and placing it in context, with observed activities and relevant data attributes detailed for each cluster.The authors present these patterns as a way to focus design efforts on data consumers’ sensemaking needs.Source: §7 · p. 23
- The paper contributes activity patterns, a framework linking patterns to data attributes, user needs for data reuse, and design recommendations.The study focuses on initial encounters with data in light of potential reuse.Source: §7 · p. 23
- The authors identify future work to validate the patterns with different data or non-research participants and to apply the findings within existing workflows, services, and platforms.They particularly note multidisciplinary data as a context for potential platform integration.Source: §7 · p. 23
- In this study, data-centric sensemaking is framed as supporting data reuse in research, making understanding data relevant to realizing reuse’s potential.The authors describe sensemaking as enabling action, with data reuse as the relevant action in this work.Source: §7 · p. 23