Utilizing Provenance as an Attribute for Visual Data Analysis: A Design Probe with ProvenanceLens

TL;DR

This work addresses the challenge of making analytic provenance actionable during visual data analysis by modeling provenance as user-accessible attributes. To operationalize this, it defines two provenance attributes, frequency and recency, with frequency $f_x = \frac{n_x}{\max_i n_i}$ and recency $r_x = \frac{\mathrm{rank}_N(\max(t_x))}{\sum_{i=1}^N n_i}$, and implements them in a system called ProvenanceLens. ProvenanceLens provides glyph-based encodings and data transformations (sort/filter) to map provenance to encodings such as $x$, $y$, fill and size, enabling in-the-moment reflection and post-hoc review. An exploratory study with sixteen participants shows high accuracy and confidence in answering provenance-based questions, with surprising insights that can prompt self-reflection; the findings yield design implications for integrating provenance as a core aspect of visual analysis. Collectively, the work demonstrates that provenance attributes can guide exploration, support auditing, and foster reflective practice, informing the design of future provenance-enabled visualization tools and collaborative analytics.

Abstract

Analytic provenance can be visually encoded to help users track their ongoing analysis trajectories, recall past interactions, and inform new analytic directions. Despite its significance, provenance is often hardwired into analytics systems, affording limited user control and opportunities for self-reflection. We thus propose modeling provenance as an attribute that is available to users during analysis. We demonstrate this concept by modeling two provenance attributes that track the recency and frequency of user interactions with data. We integrate these attributes into a visual data analysis system prototype, ProvenanceLens, wherein users can visualize their interaction recency and frequency by mapping them to encoding channels (e.g., color, size) or applying data transformations (e.g., filter, sort). Using ProvenanceLens as a design probe, we conduct an exploratory study with sixteen users to investigate how these provenance-tracking affordances are utilized for both decision-making and self-reflection. We find that users can accurately and confidently answer questions about their analysis, and we show that mismatches between the user's mental model and the provenance encodings can be surprising, thereby prompting useful self-reflection. We also report on the user strategies surrounding these affordances, and reflect on their intuitiveness and effectiveness in representing provenance.

Figures (6)

• Figure 1: Illustration of two provenance attributes, frequency and recency, modeled for each attribute and record of a dataset about movies, and scored on a scale from 0 (low) to 1 (high). Consider a user creates a scatterplot visualization of (1) Running Time$\times$(2) IMDB Rating and then clicks two datapoints one after another (1) Godzilla $\rightarrow$ (2) Kingpin, indicating interactions with two attributes and two records. Regarding data attributes, IMDB Rating and Running Time both receive a frequency score of 1.0 (each interacted once, hence maximum score), while other attributes score 0.0; for recency, IMDB Rating (most recently interacted) scores 1.0 and Running Time scores 0.5 (user mapped x axis before y axis), while other attributes score 0.0. Likewise, regarding data records, Godzilla and Kingpin both score 1.0 on frequency; but for recency, Kingpin (most recently interacted) scores 1.0 and Godzilla scores 0.5, while other records score 0.0. These scores are derived by evenly spacing the interactions between 0 and 1, based on their count and order of occurrence in the interaction history.
• Figure 2: Design illustrations of provenance attribute glyphs (A--S) for data attributes (or records) across different marks (point, text, bar), visual encodings (x, y, column, row, fill, fillOpacity, stroke, strokeOpacity, strokeWidth, size, shape, tooltip, annotation, text), and data transformations (sort, filter), including alternate configurations (e.g., --x where the range is reversed or descending sort order) and combinations (e.g., x + y + fill + size + sort). For instance, for mark=bar and encoding=fill (O): "Title" has the largest value (darkest bar) followed by "Worldwide Gross" , "Production Budget" , and "Genre" ; "id" , "Release Year" , and "Running Time" have the smallest values (lightest bars). Notice the change to the attribute sort order for the right side of the figure (P--S), compared to the unsorted attributes on the left (A--O).
• Figure 3: The ProvenanceLens user interface consisting of seven views: the Data Attributes view shows the attributes and enables transformation (e.g., sort, filter); the Marks and Encodings views specify the visualization; the Visualization view renders the specified visualization and supports filtering of data records; the Data Records view supports review and transformation (sort) of the data records shown in the visualization; the Provenance Attributes view lists the recency and frequency attributes; and the Tasks view shows the task instructions and questions, and tracks the user's progress.
• Figure 4: Five participants' different strategies to answer the same question, T6.Q1, "How similar were your interaction patterns for 'Comedy' and 'Thriller' movies? Illustrate via a visualization."$P_{14}$ created a scatterplot with "Rotten Tomatoes Rating" along x, "Running Time" along y, faceted by "Genre", colored by recency, and sized by frequency; $P_{9}$ and $P_{16}$ created a scatterplot with frequency and recency along x or y, and colored by "Genre"; $P_{12}$ created a scatterplot with frequency along x, "Genre" along y, and sized by recency; $P_{1}$ created an aggregate bar chart with "Genre" along x, total frequency along y, and colored by average recency. Except $P_{16}$, each of the above participants also applied a filter to only show movies that belong to the "Comedy" or "Thriller" genre.
• Figure 5: Co-occurrence statistics for how users map provenance attributes (only frequency, only recency, or either) to visual encoding combinations (A), as well as general preferences for visual encodings compared to filtering, and sorting (B). Note that these statistics correspond only to the recall (T2, T5) and visualize (T3, T6) tasks; we exclude the review (T1) and analyze (T4) tasks as they were more open-ended in nature.