W. Jentner, G. Lindholz, H. Hauptmann, M. El-Assady, K.-L. Ma, and D. Keim. Visual analytics of co-occurrences to discover subspaces in structured data. ACM Transactions on Interactive Intelligent Systems, 13(2):#10, 49 pages, 2023. DOI.
This Case Study Report was authored in 2026 by Wolfgang Jentner, University of Arizon, USA
This case study provides a post-hoc application of Chen’s and Ebert’s methodology [2019] to the multi-dimensional pattern exploration technique and the developed prototype by Jentner et al. [2023]. Subspace analysis of structured data is relevant for tasks such as cohort analysis of patient histories and pharmaceutical research. The data at hand consists of structured data in the form of sets, sequences, trees, or graphs, and is paired with discrete attributes such as gender, age, and patient medical conditions. The goal is to find correlations between sub-structures of the structured data and the attributes based on co-occurrence distributions. An example of cohort analysis would be that a certain subset of patients with similar medical histories correlates with attributes of men over the age of 60, having type 2 diabetes, and hypertension. This is challenging as both the sub-structures as well as combination of attributes grow exponentially. Moreover, the interestingness of correlations based on co-occurrence distributions is application-, user-, and task-dependent and cannot be easily automated to filter the computational space.
A dashboard with cross-filtering features allows the user to filter the data by multiple attributes, running algorithms and statistics only on the user-filtered subset (interaction first approaches). Therefore, the symptoms of such approaches are high interaction cost and high alphabet compression, but they simultaneously introduce high potential distortion, as users cannot execute the algorithm for every possible filter setting. Algorithm-first approaches that execute over the entire dataset require a-priori knowledge of the user to identify potential correlations, thus increasing their alphabet compression. Inappropriate parameter settings for algorithms cause an explosion in visualization costs, as an exponential number of results is generated and must be assessed by the user.
Table. Summary of the ontological framework applied to the design process of the multi-dimensional pattern exploration technique and application.
Jentner et al.’s approach is algorithm-first, but it introduces an algorithm that eliminates redundant data by linearizing one of the exponential dimensions and significantly reducing the second dimension. While this introduces some algorithmic cost, it eliminates interaction cost, as a single-pixel-based visualization can be rendered. This visualization is difficult to interpret and navigate as the information density is high (high visualization cost).
As a remedy to the high visualization cost, interactive selection was introduced, allowing users to select specific rows in either table and have the visual links highlighted, rather than displaying all links simultaneously. The application introduced zooming, panning, and tooltips to support visual navigation. These features slightly increase interaction cost but significantly reduce the risk of inattentional blindness by allowing the user to focus on manageable subsets of the pixel space. Furthermore, the application allows switching normalizations and adjusting pixel color, enabling users to reveal different visual patterns from the same underlying data. This corresponds to reducing the potential distortion of the visualization to match the user’s current analytical question. To further improve this, tables can be reordered based on user-defined criteria (e.g., interestingness measures such as support, LIFT, or confidence), placing regions of potential interest at the top or bottom and supporting navigation. Rows outside user-defined boundaries can be removed, reducing visual complexity. This complements sorting by narrowing the displayed data to regions of interest.
The approach inverts the conventional workflow of starting with attributes to filter data. Instead, it presents all sub-structures and their co-occurrences simultaneously, requiring users to adopt a new mental model. Combined with the accumulated interactive features, this steepened the learning curve for novice users, who were unsure how correlation patterns would manifest in the visualization. As a remedy, features were introduced that allow users to search for specific structured data patterns (e.g., "the car drives this road and then that one"), filter the table directly, and support hypothesis testing. This reduces the interaction cost by providing a familiar entry point into the otherwise unfamiliar visual representation. Additionally, the system lists potential correlation patterns of interest. Clicking on a suggestion highlights the relevant rows and automatically pans and zooms the canvas to the correct location. This serves a dual purpose: it reduces the cost of initial exploration and helps users learn how various correlation patterns are visually represented in the pixel space, thereby reducing the long-term cognitive cost.
Figure. A workflow for optimizing the baseline workflow enabled by a dashboard based on filtering algorithms and interactive queries.