Exploring Millions of User Interactions with ICEBOAT: Big Data Analytics for Automotive User Interfaces

Modern touchscreen-based IVISs offer many features that must be evaluated with respect to driving context, making design increasingly complex. Current evaluations rely heavily on qualitative feedback and small-scale studies that are slow, expensive, and do not scale. Despite rich telematics data in vehicles, UX experts lack access and tools for effective visualization and analysis. The paper argues that big data visualization tools that automate processing and present domain-relevant visualizations are critical to enable evidence-based decisions in automotive UX. The authors propose ICEBOAT, an interactive tool that visualizes user interactions, driving, and glance data collected from production vehicles, enabling task-specific analyses, drill-down from overview to details, and comparisons on performance- and distraction-related metrics (e.g., time-on-task, number and duration of glances). The research aims to identify UX experts’ needs, design a tool meeting those needs, and evaluate whether it empowers experts to analyze large-scale IVIS usage data independently.

The background highlights three areas: (1) Data in the design and evaluation of IVIS: Automotive UI evaluation often depends on questionnaires and controlled studies that are slow and resource intensive, limiting rapid insight generation. In contrast, web/app domains leverage continuous data collection and analysis; automotive culture and tooling lag behind, forcing UX experts to rely on data scientists. (2) Creating meaningful interactions with big data: The challenge is extracting knowledge from data and communicating results via meaningful, task- and domain-specific visualizations. Standard dashboards often serve reporting rather than exploration and don’t align with domain workflows; linked, multi-scale visualizations reduce complexity and support exploratory analysis. (3) Big data visualizations for evaluating automotive UIs: Evaluations must include distraction metrics (e.g., long off-road glances >2 s correlate with crash risk). Existing academic tools focus on controlled studies or few sensors and are not scalable to millions of OTA events. Prior work by Ebel et al. proposed multi-level visualizations for large-scale automotive interaction data but lacked a workflow-integrated tool and validation with practitioners. This work addresses that gap by specifying experts’ information and interaction needs and implementing an integrated tool.

The authors followed a mixed-methods User-Centered Design (UCD) approach: (1) Study 1: Semi-structured background interviews (n=4) with experienced automotive UX experts (5–9 years professional experience, at least 5 in UI design, >5 years at Mercedes-Benz) to confirm usefulness of prior visualizations and extract detailed information and interaction needs. The interview protocol included open-ended and follow-up questions, iterative summarization for validation, and presentation of prior visualizations to support ideation. From coded transcripts, 39 information needs were identified in seven categories (INF-1 to INF-7), and 14 interaction needs in four categories (INT-1 to INT-4). (2) Co-design: Four iterative participatory design sessions with the same experts to refine requirements and design ICEBOAT’s workflow, visualizations, and interactions (task definition, flow exploration, comparison, sequence details). (3) System design and architecture: ICEBOAT comprises a Vue.js web frontend and a backend with a FastAPI service and a PySpark-based data service connected to a data lake updated daily. The frontend integrates a Concept Database (UI metadata), an IVIS Emulator (for task recording/definition), and the backend analytics. Users can define tasks via searchable UI element selection or by recording flows with the emulator; the backend extracts matching sequences, aggregates flows, and returns metrics and visualizations. The analysis workflow includes a dashboard (onboarding, KPIs), a Task Overview Panel (adapted Sankey plus table metrics), filters (e.g., software version, car type), a Flow Comparison Panel (reduced Sankey and box plots by chosen metric), and a Sequence Details Panel (aligned interaction, glance, and driving signals over time with history/favorites). (4) Study 2: Evaluation with practitioners. Method: remote usability testing via Zoom, think-aloud, seven tasks guided by a realistic scenario (destination entry in navigation), plus interviews and a post-study survey including System Usability Scale (SUS) and a Context of Use questionnaire. Participants: N=12 (4 designers, 4 UX researchers, 4 data scientists) from Mercedes-Benz/MBition; ages 21–41 (M=29.6, SD=5.6); work experience 0.5–20 years (M=5.3, SD=5.8); all but one with college degree. Data collection included error counts, qualitative feedback, SUS, Context of Use scores, and notes on technical issues.

- Information and interaction needs: Identified 39 information needs across 7 categories (INF-1 usability/distraction metrics; INF-2 feature usage; INF-3 usage pattern visualizations; INF-4 system information; INF-5 contextual information; INF-6 input modalities; INF-7 user information) and 14 interaction needs across 4 categories (INT-1 task definition; INT-2 analysis including filters, drill-down, comparisons; INT-3 operating aids; INT-4 sharing/collaboration). Prior visualizations partially addressed INF-1, INF-2, INF-3, INF-5 but not INF-4, INF-6, INF-7. - ICEBOAT capabilities: Automates task definition (via UI element selection or IVIS emulator recording), data processing, and multi-level visualizations; supports filtering by system parameters, comparisons of flows on metrics like time-on-task, number/total duration of glances, and detailed sequence inspection aligning interactions, glance, and driving data. - Evaluation (N=12): - SUS: Mean 68.125 (MD=70, SD=16.89), overall average usability. Data scientists rated higher (mean ≈ 80) than UX experts (mean ≈ 62), mirroring differences seen in other big data platforms. - Context of Use: Mean 4.2/5 (MD=4.24, SD=0.33). Pearson correlation with SUS not significant (R=0.55, p=0.061), suggesting perceived workflow value is not directly tied to SUS. - Errors: Few when using ICEBOAT; some users initially mis-set minimum support or used checkboxes vs. filter inconsistently. More errors occurred in the emulator (5/12 captured the expected start/end), considered study-induced and not practically critical. - Qualitative: Participants reported faster insight generation and independent access to telematics data; appreciated emulator-based task definition; easily identified bottlenecks in Sankey views; could compare flows by glance metrics after brief explanation; sequence details enabled reasoning about glance–interaction–driving dependencies. Insights included suggestions to improve search suggestions to reduce typing. Overall, ICEBOAT enabled efficient, data-driven evaluation of IVIS user flows and supported exploration from aggregate patterns to individual sequences.

The findings demonstrate that a domain-specific visual analytics tool designed through UCD can bridge the gap between abundant telematics data and UX practitioners’ ability to make evidence-based decisions. ICEBOAT addresses three core challenges: (1) data-driven decision making by granting direct, independent access to large-scale interaction, glance, and driving data; (2) automotive-specific analysis by aligning task definition, flow exploration, and sequence-level inspection with IVIS workflows and safety/usability metrics; and (3) information overload by providing progressive disclosure through a drill-down workflow, pre-defined KPIs, and an emulator-based, low-overhead task definition mechanism. The tool’s average SUS coupled with high Context of Use scores suggests that while usability can be further refined—especially for non-data specialists—the tool already fits practitioners’ workflows and enables meaningful insights. The mixed feedback on interpreting certain signals (e.g., steering angle changes) highlights the need for in-tool guidance and domain annotations. Collectively, the results indicate ICEBOAT can democratize access to telematics analytics within automotive UX teams and inform both incremental UI improvements and the scoping of focused user studies to probe underlying causes.

ICEBOAT is an interactive, end-to-end analytics tool that makes millions of in-vehicle user interactions accessible for efficient evaluation of touchscreen-based IVISs. Through Study 1, the authors distilled detailed information and interaction needs of automotive UX experts, revealing a tension between the desire for rich data and the complexity of big data tools. ICEBOAT resolves this by enabling emulator-based task definition, automating data processing/cleaning while exposing key metric controls, and providing a drill-down analysis path from overview to per-sequence detail. Study 2 shows that UX experts and data scientists can effectively use ICEBOAT to visualize and evaluate large-scale automotive usage data, supporting evidence-based design decisions and contributing to the democratization of data in the automotive domain. Future work should expand modalities (e.g., speech, hardkeys), refine usability for non-specialists, and broaden validation beyond a single OEM.

- Usability disparity: Data scientists rated usability higher than UX experts, suggesting a learning curve and a need for additional onboarding/support for non-specialists. - Remote testing artifacts: Screen sharing/remote control issues occurred (4 cases; 1 participant couldn’t complete a task without assistance), potentially affecting results. - Modalities scope: The current study/tool focuses on center stack touch interactions; speech and hardkey inputs are not yet included, limiting INF-6 coverage. - Single-OEM perspective: Interviews/evaluation were conducted within one OEM; while practices may generalize, needs could be skewed. - Privacy constraints: Lack of personal data limits direct user-group analyses; only indirect target group approximations via available filters are possible for INF-7. - Interpretation variability: Some signals (e.g., steering angle changes) were interpreted inconsistently, indicating a need for guidance/annotations.