Anamaria Crisan

Anamaria Crisan, Margaret Drouhard, Jesse Vig et al. · salesforce

Deep learning models for natural language processing (NLP) are increasingly adopted and deployed by analysts without formal training in NLP or machine learning (ML). However, the documentation intended to convey the model's details and appropriate use is tailored primarily to individuals with ML or NLP expertise. To address this gap, we conduct a design inquiry into interactive model cards, which augment traditionally static model cards with affordances for exploring model documentation and interacting with the models themselves. Our investigation consists of an initial conceptual study with experts in ML, NLP, and AI Ethics, followed by a separate evaluative study with non-expert analysts who use ML models in their work. Using a semi-structured interview format coupled with a think-aloud protocol, we collected feedback from a total of 30 participants who engaged with different versions of standard and interactive model cards. Through a thematic analysis of the collected data, we identified several conceptual dimensions that summarize the strengths and limitations of standard and interactive model cards, including: stakeholders; design; guidance; understandability & interpretability; sensemaking & skepticism; and trust & safety. Our findings demonstrate the importance of carefully considered design and interactivity for orienting and supporting non-expert analysts using deep learning models, along with a need for consideration of broader sociotechnical contexts and organizational dynamics. We have also identified design elements, such as language, visual cues, and warnings, among others, that support interactivity and make non-interactive content accessible. We summarize our findings as design guidelines and discuss their implications for a human-centered approach towards AI/ML documentation.

Pouya Sadeghi, Anamaria Crisan, Jimmy Lin

Large language model (LLM) leaderboards rank AI models using standardized benchmarks and have become highly visible across computer science, despite known limitations in their reliability and robustness. Yet how they shape researchers' actual practice remains empirically uncharted. We address this gap through semi-structured interviews with eight researchers across four computer science subfields, analyzed using reflexive thematic analysis. We find a near-universal paradox of pragmatic skepticism: while participants expressed deep distrust of leaderboard rankings, they continued to use them as rough decision-making aids. Peer networks, not leaderboards, emerged as the primary model selection mechanism, and arena-based (human-voting) leaderboards were consistently preferred over static benchmark leaderboards. Leaderboard influence varied sharply across subfields, revealing that disciplinary culture, not individual attitudes, mediates engagement; for instance, NLP researchers faced state-of-the-art comparison pressure while HCI and Systems/Privacy researchers reported none. Across these differences, however, participants converged on cost transparency as the most demanded missing feature (seven of eight). We translate these findings into concrete design recommendations that align evaluation infrastructure with how researchers actually use it, such as task-specific score breakdowns, cost integration, and voter-demographic disclosure.

Matt-Heun Hong, Anamaria Crisan

Generative Large Language Models (LLMs) show potential in data analysis, yet their full capabilities remain uncharted. Our work explores the capabilities of LLMs for creating and refining visualizations via conversational interfaces. We used an LLM to conduct a re-analysis of a prior Wizard-of-Oz study examining the use of chatbots for conducting visual analysis. We surfaced the strengths and weaknesses of LLM-driven analytic chatbots, finding that they fell short in supporting progressive visualization refinements. From these findings, we developed AI Threads, a multi-threaded analytic chatbot that enables analysts to proactively manage conversational context and improve the efficacy of its outputs. We evaluate its usability through a crowdsourced study (n=40) and in-depth interviews with expert analysts (n=10). We further demonstrate the capabilities of AI Threads on a dataset outside the LLM's training corpus. Our findings show the potential of LLMs while also surfacing challenges and fruitful avenues for future research.

Anamaria Crisan, Maddie Shang, Eric Brochu

Domain experts increasingly use automated data science tools to incorporate machine learning (ML) models in their work but struggle to "debug" these models when they are incorrect. For these experts, semantic interactions can provide an accessible avenue to guide and refine ML models without having to programmatically dive into its technical details. In this research, we conduct an elicitation study using data and visual design probes to examine if and how experts with a spectrum of ML expertise use semantic interactions to update a simple classification model. We use our design probes to facilitate an interactive dialogue with 20 participants and codify their interactions as a set of target-interaction pairs. Interestingly, our findings revealed that many targets of semantic interactions do not directly map to ML model parameters, but instead aim to augment the data a model uses for training. We also identify reasons that participants would hesitate to interact with ML models, including burdens of cognitive load and concerns of injecting bias. Unexpectedly participants also saw the value of using semantic interactions to work collaboratively with members of their team. Participants with less ML expertise found this to be a useful mechanism for communicating their concerns to ML experts. This was an especially important observation, as our study also shows the different needs that correspond to diverse ML expertise. Collectively, we demonstrate that design probes are effective tools for proactively gathering the affordances that should be offered in an interactive machine learning system.

Mohammad Abolnejadian, Shakiba Amirshahi, Matthew Brehmer et al.

In decision-making conversations, experts must navigate complex choices and make on-the-spot decisions while engaged in conversation. Although extensive historical data often exists, the real-time nature of these scenarios makes it infeasible for decision-makers to review and leverage relevant information. This raises an interesting question: What if experts could utilize relevant past data in real-time decision-making through insights derived from past data? To explore this, we implemented a conversational user interface, taking doctor-patient interactions as an example use case. Our system continuously listens to the conversation, identifies patient problems and doctor-suggested solutions, and retrieves related data from an embedded dataset, generating concise insights using a pipeline built around a retrieval-based Large Language Model (LLM) agent. We evaluated the prototype by embedding Health Canada datasets into a vector database and conducting simulated studies using sample doctor-patient dialogues, showing effectiveness but also challenges, setting directions for the next steps of our work.

Anamaria Crisan, Brittany Fiore-Gartland

AutoML systems can speed up routine data science work and make machine learning available to those without expertise in statistics and computer science. These systems have gained traction in enterprise settings where pools of skilled data workers are limited. In this study, we conduct interviews with 29 individuals from organizations of different sizes to characterize how they currently use, or intend to use, AutoML systems in their data science work. Our investigation also captures how data visualization is used in conjunction with AutoML systems. Our findings identify three usage scenarios for AutoML that resulted in a framework summarizing the level of automation desired by data workers with different levels of expertise. We surfaced the tension between speed and human oversight and found that data visualization can do a poor job balancing the two. Our findings have implications for the design and implementation of human-in-the-loop visual analytics approaches.

Anamaria Crisan, Michael Correll

Topic models are widely used analysis techniques for clustering documents and surfacing thematic elements of text corpora. These models remain challenging to optimize and often require a "human-in-the-loop" approach where domain experts use their knowledge to steer and adjust. However, the fragility, incompleteness, and opacity of these models means even minor changes could induce large and potentially undesirable changes in resulting model. In this paper we conduct a simulation-based analysis of human-centered interactions with topic models, with the objective of measuring the sensitivity of topic models to common classes of user actions. We find that user interactions have impacts that differ in magnitude but often negatively affect the quality of the resulting modelling in a way that can be difficult for the user to evaluate. We suggest the incorporation of sensitivity and "multiverse" analyses to topic model interfaces to surface and overcome these deficiencies.

Anamaria Crisan, Shannah Fisher, Jennifer L. Gardy et al.

Genomic Epidemiology (genEpi) is a branch of public health that uses many different data types including tabular, network, genomic, and geographic, to identify and contain outbreaks of deadly diseases. Due to the volume and variety of data, it is challenging for genEpi domain experts to conduct data reconnaissance; that is, have an overview of the data they have and make assessments toward its quality, completeness, and suitability. We present an algorithm for data reconnaissance through automatic visualization recommendation, GEViTRec. Our approach handles a broad variety of dataset types and automatically generates coordinated combinations of charts, in contrast to existing systems that primarily focus on singleton visual encodings of tabular datasets. We automatically detect linkages across multiple input datasets by analyzing non-numeric attribute fields, creating an entity graph within which we analyze and rank paths. For each high-ranking path, we specify chart combinations with spatial and color alignments between shared fields, using a gradual binding approach to transform initial partial specifications of singleton charts to complete specifications that are aligned and oriented consistently. A novel aspect of our approach is its combination of domain-agnostic elements with domain-specific information that is captured through a domain-specific visualization prevalence design space. Our implementation is applied to both synthetic data and real data from an Ebola outbreak. We compare GEViTRec's output to what previous visualization recommendation systems would generate, and to manually crafted visualizations used by practitioners. We conducted formative evaluations with ten genEpi experts to assess the relevance and interpretability of our results.

Anamaria Crisan, Jennifer L. Gardy, Tamara Munzner

Problem-based visualization research provides explicit guidance toward identifying and designing for the needs of users, but absent is more concrete guidance toward factors external to a user's needs that also have implications for visualization design and evaluation. This lack of more explicit guidance can leave visualization researchers and practitioners vulnerable to unforeseen constraints beyond the user's needs that can affect the validity of evaluations, or even lead to the premature termination of a project. Here we explore two types of external constraints in depth, regulatory and organizational constraints, and describe how these constraints impact visualization design and evaluation. By borrowing from techniques in software development, project management, and visualization research we recommend strategies for identifying, mitigating, and evaluating these external constraints through a design study methodology. Finally, we present an application of those recommendations in a healthcare case study. We argue that by explicitly incorporating external constraints into visualization design and evaluation, researchers and practitioners can improve the utility and validity of their visualization solution and improve the likelihood of successful collaborations with industries where external constraints are more present.