{"id":"def8b878-d041-47ad-843d-19f5e0e812af","arxiv_id":"2505.07534","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper presents the HDMI Canvas, a six-flow framework characterizing how humans, data, and models both contribute to and benefit from visual analytics processes.","lead":"This position paper introduces the HDMI Canvas, a structured template that maps visual analytics systems onto six flows among humans, data, and models. It is meant to help researchers and practitioners describe existing systems and design new ones in terms that external stakeholders can understand.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'Data Exploration' arrow is not a contribution by data but an activity performed on data, breaking the six-flow symmetry and undermining descriptive power.","rationale":"I read the paper as a position piece whose central claim is that the HDMI Canvas, with its six directed flows, provides descriptive and generative power for VA. For that claim to hold, each of the six flows must correspond to a well-defined contribution or benefit by the named actor. Section 3.2 reveals that the 'Data Exploration' contribution is not performed by the data actor: the bullet list describes activities that humans and models perform on data. This is an internal inconsistency, not a matter of empirical validation. The reader's weakest assumption (unvalidated decomposition) is close, but the problem is sharper: the decomposition as specified is conceptually flawed for one arrow regardless of data. The two case studies do not test the framework; they illustrate it, and the way they score the data arrow shows the ambiguity. The paper's own Section 6 discussion of data actors highlights the unresolved question. I therefore recommend keeping the conditional verdict, but the condition should include clarifying the semantics of the data contribution arrow, either reframing it as 'data provision/feedback' or explicitly defining what active role data plays. My proposed inter-rater test would provide empirical evidence on whether the flow can be applied discriminatively.","tokens_in":11184,"tokens_out":4970,"duration_ms":45812,"concrete_test":"Use the definitions in Section 3.2 to assign a grammatical subject to each of the six arrows; if the subject of 'Data Exploration' is not 'data' (e.g., it is the human/user or the VA system), the arrow mis-assigns the contribution. Empirically, run an inter-rater study: give three raters the canvas definitions plus 10 VA systems (iPCA, IRVINE, and 8 others), and ask them to mark which of the six flows are present. Measure agreement on the 'Data Exploration' flow specifically. If agreement is low, or raters ask whether 'data exploration' means exploration by data or exploration of data, the descriptive-power claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2 the paper defines the EEE Framework as the three contributions: 'Externalization of human knowledge' (by humans), 'Data Exploration' (by data?), and 'Explanation of models' (by models). The first and third are agentive: the actor named is the subject of the verb. The middle one is not: data is the object of exploration, not its agent. The bullet list under 'Data Exploration' ('Representation of patterns...', 'Detection of clusters...', 'Relation and association discovery...') describes what analysis does to data, not what data contributes. This breaks the symmetrical 3×2 actor-contribution/benefit structure that the abstract and Section 4.2 claim. The case studies confirm the problem: iPCA's data exploration steps (1)(2) are user interactions with the visualization; IRVINE's step (1) is engineers exploring data. In neither is data the contributor. Consequently the 'Data contribution' arrow is scored present whenever any VA system supports exploration, so it cannot discriminate between passive-repository and active-actor views of data — directly undercutting the paper's descriptive-power claim. Section 6 acknowledges the 'Data Actors?' debate but does not reconcile it with the arrow's semantics. This is an internal inconsistency, not merely missing empirical validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper reviews 16 visual analytics process models and frameworks, distills nine observations from that reflection, and proposes the HDMI Canvas, a conceptual structure in which humans, data, and models are treated as actors that both benefit from and contribute to VA processes. The contribution side is formalized as the EEE Framework: Externalization of human knowledge, Exploration of data, and Explanation of models. The paper claims descriptive power for differentiating VA building blocks and generative power for designing new VA processes, and it illustrates the canvas with two case studies (iPCA and IRVINE). It also discusses limitations, including the absence of empirical evidence for the utility claim.","tokens_in":11428,"tokens_out":3490,"duration_ms":32649,"significance":"If the conceptual structure were sound, the HDMI Canvas would be a useful complement to existing VA process models, especially for communication with external stakeholders and for shifting attention from commonalities to differences among systems. The paper's strengths include its broad literature reflection, the clear visual metaphor of six directed flows, and the inclusion of modern human-centered concerns such as knowledge externalization, feedback loops, and explainable AI. The two case studies help illustrate how existing systems can be mapped onto the canvas. However, the central descriptive and generative claims rest on a semantic asymmetry in the EEE Framework and on illustrative rather than evaluative case studies, as the author explicitly acknowledges in Section 6.","major_comments":[{"comment":"The EEE Framework's 'Data Exploration' arrow is not a contribution by data. 'Externalization of human knowledge' and 'Explanation of models' name the actor as the subject of the verb, but 'Data Exploration' names an activity performed on data; the bullet list under it (representation of patterns, detection of clusters, relation discovery, etc.) describes what analysts or models do to data, not what data contributes. The case studies confirm this reading: iPCA's data exploration steps (1) and (2) are user interactions with the visualization, and IRVINE's step (1) is engineers exploring data. As a result, the data-contribution arrow is satisfied by any VA system that supports exploration, so it cannot discriminate between a passive-repository view of data and an active-actor view, directly undercutting the paper's descriptive-power claim. Section 6's 'Data Actors?' discussion acknowledges the debate but does not reconcile it with the arrow's semantics. The paper should either redefine the arrow as a genuine data contribution (e.g., data affordances, query responses, data-driven constraints) or explicitly reframe it as a process contribution and accept the resulting asymmetry.","section":"Section 3.2 / Figure 1"},{"comment":"The abstract states that utility is 'demonstrated through two preliminary case studies,' but Section 6 says 'empirical data does not yet exist to back up the utility claim made.' These statements are in tension. The two case studies show that existing systems can be mapped onto the canvas, but they do not show that the canvas eases differentiation or guides design better than alternative taxonomies. Since descriptive and generative power are the paper's central claims, the paper should either weaken the wording to 'illustrated' and 'hypothesized' or add an evaluation protocol, such as independent analysts applying the canvas to a diverse set of systems and measuring agreement, coverage, and discriminative value.","section":"Abstract / Section 6"},{"comment":"The IRVINE case study is drawn from the author's own prior system, which raises a mild self-reference concern. Because the canvas was developed after that system, the mapping may be post hoc and self-confirming. The paper should acknowledge this selection bias and ideally add at least one third-party system mapped by researchers not involved in the canvas design to strengthen the descriptive-power argument.","section":"Section 5.2"},{"comment":"The paper claims generative power for the canvas, but the only supporting evidence is retrospective mapping of existing systems; no prospective example is given of the canvas being used to design a new VA process or system. As a position paper, this may be acceptable if framed as a hypothesis, but the current phrasing in Section 4.3 and the Conclusions overstates what has been shown.","section":"Section 4.3"}],"minor_comments":[{"comment":"There is a typo in 'dynamic paramete fine-tuning'; it should read 'dynamic parameter fine-tuning.'","section":"Section 5.2"},{"comment":"The term 'Supertool' is used to describe a fully integrated VA system, but it is not defined before its first use; a brief definition or reference would help readers.","section":"Section 4.3"},{"comment":"In the 'What is Knowledge?' discussion, the relationship between the 'knowledge and preferences as a tandem' statement and the human-benefit bullet 'Insight, evidence, and knowledge generation' could be clarified, since the canvas elsewhere seems to treat knowledge generation as a human benefit.","section":"Section 6"},{"comment":"The EEE acronym is memorable, but the phrase 'Exploration of data' should be reworded or qualified to make clear whether the contribution is made by data, by the VA process, or by human and model actors acting on data; this wording issue is related to the major semantic concern above.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The core conceptual issue is the semantic asymmetry of the 'Data Exploration' arrow, which is load-bearing for the descriptive-power claim. This is fixable by redefining the arrow or reframing the framework's claims, so I recommend major revision rather than rejection. The lack of empirical validation is acceptable for a position paper only if the claims are consistently framed as hypotheses; the abstract's 'demonstrated' wording currently overreaches."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The HDMI Canvas is a useful piece of synthesis: it reorganizes known VA process models into a single visual vocabulary and gives the community a shared way to talk about contributions and benefits of humans, models, and data. The six-flow structure is new relative to the 16 cited models, and the paper is honest about the absence of empirical validation. If it works, it's a helpful outreach and design aid.\n\nBut there is a real internal inconsistency in the 'Data Exploration' arrow. The EEE framework pairs Externalization of human knowledge and Explanation of models as agentive contributions: the actor named is the subject of the verb. Data Exploration is not that. Data is the object, not the agent. The bullet list under it ('detection of clusters,' 'relation discovery') describes what the analyst does to data. The case studies confirm it: iPCA's exploration is user interaction; IRVINE's step (1) is engineers exploring data. So the arrow gets checked for any VA system that supports exploration, which means it cannot discriminate between passive-repository and active-actor views of data. That undercuts one of the paper's main claims. Section 6 acknowledges the 'Data Actors?' debate but does not reconcile it with the arrow's semantics. This is a conceptual flaw, not just a missing evaluation.\n\nThe other soft spots are milder. The utility claim is unvalidated, but the author says so explicitly. One of the two case studies is the author's own prior system, which is routine in a position paper but does weigh lightly on the evidence. The building-block lists are hand-assembled and could be more systematically derived, though they're reasonable as a starting vocabulary.\n\nFor a workshop position paper, this is a solid, honest submission. It doesn't overclaim, and the framework is mostly coherent. The data arrow problem is fixable: rename or reframe it as 'data-driven exploration' or 'data accessibility' and be explicit about whether data is a passive informant or an active contributor. With that fixed and the claims softened, it's a good contribution to EuroVA.\n\nI'd send this to a serious referee. It deserves peer review, not a desk reject. The framework is likely to be discussed and used regardless, so better to have it reviewed and sharpened.","headline":"A useful VA synthesis canvas whose 'data contribution' arrow is semantically broken; honest position paper worth peer review with a requested revision.","tokens_in":11910,"tokens_out":3492,"would_cite":true,"duration_ms":31831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that visual analytics can be mapped as six flows between humans, data, and models, and that this map tells systems apart and guides new designs.","keywords":["visual analytics","human-data-model interaction","HDMI Canvas","EEE framework","knowledge externalization","explainable AI","process models","design space"],"falsifier":"A survey of a broad sample of visual analytics systems that yields a substantial fraction of systems that cannot be assigned to the six flows, or a study in which independent analysts using the HDMI Canvas produce conflicting descriptions of the same system, would show the canvas does not reliably differentiate approaches.","tokens_in":1575,"feed_emoji":"🧭","tokens_out":2309,"duration_ms":57702,"temperature":0.7,"pith_summary":"This position paper argues that visual analytics has been underserved by process models that mainly describe a one-way flow from data to insights. It proposes a more balanced picture in which humans, data, and models are all actors who both contribute to and benefit from a visual analytics process, connected by six directed flows. If the canvas is adopted, it would give practitioners a common language to distinguish concrete visual analytics systems rather than only stating general principles, and it would give external stakeholders a way to see what visual analytics can offer.","feed_headline":"Three actors, six arrows: a canvas for visual analytics","feed_subtitle":"The map shows what humans, data, and models each give and get, helping designers tell systems apart.","key_machinery":"The central object is the HDMI Canvas, a 3 by 2 arrangement of six directed arrows connecting humans, data, and models to the visual analytics process. The three incoming arrows are the EEE contributions: Externalization of human knowledge and preferences, Exploration of data, and Explanation of models; the three outgoing arrows are the benefits each actor receives. The canvas carries the argument by turning each arrow into a concrete list of visual analytics building blocks, so that a system can be described by which flows it uses and a designer can use the lists as a checklist for constructing new systems.","core_discovery":"The central claim is the HDMI Canvas: a structured perspective on visual analytics that positions humans, data, and models as key actors, with three incoming arrows representing contributions to the process and three outgoing arrows representing benefits from the process. The contributions are the EEE framework: Externalization of human knowledge and preferences, Exploration of data, and Explanation of models. The benefits are the outcomes each actor receives, such as knowledge and insight for humans, enriched and reusable data for data, and improved, explained, and trusted models for models. The paper argues that this six-flow structure gives visual analytics enhanced descriptive power, because it can differentiate between concrete building blocks of existing systems, and generative power, because it can guide the design of novel visual analytics processes. Two case studies, iPCA and IRVINE, are used to show how systems can be characterized by which of the six flows they support.","pith_inferences":["A natural extension would be to turn the canvas into an evaluation rubric: future work could code a corpus of visual analytics papers by the six flows and test whether that coding predicts system adoption, user satisfaction, or analytical success.","Because the canvas treats data as an active actor rather than a passive repository, it could connect to broader debates about data feminism and data humanism, though the paper leaves that link implicit.","The building-block lists could be tested empirically through a card-sorting or annotation study, measuring whether independent analysts agree on which blocks a given system uses.","The six-flow structure could be extended with additional arrows for ethics, governance, or societal impact, but the paper does not propose such an extension."],"forward_implications":["Existing visual analytics systems can be compared by which of the six contribution and benefit flows they realize, as iPCA realizes five flows and IRVINE realizes all six.","The canvas provides a generative design checklist: new visual analytics processes can be generated by deciding which actors contribute and benefit, with a fully integrated human-data-model Supertool when all six flows are present.","Human knowledge externalization becomes a first-class contribution alongside data exploration and model explanation, aligning the field with modern feedback loops, interactive labeling, and explainable AI.","The canvas vocabulary is intended to be accessible to external stakeholders, so it can support interdisciplinary project scoping and user-centered design without requiring prior visual analytics training.","The two case studies suggest the canvas can describe both research prototypes and real industrial systems, supporting its use for communication and outreach."],"supporting_citations":[{"why":"Defines visual analytics as a process from data to knowledge and serves as the baseline model the canvas extends.","marker":"[KAF*08]"},{"why":"The knowledge generation model is the dominant recent process model that the canvas complements.","marker":"[SSS*14]"},{"why":"The KDD pipeline exemplifies the traditional data-to-insight flow that the paper observations critique.","marker":"[FPSS96]"},{"why":"Applies knowledge conversion processes to a visual analytics system and grounds the human knowledge externalization contribution.","marker":"[WJD*09]"},{"why":"Provides a conceptual model of knowledge-assisted visual analytics that supports the externalization arrow.","marker":"[FWR*17]"},{"why":"Introduces the human-is-the-loop framing that broadens the feedback-loop concept in the canvas.","marker":"[EHR*14]"},{"why":"The VIS4ML ontology supplies building blocks for how models both contribute to and benefit from visual analytics.","marker":"[SKKC19]"},{"why":"The iPCA case study demonstrates a system that realizes five of the six canvas flows.","marker":"[JZF*09]"},{"why":"The IRVINE case study demonstrates a real industrial system that realizes all six flows.","marker":"[EBJ*22]"},{"why":"The Business Model Canvas provides the structural inspiration for using a canvas format to organize strategic analysis.","marker":"[OP10]"}],"fun_headline_variants":["Canvas maps the give-and-take of humans, data, and models","HDMI Canvas: descriptive and generative power for VA","Six flows that define visual analytics processes","A fresh perspective on roles in visual analytics","The HDMI Canvas: who contributes, who benefits in VA"],"cache_read_input_tokens":14080,"weakest_assumption_plain":"The load-bearing premise is that visual analytics processes can be meaningfully decomposed into exactly these six predefined contribution and benefit flows, and the paper explicitly states in Section 6 that no empirical data yet backs the utility claim.","fun_headline_variants_meta":{"raw":{"variants":["Canvas maps the give-and-take of humans, data, and models","HDMI Canvas: descriptive and generative power for VA","Six flows that define visual analytics processes","A fresh perspective on roles in visual analytics","The HDMI Canvas: who contributes, who benefits in VA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1267,"prompt_tokens":892,"completion_tokens":375,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":299}},"tokens_in":508,"tokens_out":375,"duration_ms":3987,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:13:43.639311+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A survey of a broad sample of visual analytics systems that yields a substantial fraction of systems that cannot be assigned to the six flows, or a study in which independent analysts using the HDMI Canvas produce conflicting descriptions of the same system, would show the canvas does not reliably differentiate approaches.","supporting_citations":[],"review_version":1}