{"id":"a585a07d-c4b8-44fb-aa1a-2b733011d6cf","arxiv_id":"2605.23540","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Data points genuinely similar to multiple dissimilar neighborhoods are detected as local articulation points of a sparsified high-dimensional graph and split into multiple projected copies, one per neighborhood.","lead":"This paper shows that some data points genuinely belong to two different groups at once, and introduces a way to detect such 'ambiguous' points and draw them as multiple linked copies in a data map — one copy per group. It matters because standard projections silently hide these dual memberships, misleading analyses of classifier mistakes and single-cell biology.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sparsification can manufacture local articulation points: Eq. (1) guarantees only global Laplacian approximation, not radius-r connectivity, and no experiment checks whether LAPs in the sparsified graph or the resulting split decisions are stable or present in the original graph.","rationale":"The most load-bearing step is not the split placement or the visual encoding; it is the decision about which vertices count as ambiguous. Def. 3.2 defines ambiguity relative to Gbar, not G, and Gbar is connected to G only by a global spectral inequality. A local articulation point is a discrete, non-Lipschitz property: deleting a single edge can create one without changing the Laplacian quadratic form appreciably. Since the paper's quantitative support (Sec. 4.1) measures kNN overlap rather than LAP overlap or split stability, it cannot distinguish genuine ambiguity from an artifact of aggressive sparsification. This is more fundamental than the acknowledged 'no two ambiguous instances' limitation: that limitation narrows the scope, whereas an unchecked sparsifier could invalidate every reported example. The proposed synthetic test is cheap and would settle the issue without requiring new theory. The reader already identified the global-to-local sparsification transfer as a secondary premise; I elevate it to the primary concern, hence 'partial' agreement. Since the reader's verdict was CONDITIONAL and this concern reinforces that, I recommend no change to the verdict.","tokens_in":18267,"tokens_out":6228,"duration_ms":74915,"concrete_test":"For each of the four datasets, compute LAP_r and the split set S(v,r) on G and on Gbar (same r and parameters), repeating spectral sparsification with 10 random seeds. Report (i) Jaccard overlap between LAP_r(G) and LAP_r(Gbar), (ii) the distribution of split counts per vertex across sparsification seeds, and (iii) the same on a synthetic control graph that has no articulation points but contains two dense clusters joined by multiple parallel paths. If LAPs appear in Gbar but not G, or split decisions vary strongly across sparsification seeds, or the control graph yields LAPs after sparsification, then Definition 3.2 is detecting artifacts of the sparsifier and the central claim is undermined. If overlap is high, split decisions are stable, and the control produces no LAPs, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 3.2 makes the whole pipeline rest on LAP detection in the sparsified graph Gbar: a vertex is 'ambiguous' iff it is an r-step articulation point of Gbar, and Def. 3.3 then gives it one copy per remaining component. Nothing upstream justifies that this discrete local property survives spectral sparsification. Eq. (1) is a global guarantee over all quadratic forms; it does not control the number of connected components of Gbar_r(v) after deleting v, nor whether the same vertices are LAPs at radius r. Spielman–Srivastava samples edges by effective resistance and can delete an individual edge whose removal has small global effect but disconnects a small local neighborhood; conversely it can retain spurious edges that fuse neighborhoods. The paper's own parameter choices make this more than a theoretical nuance: epsilon is set to 0.7–0.9 (Sec. 4), and the text says that a more sparse graph 'allows us to identify more ambiguous instances' — i.e., the detector's output is deliberately sensitive to a preprocessing step that has no local-connectivity guarantee. The empirical check in Sec. 4.1 (rho_HD, Fig. 10) compares kNN overlap of final embeddings from G and Gbar; even perfect kNN overlap does not imply equal articulation structure. With only five UMAP runs per graph, the confidence interval is too weak to certify preservation of local topology. Hence the core identification step — and therefore every 'revealed membership' in Figs. 1, 6–8 — may be reporting sparsification artifacts rather than genuine multi-neighborhood membership.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new visual artifact in dimensionality reduction, 'partial neighborhood embedding,' caused by 'ambiguous instances'—points highly similar to multiple mutually dissimilar high-dimensional neighborhoods. The authors formalize ambiguity as r-step local articulation points (LAPs) on a spectrally sparsified relationship graph (Def. 3.2), split each such vertex into one copy per remaining component (Def. 3.3), and re-embed the modified graph with UMAP. The pipeline is demonstrated on SVHN, MNIST, a text corpus, and single-cell RNA-seq data, with additional experiments on sparsification fidelity and on the insensitivity of standard DR quality metrics. The graph-theoretic core (Lemma 3.1) is simple and correct, but the load-bearing link between spectral sparsification and local LAP detection is only empirically supported by a weak kNN-overlap check.","tokens_in":18549,"tokens_out":2738,"duration_ms":34403,"significance":"If the identified instances genuinely correspond to multi-neighborhood membership, the work gives local DR analysis a principled way to surface structure that standard single-point projections hide. Formalizing ambiguity through graph articulation is novel and the vertex-splitting approach is plausible. The paper also correctly notes that trustworthiness and continuity cannot distinguish ambiguity-driven artifacts from other distortions. The central claim is interesting and potentially useful, but it currently rests on an unvalidated assumption about sparsification preserving local connectivity.","major_comments":[{"comment":"The detector defines ambiguity as LAP in the sparsified graph Gbar, but Eq. (1) is a global Laplacian guarantee. It does not control the number of connected components of the induced r-neighborhood after vertex removal, so sparsification can create or destroy LAPs at radius r. The paper's own parameter choices (ε=0.7–0.9, and the statement that a sparser graph 'allows us to identify more ambiguous instances') make this a practical risk, not a theoretical nicety. The empirical check in Sec. 4.1 (ρHD, Fig. 10) reports kNN overlap of final embeddings, which does not imply equal articulation structure. This gap is load-bearing: if sparsification manufactures LAPs, every 'revealed membership' in Figs. 1, 6–8 may be an artifact. The authors should either prove a local-connectivity transfer lemma or experimentally validate LAP stability across sparsification replicates and against the original","section":"Sec. 3.1, Def. 3.2, Eq. (1)"},{"comment":"The paper concedes that 'no two ambiguous instances can exist within the same neighborhood.' This rules out pairs or chains of adjacent borderline points, which are exactly what one expects in developmental single-cell trajectories—the very scenario highlighted in Fig. 8. Consequently, the method cannot detect the most common form of multi-neighborhood membership in continuous processes, and the claim that the approach 'reveals... hidden neighborhood memberships' in such data is unsupported. This limitation is not merely a boundary case; it is a structural restriction of Definition 3.3 and the LAP-LAP deletion rule. The paper should either extend the approach to handle adjacent ambiguities or explicitly narrow its claims to the single-vertex-articulation setting.","section":"Sec. 5, Limitations"},{"comment":"The LAP-LAP rule deletes edges between two LAPs as 'spurious bridges' without formal justification. This presumes that genuine ambiguous vertices are never adjacent, which is the same restrictive assumption noted in Sec. 5. If two vertices form a bridge between neighborhoods (e.g., a differentiation continuum), deleting their connecting edge removes real structure. The heuristic may be reasonable for the examples, but it is not grounded in the formal definition and can erase real multi-neighborhood relationships. Please provide a formal characterization or additional experiments isolating this rule's effect on results.","section":"Sec. 3.2, Decision rules"},{"comment":"The quantitative sparsification analysis uses only five UMAP runs per graph, and the reported ratios are point estimates without confidence intervals or hypothesis tests. For the RNA-seq dataset, ρ2D is low, but this is attributed to UMAP optimization variance; the same observation could indicate that sparsification changes the layout substantially. More importantly, the analysis never measures the quantity that matters for detection—preservation of the set of LAPs or of component counts in r-neighborhoods—so the conclusion that 'sparsification does not affect local neighborhood structures' is not established at the level needed for the pipeline's core step.","section":"Sec. 4.1, Fig. 10"}],"minor_comments":[{"comment":"The sentence 'Note that normax can exceed the graph diameter' appears to be a typo; for any vertex, the farthest distance cannot exceed the graph diameter. It should read 'cannot exceed.'","section":"Sec. 3.1, after Def. 3.1"},{"comment":"In the definition of preservedNN@k, the formatting of the equation is broken (missing fraction bar). Please fix the typesetting.","section":"Sec. 4.1"},{"comment":"The comparison to GhostUMAP2 is brief; since both methods produce multiple points per instance, it would help to clarify—beyond 'instability vs. data ambiguity'—whether the two phenomena are distinguishable in practice and whether GhostUMAP2 could also surface ambiguous instances.","section":"Sec. 2, Related Work"},{"comment":"The claim of 'near-linear time complexity in practice' is supported only by a heuristic argument and the observed sparsity of LAPs. A brief empirical runtime plot or a more precise complexity bound (e.g., O(n + m + Σ_v |Γ(v)|) under a sparsity assumption) would strengthen the contribution.","section":"Sec. 3.4 / Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising but the central identification step is currently under-validated: spectral sparsification has no local guarantee, and the provided empirical check does not test the needed quantity. The LAP-LAP restriction and the 'no two ambiguous instances' limitation also undermine the applicability to trajectory-like data, which the paper itself highlights. These issues are fixable with additional experiments (LAP stability across sparsification runs, comparison to unsprsified LAPs, and a study of the LAP-LAP rule) and by tempering the claims. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper names a real DR artifact (partial neighborhood embedding), gives it a clean graph-theoretic operationalization (local articulation point on the sparsified neighborhood graph), and builds a working splitting pipeline with several genuinely illustrative case studies. The graph theory is mostly correct: Lemma 3.1's monotonicity proof holds, and the Spielman-Srivastava bound is stated right. It also demonstrates a real gap: trustworthiness/continuity cannot flag ambiguous points (Fig 11). Credit where due: this is a competent, honest contribution, and the limitations section is admirably candid.\n\nThe soft spots, in proportion. The biggest one is exactly what the stress-test note says: Eq. (1) is a global quadratic-form guarantee; it does not control whether radius-r connectivity and local articulation points survive sparsification. The paper's own parameter choices make this concrete — ε=0.7–0.9, and the text says a sparser graph 'allows us to identify more ambiguous instances.' That means the detector's output is deliberately sensitive to a preprocessing step with no local-connectivity guarantee. The empirical check (ρ_HD, Fig 10) is kNN overlap, not articulation structure, and with five UMAP runs it is too weak to certify that LAPs are stable. So the identified 'ambiguous instances' could be sparsification artifacts in some cases. That is not a fatal flaw for a visualization tool, but it should be tested explicitly on the original graph.\n\nSecond, membership revelation is partly by construction: Def. 3.3 attaches each copy to one specific remaining component, so the mapping phase places it there mechanically. The examples are reassuring, but they are hand-picked and few. Third, the abstract claims 'quantitative support' for reducing PNE, yet Sec. 4.1 measures neighborhood preservation of sparsification, not PNE reduction — the paper itself concedes no current metric can verify split projections. That overstates the evidence. Fourth, parameters (ε, τw, r) are tuned per dataset. Fifth, no code or data supplement. Finally, the concession in Sec. 5 that no two ambiguous instances can exist in the same neighborhood is a structural blind spot, especially for the single-cell trajectories highlighted in Fig. 8.\n\nWho this is for: visualization researchers working on DR quality and XAI audits; they will get a useful framing and a plausible first tool. It deserves a serious referee, but the right verdict after review is major revision. The sparsification-to-LAP transfer needs either a local-preservation theorem or an explicit stability check, and the PNE-reduction claim needs a metric or careful rewording.","headline":"A competent, honest visualization paper with a real artifact and a sound graph-theoretic core, but the load-bearing sparsification-to-LAP step lacks a local guarantee and the 'revelation' is partly by construction.","tokens_in":19204,"tokens_out":2114,"would_cite":true,"duration_ms":22809,"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 claims that some data points genuinely belong to multiple dissimilar neighborhoods, that standard dimensionality reduction shows each in only one, and that splitting those points into copies—one per neighborhood—reveals membershi","keywords":["Dimensionality reduction","Ambiguous instances","Partial neighborhood embedding","Local articulation points","Graph sparsification","Vertex splitting","Neighborhood preservation","UMAP"],"falsifier":"Take two well-separated clusters in a high-dimensional space and connect them by two adjacent points, each lying between the clusters and similar to both. For any radius r that includes both clusters, removing either bridging point alone leaves the local subgraph connected, so the method reports no ambiguous instances; if ground-truth labels say both points belong to both clusters, partial neighborhood embedding persists despite the method.","tokens_in":17953,"feed_emoji":"🔀","tokens_out":6026,"duration_ms":67768,"temperature":0.7,"pith_summary":"The paper identifies a specific, previously unnamed source of distortion in dimensionality reduction: ambiguous instances, points that are highly similar to several mutually dissimilar high-dimensional neighborhoods. Because a standard projection assigns each instance exactly one coordinate, such a point lands in one neighborhood (or none), so the projection silently drops its other memberships—an artifact the authors call partial neighborhood embedding. The paper proposes to detect such instances in the graph that a local DR method builds during its relationship phase, then to split each ambiguous vertex into one copy per true neighborhood before the mapping phase. On image, text, and single-cell examples, the approach places copies of the same instance in each of its neighborhoods, exposing memberships that standard projections hide; existing quality metrics such as trustworthiness and continuity do not flag these cases. If the claim holds, instance-level visual analysis—misclassification audits, rare subpopulation discovery, trajectory interpretation—gains a principled way to see multi-membership structure that is real under the chosen metric.","feed_headline":"Split ambiguous points to reveal every neighborhood they belong to","feed_subtitle":"A graph-based method detects instances with multiple dissimilar neighborhoods and places a copy in each.","key_machinery":"The load-bearing object is the r-step local articulation point (LAP): a vertex whose removal disconnects its own r-hop induced subgraph. After effective-resistance spectral sparsification strips away incidental edges, each LAP is treated as ambiguous and split into as many copies as the number of connected components left behind, subject to two filters: edges between two LAPs are dropped, and a component is retained only if its aggregate edge weight reaches a fraction τw of the strongest component. The resulting graph is then embedded by the same local-DR mapping phase, producing one visual copy per true neighborhood.","core_discovery":"The central claim is that ambiguity is a structural property of the data-plus-metric, not a layout artifact, and it can be read off the DR relationship graph. After spectral sparsification removes incidental edges, an instance is called ambiguous at radius r when it is a local articulation point: deleting it increases the number of connected components inside its r-step neighborhood. Each such vertex is replaced by as many copies as there are remaining components, with two safeguards—edges linking two ambiguous vertices are removed, and a component is retained only if its total edge weight reaches a user-set fraction τw of the strongest component. The disambiguated graph is then embedded wit","pith_inferences":["Editorial inference: because the definition is metric- and representation-dependent, the same dataset can split under one embedding and not another; that makes the method a diagnostic for whether the chosen representation supports multi-membership, with applications in representation auditing.","Editorial inference: the single-vertex LAP criterion cannot see bridges formed by two adjacent borderline points, so in continuous developmental trajectories the most graded transitions are likely the ones the method misses; a trajectory- or pair-aware detector would be a natural next step.","Editorial inference: once instances are duplicated, downstream analyses that count points or run clustering on the disambiguated graph need a way to treat copies as soft memberships rather than independent observations; a testable extension is to measure whether soft-membership downstream models improve over hard assignments.","Editorial inference: a synthetic benchmark with known ground-truth multi-membership labels—for instance, Gaussian clusters connected by one point versus by a chain—would let users calibrate r and τw and would turn the qualitative examples into an operating-characteristic curve."],"forward_implications":["Analysts auditing a single misclassification can now see both sides: the SVHN '1' misclassified as '7' is placed among sevens and among ones, so the classifier's mistake looks less confident and more explainable.","In single-cell visualizations, cells with split copies mark potential transitional or rare intermediate states that a fixed-position projection would have hidden behind one cluster.","Because the pipeline operates on the weighted graph produced in the relationship phase, the same detection-and-splitting step carries over to other local graph-based DR techniques, not just the one used for the examples.","Trustworthiness and continuity scores are shown to be blind to ambiguity, which motivates a new class of DR quality metrics that accept multiple points per instance.","The number of splits as a function of neighborhood size and sparsification is informative: if splits appear only under very aggressive sparsification, the paper suggests ambiguity may not genuinely exist in the data."],"fun_headline_variants":["Replicate ambiguous points to preserve all neighborhood ties","Split local articulation points to fix projection artifacts","Ambiguous instances are copied into each relevant neighborhood","Graph-based disambiguation reduces partial embedding errors","Reveal hidden memberships by splitting ambiguous data points"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That an ambiguous instance is a single vertex whose removal isolates its true neighborhoods: if two or more borderline points lie between the same neighborhoods, the detector will not fire, so the method misses exactly the graded transitions it highlights.","fun_headline_variants_meta":{"raw":{"variants":["Replicate ambiguous points to preserve all neighborhood ties","Split local articulation points to fix projection artifacts","Ambiguous instances are copied into each relevant neighborhood","Graph-based disambiguation reduces partial embedding errors","Reveal hidden memberships by splitting ambiguous data points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1103,"prompt_tokens":740,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":302}},"tokens_in":484,"tokens_out":363,"duration_ms":5063,"temperature":1.0,"reasoning_tokens":302,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:21:23.153805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two well-separated clusters in a high-dimensional space and connect them by two adjacent points, each lying between the clusters and similar to both. For any radius r that includes both clusters, removing either bridging point alone leaves the local subgraph connected, so the method reports no ambiguous instances; if ground-truth labels say both points belong to both clusters, partial neighborhood embedding persists despite the method.","supporting_citations":[],"review_version":2}