{"id":"5a1d04f8-7f8b-467a-ad92-cd4cab143bbe","arxiv_id":"2608.11269","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CosMAP combines cosine-similarity neighborhoods, a temperature-scaled affinity graph, and a two-phase embedding refinement to produce low-dimensional visualizations that the authors find clearer than UMAP, t-SNE, LocalMAP, and related methods on benchmark and omics data.","lead":"CosMAP is a new dimensionality reduction method for sparse high-dimensional data, built by replacing UMAP's Euclidean neighborhood graph with cosine-based, temperature-smoothed affinities and adding a two-stage refinement that rebuilds the graph from an intermediate embedding. The authors report cleaner cluster separation on handwriting, single-cell, and Quebec kinship datasets, making it a candidate tool for exploratory omics visualization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of improved neighbourhood preservation is asserted but never directly measured; the only quantitative evidence is NMI on k-means clusters, which rewards separation and can hide neighbourhood distortion, so the load-bearing empirical premise is untested.","rationale":"The reader's weakest_assumption focuses on the cosine k-NN graph with fixed tau=0.5 plus intermediate refinement being more faithful than UMAP's adaptive-bandwidth Euclidean graph. My concern is closely related but more targeted: the paper never actually measures faithfulness or neighbourhood preservation in the experiments, so the premise is not tested at all. The reader noted that 'the abstract overstates neighborhood-preservation gains that are not measured directly,' which is the same gap I identify. I partially agree because the reader also emphasized the two-phase refinement amplifying noise on Cortex, which I incorporate into the concern. However, I locate the load-bearing weakness in the evaluation design rather than in the graph-construction premise per se. There is no internal inconsistency in the method's mathematics; Eq. (7) is a standard BCE loss with a UMAP-style kernel, and the public code supports reproducibility. The concern is about the empirical support for the central claim. A direct neighbourhood-preservation benchmark would settle it. Therefore the verdict stays CONDITIONAL: the method is plausible and reproducible, but the headline claim should not be accepted as stated without the missing measurement.","tokens_in":25467,"tokens_out":1626,"duration_ms":14357,"concrete_test":"Recompute all real-data comparisons (MNIST, USPS, Retina, Cortex, Heart, PBMC, kinship) using direct neighbourhood-preservation metrics on the 2D embeddings: trustworthiness and continuity (e.g., K=50 and K=100) between the original feature space (or kinship matrix) and the embedding, using identical preprocessing and default parameters for all methods. If CosMAP does not rank first or statistically tied for first on these metrics, the abstract's neighbourhood-preservation claim should be removed or weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's central claim is that CosMAP 'improves neighbourhood preservation' and provides 'more coherent visual representations.' However, the paper reports no direct neighbourhood-preservation metric on any real dataset. The only quantitative evidence is NMI between ground-truth labels and k-means clusters of the 2D embedding (Eq. 14, Figures 7, 9, 12). NMI measures cluster-label agreement, not neighbourhood preservation; a method that aggressively separates classes can score high NMI while destroying local neighbourhood geometry (e.g., by fragmenting a class or over-separating). The Cortex result in Section 4.3 is direct evidence of this failure mode: CosMAP is visually competitive but PaCMAP has the highest NMI, and the authors note CosMAP splits the endothelial-mural population into two subclusters (over-refinement). Thus the headline claim 'improves neighbourhood preservation' is unsupported by the reported evidence, and the two-phase refinement (Definition 1, Algorithm 3) that drives the claimed improvement can amplify spurious separation exactly in the regime where neighbourhood preservation would fail. Additionally, the genealogy evaluation (Section 4.4) explicitly uses a per-method tuning and inspection protocol with Euclidean distance fixed across methods, so its NMI advantage is not a clean default-settings comparison and does not establish preservation of kinship geometry. The load-bearing premise — that cosine k-NN graphs with fixed temperature tau=0.5 plus intermediate refinement give more faithful neighbourhoods than UMAP's adaptive-bandwidth Euclidean graph — is therefore asserted but never directly tested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CosMAP is a graph-based dimensionality-reduction method that extends UMAP by building k-NN graphs with cosine similarity, defining a temperature-scaled local affinity distribution (Eq. 4), and optimizing a binary cross-entropy objective with negative sampling (Eqs. 7–9, Algorithm 1). Its main methodological novelty is a two-phase refinement procedure (Definition 1 and Algorithm 3): a r-dimensional intermediate embedding is first learned, then a new k-NN graph is reconstructed from that embedding, and the final 2D embedding is optimized from this refined graph. The authors evaluate CosMAP on MNIST and USPS, two scRNA-seq datasets (mouse retina and cortex), and a BALSAC–CARTaGENE kinship matrix, comparing against t-SNE, UMAP, PaCMAP, LocalMAP, and other methods. The abstract claims that CosMAP 'produces more coherent visual representations, improves neighbourhood preservation, and provides clearer global organization' of the data.","tokens_in":25797,"tokens_out":4523,"duration_ms":42124,"significance":"If substantiated, CosMAP would provide a practical tool for exploratory visualization of sparse high-dimensional omics and genealogical data, and its two-phase graph-refinement idea could inform future neighbor-embedding designs. The paper has clear strengths: the method is described in algorithmic detail, the implementation is publicly available, scRNA-seq preprocessing is standardized, and a metric-oriented ablation is included. However, the central empirical claim of improved neighbourhood preservation is not directly measured anywhere in the manuscript; the only quantitative evidence is NMI between k-means clusters and annotated labels, which is not a neighbourhood-preservation metric. In addition, the two-phase refinement—the paper's main novelty—is not ablated on real datasets. These gaps currently leave the headline claims under-supported.","major_comments":[{"comment":"The abstract's claim that CosMAP 'improves neighbourhood preservation' is not supported by any direct neighbourhood-preservation metric. The only quantitative evaluation reported is NMI between ground-truth labels and k-means clusters of the 2D embeddings (Eq. 14). NMI measures cluster-label agreement, not local neighbourhood fidelity; a method that aggressively separates classes or fragments a coherent population can achieve high NMI while destroying local geometry. The Cortex result in Section 4.3 is a concrete example: CosMAP splits the endothelial-mural population into two subclusters, and PaCMAP obtains the highest NMI. To support the stated claim, the authors should report neighbourhood-preservation metrics (e.g., trustworthiness, continuity, co-k-nearest-neighbour agreement, or mean relative rank error) on at least MNIST, Retina, and Cortex, or temper the abstract's wording to reflect the evidence actually provided.","section":"Section 4, Eqs. (13)–(14)"},{"comment":"The per-method tuning protocol for the kinship dataset weakens the quantitative comparison. The text states that 'a per-method hyperparameter search' was performed and that Euclidean distance was fixed as the working metric because it is supported by all methods. As a result, the reported NMI advantage of CosMAP (Fig. 12) reflects tuned configurations, not a clean comparison of default settings. Furthermore, because the input is a pairwise kinship matrix, it is not specified whether the methods treat each individual as a feature vector (rows of the matrix) or use a precomputed dissimilarity structure; the construction of the k-NN graph for this dataset must be clarified. Without a controlled protocol (e.g., identical hyperparameter budgets or a sensitivity analysis over k), the claim that CosMAP provides 'clearer global organization' of genealogical patterns is not established.","section":"Section 4.4, Genealogical dataset"},{"comment":"The two-phase refinement is the central methodological contribution, but its benefit is not validated on real data. The paper itself states in Section 5 that the full refinement 'does not always provide a clear advantage' and can over-separate small datasets, as observed on Cortex in Section 4.3. The only quantitative ablation in Appendix 7.1 uses a synthetic make_blobs dataset and varies the metric and temperature; it does not compare the one-phase and two-phase versions. A head-to-head comparison of CosMAP with and without the second phase on MNIST, Retina, and Cortex would directly test whether reconstructing the k-NN graph from the intermediate embedding actually improves local and global structure. This is a load-bearing component of the contribution and should be addressed.","section":"Section 5 and Algorithm 3"}],"minor_comments":[{"comment":"There are several typographical issues: 'dimesional' in Section 2, 'V AEs' with unwanted spaces, 'InfoCE-t-SNE' instead of 'InfoNCE-t-SNE', inconsistent 'Cartagène'/'CARTaGENE', and 'the our experiments' in Appendix 7.2. These should be corrected in a revision.","section":"Throughout"},{"comment":"The term 'NT-Xent' is potentially misleading because Eq. (4) normalizes over the k-neighbourhood only, whereas the original NT-Xent loss normalizes over all samples in a batch. Suggest calling this a temperature-scaled local softmax to avoid conflating the two.","section":"Section 3.1, Eq. (4)"},{"comment":"For MNIST the text states that no normalization or standardization was applied, but the preprocessing for USPS is not described. Please state explicitly whether the same raw-feature treatment was used for USPS.","section":"Section 4.1"},{"comment":"The phrase 'the single choice supported by all methods' is ambiguous: should be clarified as Euclidean distance computed on rows of the kinship matrix, or on a transformed dissimilarity, so that readers understand what input each method actually receives.","section":"Section 4.4"},{"comment":"The statement that 'BC3A/BC3B-related structure becomes partially visible' is vague; it would benefit from a quantitative measure or a more precise description of which cells are separated and under what conditions.","section":"Section 4.2, Retina discussion"},{"comment":"The Heart Cell Atlas, PBMC, and COIL-20 results appear only in an appendix and are not mentioned in the abstract or conclusion. Either integrate them into the main evaluation or explicitly state why they are excluded from the summary claims.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper describes a plausible and well-documented method, and the public code is a clear asset. The main concern is that the headline claims outrun the evidence: improved neighbourhood preservation is never measured directly, and the key two-phase refinement is not ablated on real data. The genealogy comparison also needs a clearer and more controlled protocol. These issues are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Basil, here's my take on CosMAP. The paper is worth reading: it's a clean, well-documented addition to the neighbor-embedding family, and the authors do most things right. The method itself combines cosine kNN graphs with temperature-scaled softmax affinities, a UMAP-style kernel, and a two-phase refinement step; the equations are consistent, the code is public, and the ablation and sensitivity analysis are more careful than we usually see in this area. The genealogy application is a nice touch.\n\nThe soft spot is the gap between the abstract's claim — 'improves neighbourhood preservation' — and what's actually measured. The only quantitative metric is NMI between ground-truth labels and k-means clusters on the 2D embedding. NMI rewards clean separation and can hide local distortion. The stress-test note is right: they never measure trustworthiness, continuity, or anything that directly tests whether local neighborhoods are preserved. The Cortex example makes this concrete: CosMAP splits endothelial-mural into two subclusters, and PaCMAP has the highest NMI. So the load-bearing premise of the paper — that the cosine graph plus refinement gives more faithful neighborhoods — remains an assertion, not a demonstrated result.\n\nThe kinship evaluation is also weaker than it looks, since they tune each method by inspection and fix Euclidean distance everywhere; that's transparent, but it doesn't give a clean default-settings comparison. The best method varies across datasets, which is fine, but it cuts against the broad claim.\n\nThese are fixable. Adding a couple of standard neighborhood-preservation metrics would directly address the main concern; redoing the genealogy comparison with a consistent protocol, or at least reporting sensitivity to the tuning, would help. The paper is already useful as a well-documented implementation and a careful visual comparison. It deserves a serious referee, so I'd send it to review, but with the expectation that the authors tighten the evaluation before publication.","headline":"A solid, well-documented new DR method, but the headline claim of improved neighborhood preservation is asserted, not directly measured.","tokens_in":26359,"tokens_out":2780,"would_cite":true,"duration_ms":25306,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that CosMAP, a cosine-graph dimensionality-reduction method with a two-phase refinement, produces more coherent and neighborhood-faithful embeddings than existing methods on omics and genealogical data.","keywords":["dimensionality reduction","contrastive learning","cosine similarity","neighbor embedding","single-cell RNA sequencing","genealogical kinship","two-phase refinement","UMAP extension"],"falsifier":"Compute a dataset with known ground-truth neighborhoods (for example, points on a low-dimensional manifold with added sparse noise), count the precision and recall of edges in the cosine k-NN graph and in the rebuilt graph from the 30-D intermediate embedding against the true edges, and compare downstream NMI after clustering the 2-D embeddings. If the rebuilt graph has no higher edge fidelity than the raw cosine graph, or if CosMAP's NMI advantage disappears when UMAP's Euclidean graph is given matched preprocessing, the central claim fails.","tokens_in":25266,"feed_emoji":"🧬","tokens_out":5477,"duration_ms":47211,"temperature":0.7,"pith_summary":"CosMAP is a new graph-based dimensionality-reduction method for sparse, high-dimensional data such as single-cell RNA sequencing and genealogical kinship matrices. The paper argues that building the neighborhood graph from cosine similarities with a temperature-normalized contrastive affinity, then rebuilding the graph from an intermediate 30-dimensional embedding before the final 2-D projection, yields embeddings that are more coherent and better preserve local neighborhoods than current methods. On handwritten digits, mouse retina and cortex cell populations, and a Quebec kinship matrix, CosMAP is reported to separate known groups more cleanly while keeping global structure readable. The intended payoff is a practical visualization tool for exploratory analysis when labels are absent or unreliable.","feed_headline":"Two-stage cosine graph rebuild sharpens 2-D cluster maps","feed_subtitle":"CosMAP denoises its neighborhood graph in an intermediate space before the final projection, separating digit, cell, and kinship groups…","key_machinery":"The central object is a temperature-normalized cosine affinity graph combined with a two-phase refinement pipeline. For each point, CosMAP computes a conditional distribution over its k nearest neighbors, P_{j|i}=exp(sim(x_i,x_j)/τ) normalized over N_k(i), symmetrized to p_ij, and then matched in the embedding by a heavy-tailed kernel q_ij=1/(1+a||y_i-y_j||^{2b}) through a binary cross-entropy loss with negative sampling. Definition 1 formalizes the two-phase refinement: an intermediate 30-D CosMAP embedding is used to reconstruct a more reliable graph, and a coordinate projection initializes the final 2-D optimization. This mechanism is what carries the claim that noisy high-dimensional edges can be corrected before final visualization.","core_discovery":"The paper's central claim is that CosMAP produces more faithful and interpretable low-dimensional embeddings than existing neighbor-embedding methods for omics and genealogical data. The method uses a k-nearest-neighbor graph under cosine similarity, converts edge strengths into a temperature-scaled conditional distribution with τ=0.5, symmetrizes it, and optimizes a binary cross-entropy attraction–repulsion objective with negative sampling. The distinctive step is a two-phase refinement: it first learns an intermediate r-dimensional embedding with r=30 by default, rebuilds the k-NN graph from that embedding, and only then optimizes the final 2-D projection, initialized by projecting the first two coordinates of the intermediate embedding. The reported results place CosMAP ahead of UMAP, t-SNE, PaCMAP, LocalMAP, PHATE, TriMAP, NCVis, h-NNE, and contrastive t-SNE variants on cluster coherence and neighborhood preservation, with LocalMAP as the closest competitor.","pith_inferences":["A natural test the paper does not run is to feed the rebuilt intermediate graph back into UMAP's own optimizer; if CosMAP's gains come mainly from graph refinement rather than the cosine affinities, UMAP would match CosMAP with the same two-phase graph.","The temperature τ=0.5 is fixed rather than learned; an adaptive temperature per point might remove the remaining over-splitting seen on Cortex, since small datasets with subtle boundaries appear to need a sharper or smoother distribution.","The same two-phase recipe could be applied to other input types, such as spatial transcriptomics or scATAC-seq, where the initial feature space is even sparser and cosine affinities may behave differently.","Because the kinship matrix itself is the input rather than a feature matrix, CosMAP effectively treats genealogical relatedness as a manifold; if useful, this suggests extending the method to arbitrary precomputed similarity matrices with explicit handling of non-metric affinities."],"forward_implications":["If CosMAP's graph construction is more faithful, exploratory single-cell analyses should see fewer spurious cell clusters and cleaner separation of rare populations without any labeled training data.","The two-phase refinement implies a general recipe: any neighbor-embedding method can be improved by first projecting to a moderate dimension, rebuilding its graph, and then projecting to 2-D, which is what the paper's ablation-style comparison suggests for CosMAP.","For genealogical kinship matrices, CosMAP's embedding recovers regional founder-effect structure directly from pairwise relatedness, so the method could replace hand-tuned visualization pipelines in population genetics.","Because CosMAP uses negative sampling and batching, it scales to data sizes used in current single-cell studies, making the claimed improvements available on datasets of tens of thousands of cells.","The observed over-splitting on the small Cortex dataset implies that the refinement should be used selectively; the paper itself flags that one-phase optimization can be more coherent for small datasets."],"supporting_citations":[{"why":"Supplies the graph-based neighbor-embedding framework and the heavy-tailed low-dimensional kernel that CosMAP extends.","marker":"[35]"},{"why":"Provides the critique of unreliable fixed k-NN graphs that motivates CosMAP's two-phase refinement.","marker":"[10]"},{"why":"Supplies the temperature-scaled cosine contrastive normalization (NT-Xent) adapted into Eq. (4).","marker":"[8]"},{"why":"Supplies the negative-sampling strategy used to approximate the repulsive term.","marker":"[18]"},{"why":"Shows t-SNE, UMAP, and LargeVis losses as contrastive objectives, placing CosMAP in that spectrum and providing the InfoNCE-t-SNE and Neg-t-SNE baselines.","marker":"[36]"},{"why":"Defines the crowding problem and heavy-tailed t-distribution used for low-dimensional similarities.","marker":"[32]"},{"why":"Reports that LocalMAP over-fragments small neighborhoods, supporting the paper's caution about aggressive graph adjustment.","marker":"[12]"},{"why":"Provides evidence that the choice of proximity metric (cosine vs Euclidean) changes neighborhood structure in single-cell data.","marker":"[13]"}],"fun_headline_variants":["CosMAP sharpens 2-D maps by rebuilding graphs in a higher space","Two-stage embedding sharpens cluster separation in omics and kinship","CosMAP: contrastive affinities and two-phase refinement improve embeddings","Better cluster maps via cosine graphs and a two-stage optimization","CosMAP re-learns neighborhoods from an intermediate embedding for 2-D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole advantage rests on the premise that a k-nearest-neighbor graph built from cosine similarities with a fixed temperature τ=0.5, and then rebuilt from a 30-dimensional intermediate embedding, is more faithful to true structure than the adaptive-bandwidth Euclidean graph used by UMAP.","fun_headline_variants_meta":{"raw":{"variants":["CosMAP sharpens 2-D maps by rebuilding graphs in a higher space","Two-stage embedding sharpens cluster separation in omics and kinship","CosMAP: contrastive affinities and two-phase refinement improve embeddings","Better cluster maps via cosine graphs and a two-stage optimization","CosMAP re-learns neighborhoods from an intermediate embedding for 2-D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000562,"raw_usage":{"total_tokens":2709,"prompt_tokens":1027,"completion_tokens":1682,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1591}},"tokens_in":643,"tokens_out":1682,"duration_ms":11981,"temperature":1.0,"reasoning_tokens":1591,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:21:54.749193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute a dataset with known ground-truth neighborhoods (for example, points on a low-dimensional manifold with added sparse noise), count the precision and recall of edges in the cosine k-NN graph and in the rebuilt graph from the 30-D intermediate embedding against the true edges, and compare downstream NMI after clustering the 2-D embeddings. If the rebuilt graph has no higher edge fidelity than the raw cosine graph, or if CosMAP's NMI advantage disappears when UMAP's Euclidean graph is given matched preprocessing, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides evidence that the choice of proximity metric (cosine vs Euclidean) changes neighborhood structure in single-cell data."}],"review_version":1}