{"id":"7a66e6b3-2daa-4328-9ebe-5b3dc9e9791b","arxiv_id":"2511.09486","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Using ABIDE's local neighborhood sizes and intrinsic dimension as plug-in hyperparameters improves LLE, spectral clustering, and UMAP embeddings on benchmark datasets.","lead":"This paper plugs an intrinsic-dimension estimator (ABIDE) into LLE, spectral clustering, and UMAP, using its per-point neighborhood sizes and dimension estimate as automatic hyperparameters. Experiments suggest gains over default settings on four datasets, though some comparisons are confounded by different output dimensions and missing error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ABIDE's uniformity-based k* is not shown to be optimal for the downstream DR objective; the 'optimal tuning' claim rests on an untested equivalence.","rationale":"The reader's weakest assumption identified precisely that ABIDE's uniformity criterion may not align with the downstream DR objective. I agree: this is the most load-bearing gap. The paper's central claim is normative ('optimal hyper-parameter tuning'), but the method is built on a statistical criterion—local Poisson-process homogeneity—that has no proven or tested connection to linear reconstruction error, spectral embedding quality, or UMAP's cross-entropy objective. The experiments give some evidence that the adaptive method beats default configurations, and the simulated-manifold results are encouraging, but they do not close the gap because the baselines are not oracle-tuned and the output dimensionality is often not matched. This is not a fatal flaw; it is a missing justification for the word 'optimal.' The appropriate verdict remains CONDITIONAL: the authors could either prove a formal connection (e.g., under which curvature and sampling conditions k* minimizes LLE reconstruction error) or provide a comprehensive benchmark against oracle-tuned baselines. No code was released, which makes the empirical claims harder to verify, but that is secondary to the conceptual gap. I do not see an internal inconsistency; the concern is about the strength of the central claim relative to the evidence.","tokens_in":12134,"tokens_out":3634,"duration_ms":43544,"concrete_test":"Construct a synthetic dataset on a manifold with roughly constant density but strong curvature variation (e.g., a helix with changing radius or a sinusoidally bent sheet). Compute ABIDE's k*_i from Eq. (3) and, for each point, the k that minimizes the LLE reconstruction error in Eq. (7) (or that maximizes embedding recovery against ground-truth coordinates). Compare the two k distributions. If they diverge systematically, uniformity-optimal neighborhoods are not DR-optimal, directly undermining the Abstract's 'optimal hyper-parameter tuning' claim. Additionally, rerun Tables 2-4 with all baselines using the same output dimensionality d* and a grid-selected k, to check whether the adaptive method beats the best fixed-k baseline rather than only the sklearn/uwot defaults.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that ABIDE's k*_i and d* provide 'optimal hyper-parameter tuning' for NN-based DR (Abstract). But k*_i is defined solely by a likelihood-ratio test for local homogeneity of the Poisson process intensity (Section 2.1, Eq. 3), i.e., k* is optimal for density uniformity, not for the objective of the downstream DR algorithm. For LLE, the reconstruction error in Eq. (7) depends on how well each point is linearly reconstructible from its neighbors, which depends on manifold curvature, not on sampling uniformity. A neighborhood that is uniform in density can still be far from linear if the manifold bends sharply, so using k* may worsen LLE*, SC*, or UMAP* relative to a better-tuned fixed k. The paper never proves or systematically tests that uniform-density neighborhoods coincide with neighborhoods that minimize reconstruction error or maximize clustering/visualization quality. The experiments in Tables 2-4 compare against default hyperparameters, not against the oracle best fixed-k configuration with matched output dimensionality. Figure 3 partially addresses this by fixing d*, but 'better than defaults' is weaker than 'optimal'. Thus the claim of optimal tuning is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a general framework for adaptive nonparametric dimensionality reduction: it uses ABIDE, an intrinsic-dimension estimator that also returns a per-point optimal number of nearest neighbours k*_i, to set the neighbourhood size and projection dimensionality of NN-based DR algorithms. The authors instantiate this for LLE (LLE*), spectral clustering (SC*), and UMAP (UMAP*), and evaluate the resulting embeddings through k-means clustering metrics on MNIST, News Articles, synthetic manifolds, and Iris, plus a supervised logistic-regression experiment on the embedded representations. The central claim is that ABIDE's k* and d* provide an 'optimal hyper-parameter tuning' for any DR method that relies on local neighbourhoods, and that the adaptive versions outperform default implementations on the tested benchmarks.","tokens_in":12410,"tokens_out":4923,"duration_ms":61448,"significance":"If the claim were fully established, the framework would be a practically useful replacement for manual tuning of n_neighbours and n_components in several common DR methods, and the connection between intrinsic-dimension estimation and neighbourhood selection is conceptually appealing. The paper rightly capitalises on ABIDE's consistency result from prior work, and it reports a broad set of experiments across four datasets and three downstream tasks, with computational overhead quantified in Table 1. The proof-of-concept extensions to SC and UMAP are useful. However, the evidence as presented does not support the word 'optimal': k*_i is selected for local Poisson-intensity homogeneity, not for any DR objective, and the empirical comparisons do not separate the effect of neighbourhood adaptivity from that of using a higher-dimensional projection. The paper is a reasonable candidate for publication after substantial revision, but the current overclaim and the confounded experimental design are load-bearing.","major_comments":[{"comment":"The abstract and §2.2 claim that the framework provides 'optimal hyper-parameter tuning' for NN-based DR. This is not established. The k*_i in Eq. (3) is the smallest k at which a likelihood-ratio test rejects homogeneity of the Poisson process intensity; it is optimal for a uniformity criterion, not for reconstruction error, clustering separability, or visualisation quality. For LLE, the objective in Eq. (7) depends on how well each point is linearly reconstructible from its neighbours, which is governed by manifold curvature and noise, not by sampling uniformity. Uniform-density neighbourhoods on a strongly curved manifold can be poor for LLE. The paper provides no theorem and no experiment tying the uniformity criterion to the downstream DR objective. Please either weaken the 'optimal' claim to 'adaptive heuristic' throughout, or add a direct comparison between k* and an oracle/strong","section":"§2.1, Eq. (3); §2.2, Eq. (5); §2.3, Eq. (7)"},{"comment":"The main quantitative comparisons confound neighbourhood adaptivity with output dimensionality. In Tables 2–4, the adaptive methods embed into d* components (e.g., 11.56 for MNIST, 10.94 for News Articles in Table 1), while the default baselines use the standard 2D visualisation setting for LLE and UMAP; the SC comparison similarly does not hold the embedding dimension fixed. A large part of the reported improvement may simply be the benefit of using more than 2 components. Figure 3 partially addresses this for LLE by fixing n_components=d* and varying n_neighbours, but it does not cover SC or UMAP, and none of Tables 2–4 or Figure 3 report repeated runs or error bars despite k-means being stochastic. At minimum, add matched-dimension baselines (default method with n_components=d* and n_neighbours equal to the mean or median of k*) and report mean±std over multiple random initialisations","section":"Tables 2–4; §3.2.2, Fig. 3"},{"comment":"The method is presented as avoiding hyper-parameter tuning, but it still contains a free significance level α in the definition of k*_i. The text itself notes that 'small values of α imply the identification of larger neighbourhoods' and calls this a 'choice of the researcher', yet no default α is stated and no sensitivity analysis is reported for any dataset. Since k*_i directly controls the neighbourhood graph, the results may depend non-negligibly on this choice. Please state the value of α used in all experiments and provide a sensitivity analysis (e.g., α ∈ {0.01, 0.05, 0.1}) for at least one dataset, or justify a principled default.","section":"§2.1, Eq. (3); §3.2.2"},{"comment":"The out-of-sample extension used for supervised learning is under-specified. Step 2 states 'Apply ABIDE: set intrinsic dimension to d*' for the augmented training-plus-test dataset, but ABIDE is a joint iterative estimator of d and k; fixing d to the training value changes the neighbourhood-selection criterion relative to the training procedure. The derivation of the test point's k*_test under this fixed-d variant is not given. In addition, Figure 5(B) compares LLE* to a grid of non-adaptive LLE configurations, but the grid uses different n_components values from d* (e.g., MNIST d*≈11.56 while the grid only goes to 12), so the supervised comparison is again not fully dimension-matched. Please clarify the out-of-sample algorithm and, if possible, compare against non-adaptive LLE with n_components=d* and n_neighbours equal to the summary statistics of k*.","section":"§3.3, Algorithm 1"}],"minor_comments":[{"comment":"Typo: 'obatined' should be 'obtained'.","section":"§3.2.4"},{"comment":"Several entries in the News Articles row are run together without spaces ('SC 0.447 0.5570.6940.635'); please fix the numerical formatting.","section":"Table 3"},{"comment":"The GitHub repository is mentioned but no URL or permanent DOI is given; provide one for reproducibility.","section":"§3.1"},{"comment":"The sentence in the last paragraph of §2.3 claims that locally linear reconstructions are 'much more accurate' on uniform-sampling neighbourhoods; as noted in the major comments, this is an assertion rather than a result. Please either prove it under explicit conditions or soften it to a heuristic motivation.","section":"§2.3"},{"comment":"The description of the SC baseline is incomplete: it is not clear what n_components or n_clusters are used for the default scikit-learn spectral clustering, or how these are matched to the adaptive version. Clarify the exact hyperparameter settings.","section":"§3.2.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially an application of the authors' own ABIDE estimator to DR hyperparameter selection. That is legitimate and the ABIDE consistency proof is external support, but the 'optimal' wording and the confounded comparisons are likely to draw criticism in review. I would encourage the editor to ask for the matched-dimension baselines and error bars as a condition of acceptance; with those additions, the manuscript would make a modest but useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead Di Noia et al. on using ABIDE's adaptive neighborhoods and intrinsic dimension estimate to set k and d for LLE, spectral clustering, and UMAP. The core idea is simple and genuinely new: instead of a global n_neighbors, let each point have its own k* chosen by a likelihood-ratio test for local Poisson homogeneity, and use the ABIDE dimension estimate as the target embedding dimension. That is a clean, reproducible plug-in, and the paper earns credit for extending it to three methods and for including an out-of-sample projection scheme.\n\nWhat the paper does well: the framework is clearly stated, the integration with LLE is explicit, and Figure 3 is a step in the right direction — fixing d* and sweeping k shows that LLE* is often at or above the grid-search curve. The supervised extension is thoughtful.\n\nWhere it falls short. First, the headline claim of \"optimal hyper-parameter tuning\" is not supported. k* optimizes a statistical criterion — local uniformity of the Poisson intensity — not the reconstruction error or clustering quality of the downstream method. The stress-test note is on point: a uniform-density neighborhood can still be highly nonlinear if the manifold bends, so there is no reason to expect k* to be optimal for LLE/UMAP objectives. The paper never tests that link directly. It only shows \"better than sklearn defaults,\" which is a low bar.\n\nSecond, the main comparisons in Tables 2–4 confound the neighborhood adaptivity with the output dimension. Adaptive methods use d* (e.g., 11.56 for MNIST) while the defaults project to 2D. For clustering metrics, more dimensions usually help. Figure 3 partially fixes this for LLE, but the headline tables do not.\n\nThird, the experimental reporting is thin: no error bars for stochastic k-means, no stated alpha for the likelihood-ratio test (which the authors themselves call a researcher choice), and no code released. These are fixable, not fatal, but they matter.\n\nThe citation pattern is honest: ABIDE is self-cited but it is the method being plugged in, and the paper cites prior adaptive LLE work without pretending to be the first adaptive neighborhood idea. The central claim is not circular.\n\nBottom line: this is a plausible, useful engineering contribution with a promising evaluation direction, but it currently overclaims optimality and underdelivers controlled experiments. With a matched-dimension baseline, variance reporting, a stated alpha, and a direct test of whether uniformity-optimal neighborhoods are indeed better for reconstruction, it would be a solid paper.\n\nI'd send it to a serious referee — it deserves real scrutiny rather than a desk reject. I wouldn't cite it yet for the optimality claim, but I'd keep an eye on the revision.","headline":"Useful plug-in idea with real promise, but the 'optimal tuning' claim outruns the experiments: the evaluation confounds d* with k* and never shows that uniformity-optimal neighborhoods are reconstruction-optimal.","tokens_in":12879,"tokens_out":1868,"would_cite":false,"duration_ms":20598,"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":"A single data-driven rule for neighborhood sizes can replace manual tuning in nearest-neighbor dimensionality reduction methods.","keywords":["dimensionality reduction","intrinsic dimension estimation","locally adaptive neighborhoods","nearest neighbors","hyperparameter tuning","locally linear embedding","spectral clustering","UMAP"],"falsifier":"Take a dataset with known manifold structure and deliberately non-uniform sampling density. Run LLE* and, for every fixed k in a wide grid, run LLE with n_components = d* and evaluate a downstream clustering metric with ground-truth labels. If any fixed k (or a per-point k chosen by a different criterion, e.g., minimizing local reconstruction error) beats LLE* by a substantial margin on most such datasets, the claim that ABIDE's k* is optimal for dimensionality reduction is falsified.","tokens_in":12025,"feed_emoji":"🎛️","tokens_out":4891,"duration_ms":48979,"temperature":0.7,"pith_summary":"This paper argues that the two most fragile hyperparameters in nearest-neighbor-based dimensionality reduction—the number of neighbors defining each local neighborhood and the dimension of the target space—can be set automatically by a joint estimator of intrinsic dimension. The estimator, ABIDE, assigns each data point its own neighbor count k*, chosen so the sampling distribution in that neighborhood is approximately uniform, and returns the overall intrinsic dimension d*. Feeding these per-point values into LLE, spectral clustering, and UMAP yields embeddings that beat default configurations on clustering metrics on MNIST, news articles, and simulated manifolds, and improves downstream classification. If correct, this gives practitioners a principled substitute for grid search in unsupervised settings where labels are absent.","feed_headline":"Adaptive neighborhoods beat default settings in three DR methods","feed_subtitle":"A likelihood-ratio test sets each point's own neighbor count, removing grid search and lifting clustering accuracy on real and simulated dat","key_machinery":"ABIDE (Adaptive Binomial ID Estimator): a likelihood-based estimator that jointly estimates intrinsic dimension d and per-point optimal neighbor counts k*_i. The key mechanism is a likelihood-ratio test on the spacings between successive nearest-neighbor distances; k*_i is the smallest k at which the test rejects the null that the local Poisson intensity is constant, so the chosen neighborhood is the largest one that still looks uniformly sampled. The paper's framework treats this k*_i as the 'optimal locally adaptive neighborhood structure' and uses it to build the adjacency matrix of any NN-based projection method, with d* fixing the projection dimension.","core_discovery":"The central claim is that optimal hyperparameters for NN-based dimensionality reduction are not a single global pair (n_neighbors, n_components) but a per-point neighborhood structure. For each point x_i, ABIDE grows a ball until a likelihood-ratio test detects a break in the local homogeneity of the underlying Poisson process; the largest k before the break, k*_i, is declared the optimal neighborhood, and the binomial-count estimator at convergence yields d*. The paper shows that plugging k*_1:n and d* into LLE, spectral clustering, and UMAP—replacing fixed k and hand-set dimension—produces lower-dimensional representations that cluster more cleanly and classify more accurately than default","pith_inferences":["The paper defines 'optimal' by statistical uniformity of sampling, not by any direct measure of embedding quality; whether uniform neighborhoods also minimize reconstruction error or downstream task error is an empirical assumption worth testing head-to-head against oracle-tuned k.","The spread of the k* distribution itself could serve as a diagnostic: a wide spread flags strongly non-uniform density, where fixed-k methods are most unreliable and adaptive structure matters most.","A natural stress test is to compare LLE* against LLE with a grid of fixed k values chosen by an oracle (best ARI on test labels) on the same datasets; the paper's grid comparisons suggest LLE* is near or above average, but an oracle comparison would quantify the remaining gap.","The same k*_i could be plugged into density-based clustering (DBSCAN) or graph-based label propagation, turning this from a dimensionality-reduction tuning trick into a general local-bandwidth selection rule."],"forward_implications":["Replacing fixed n_neighbors with per-point k*_i improves LLE, spectral clustering, and UMAP on clustering metrics across MNIST, news articles, and simulated manifolds.","The framework removes the need for grid search or manual hyperparameter choice in unsupervised dimensionality reduction, where no labels exist to validate a choice.","The gains are largest when data density is non-uniform; on roughly uniform data, a fixed neighborhood near the median of the k* distribution performs comparably, but the adaptive version is safer because density is unknown.","An out-of-sample extension projects new points by estimating their own k* via the trained d, enabling supervised use of the adaptive embeddings.","Any NN-based dimensionality reduction algorithm—the paper mentions DBSCAN, label propagation, and k-NN as further candidates—can inherit the same adaptive neighborhood structure."],"fun_headline_variants":["Adaptive neighborhoods beat defaults in DR","Per-point neighbor tuning lifts DR clustering","No grid search: data-driven neighborhoods improve DR","Adaptive neighborhood sizes sharpen LLE and UMAP"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a neighborhood in which the sampling process looks locally uniform is the right neighborhood for faithful dimensionality reduction—if uniform sampling does not coincide with low reconstruction error, the 'optimal' tuning is optimal only for uniformity, not for the embedding.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive neighborhoods beat defaults in DR","Per-point neighbor tuning lifts DR clustering","No grid search: data-driven neighborhoods improve DR","Adaptive neighborhood sizes sharpen LLE and UMAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000495,"raw_usage":{"total_tokens":2237,"prompt_tokens":686,"completion_tokens":1551,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":1495}},"tokens_in":430,"tokens_out":1551,"duration_ms":12596,"temperature":1.0,"reasoning_tokens":1495,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:35:35.251423+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with known manifold structure and deliberately non-uniform sampling density. Run LLE* and, for every fixed k in a wide grid, run LLE with n_components = d* and evaluate a downstream clustering metric with ground-truth labels. If any fixed k (or a per-point k chosen by a different criterion, e.g., minimizing local reconstruction error) beats LLE* by a substantial margin on most such datasets, the claim that ABIDE's k* is optimal for dimensionality reduction is falsified.","supporting_citations":[],"review_version":1}