{"id":"c9a98826-e637-410e-b497-5762bf45730f","arxiv_id":"2505.11780","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"pdsCART is described as a MapReduce parallelization of dsCART, claimed to produce identical trees at lower runtime, but the supporting experiments are not reproducible.","lead":"This preprint reviews pdsCART, a parallel version of the dsCART streaming decision tree that runs on MapReduce. It claims identical trees to dsCART at lower runtime, but offers no code, data, or units for its timing results.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exact-tree-equivalence claim rests on an unstated correctness condition about deferred split evaluation that the paper neither derives nor tests.","rationale":"I agree with the reader's rejection: the equivalence claim is the strongest claim and the least supported. The paper repeatedly hedges in the text (e.g., \"often closely aligns\", \"nearly identical\") while the experiments claim exact identity, and the reported metrics cannot distinguish exact identity from approximate agreement. The central risk is not that the algorithm is wrong about the world, but that the manuscript has not established its own headline result; the paper itself states it is not introducing a new model, and no external artifact is provided. A reproducibility check on tree structure is the minimal experiment that would decide whether the exact-equivalence statement is true or merely an artifact of coarse evaluation. I therefore keep the reader's REJECT verdict unchanged.","tokens_in":6853,"tokens_out":2495,"duration_ms":24951,"concrete_test":"Independently implement dsCART and the pdsCART batch-histogram variant (or obtain the authors' implementation) and rerun the Table 2 configurations—D1, D5, and D* with record counts 1/20/40/60/80, 1/200/400/600/800, and 1/200/400/600/800 respectively, using the same Θ—then compare full tree structures (split feature, split threshold, leaf class distribution) and exact accuracy. If any tree differs between the two methods on any configuration, the Section 4 universal-equivalence claim is false; if all structures match, the concern is settled in the paper's favor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central and only quantitative claim is in Section 4: across all test cases, pdsCART produces \"the exact same decision trees\" as dsCART with \"identical accuracy levels,\" making it a drop-in parallel replacement. For that claim to be true, every split decision made by dsCART from per-instance updates must be reproduced when split evaluation is deferred until a batch boundary and based only on merged histograms. Section 3.1 asserts only that the feature selected from a stream segment \"often closely aligns\" with the full-data choice, and Section 3.2 says that a suitably chosen threshold Θ yields \"nearly identical\" trees. Neither a probabilistic bound nor a structural invariance argument is provided. The tables that follow report accuracy and runtime, not tree structure, so they cannot substantiate exact equality; equal rounded accuracies (e.g., 83.11%) are compatible with different split choices. No code, datasets, or procedure details are included, and the original pdsCART source is not cited, so the \"all test cases\" claim is not independently auditable. The load-bearing assumption is therefore the unstated correctness condition that deferred split evaluation preserves each split feature and split point; the manuscript gives no reason to believe this holds exactly rather than approximately.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents and evaluates pdsCART, a MapReduce-based parallelization of the dsCART streaming decision tree algorithm. The proposed method maintains per-mapper local histograms of feature/class counts, merges them in the reduce phase, and uses the merged histograms to evaluate candidate splits only at batch boundaries, controlled by a threshold parameter Θ. The paper reports experiments on six datasets (five synthetic, one real-world), summarized in three tables, and claims in Section 4 that pdsCART produces the exact same decision trees and identical accuracy as dsCART across all test cases, so the evaluation focuses on execution time rather than predictive performance.","tokens_in":7055,"tokens_out":6625,"duration_ms":63166,"significance":"If the central equivalence claim were substantiated, pdsCART would be a useful engineering result: a drop-in parallel replacement for dsCART with identical model behavior and lower runtime. The paper's scope is honestly stated (adaptation, not a new model), and the runtime trends in Tables 3 and 4 are plausible and worth reporting. However, the main claim is load-bearing and is not supported: no correctness proof for deferred split evaluation is given, no tree-structure comparison is reported, no reproducibility artifacts are provided, and the threshold Θ is tuned to match dsCART. The manuscript therefore does not currently establish its key contribution, and its empirical part cannot be independently audited.","major_comments":[{"comment":"The central claim of the paper, that pdsCART and dsCART generate \"the exact same decision trees\" with \"identical accuracy levels\" in all test cases, is not evidenced by the experimental section. No tree-structure comparison (split feature, split point, tree depth, node counts, or any tree-equality metric) is reported; Tables 2-4 report accuracy and runtime only, and constant two-decimal accuracies such as 83.11% are compatible with different split choices. The claim is asserted, not demonstrated, and it is precisely the load-bearing result that lets the paper direct its evaluation to runtime.","section":"Section 4, first paragraph and Tables 2-4"},{"comment":"The exact-equivalence claim requires that every split decision made by dsCART from per-instance updates is exactly reproduced when split evaluation is deferred to batch boundaries and based only on merged histograms. Section 3.1 states only that the feature selected from a stream segment \"often closely aligns\" with the full-data choice, and Section 3.2 states that an appropriate threshold Θ yields \"nearly identical\" trees. Neither a probabilistic bound nor a structural invariance argument is provided, so the correctness condition underlying the paper's central claim is unstated and untested.","section":"Sections 3.1-3.2"},{"comment":"There is a potential algorithmic mismatch between the baseline and the parallel method. dsCART is described in Section 2.1.2 as using Gaussian estimation to select splitting features, whereas pdsCART is described in Sections 3.1-3.2 as using histograms and the Gini index to compute split improvements. The paper never explains how the histogram/Gini-based computations reproduce the Gaussian-based split decisions of dsCART; this gap must be closed before the identical-tree claim can be evaluated.","section":"Sections 2.1.2, 3.1-3.2"},{"comment":"The quantitative illustration for dataset D4 is internally inconsistent. Processing 4 million records in batches of 22 yields about 181,818 batches, not \"around 20,000 computations,\" while batches of 800 yield 5,000 batches, which matches the quoted figure only in the second case. This inconsistency affects the only concrete illustration of how deferred evaluation reduces the number of split computations and weakens the paper's efficiency argument.","section":"Sections 3.2 and 4 (D4 example)"},{"comment":"The threshold parameter Θ and the phrase \"with appropriate tuning of the parameters\" indicate that the claimed equivalence is partly enforced by construction: the algorithm's split behavior is tuned so that the produced trees match dsCART. The manuscript gives no independent calibration rule for Θ, no sensitivity analysis over it, and no definition of \"appropriate tuning,\" so the reported \"identical trees\" observation cannot be distinguished from an artifact of parameter choice.","section":"Sections 3.2 and 4"}],"minor_comments":[{"comment":"The table is difficult to parse because the first column of each dataset block is not labeled; it appears to denote the number of records used per split estimation, but the same symbol \"1\" is used with different implied scales across datasets, and the intended units are not stated.","section":"Table 2"},{"comment":"The units of the reported \"Time\" values are never specified, so the runtime claims cannot be interpreted quantitatively.","section":"Tables 2-4"},{"comment":"The table introduces datasets D2a, D5a, D10a, and D20a that are not defined in Table 1, and the row labels vary tree depth and attribute count simultaneously, confounding the interpretation of the runtime difference.","section":"Table 3"},{"comment":"The sentence stating that prior methods were not tailored to \"directional decision trees like GBRT\" is confusing because GBRT is not a directional decision tree; please rephrase or remove the comparison.","section":"Section 2.2"},{"comment":"The claim of constructing a decision tree in a \"single pass over the data\" is not qualified against the round-based MapReduce procedure in Section 3.2, where mappers repeatedly traverse the current tree and the controller reads output files between rounds.","section":"Conclusion, Section 5"},{"comment":"Reference [18] is cited for the interpretability of decision trees, but the cited chapter is titled \"Clustering methods,\" which appears to be a citation error.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript reads more like an extended abstract than a complete journal paper. The central equivalence claim is not supported by any derivation, code, or reproducible experimental protocol, and the numerical inconsistency in Section 4 suggests that the efficiency argument has not been carefully checked. In my view, this is not salvageable by a normal revision within the paper's current scope; it would require a new correctness analysis, a full experimental rework with code and data release, and a precise specification of the threshold selection procedure. I would advise against sending the manuscript to full review in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the review request. Quick take: the paper is a short, sometimes clear exposition of pdsCART, a MapReduce parallelization of dsCART, but the main scientific claim in Section 4—that pdsCART generates \"exact same decision trees\" as dsCART with identical accuracy—is not established. There is no code, no dataset details, no units on the reported times, no tree-structure comparison, and no reference to the original pdsCART source. The paper explicitly says the goal is not to introduce a new model, and it reads as a report on an implementation rather than a research preprint.\n\nWhat is good: the high-level survey of parallel decision-tree and streaming algorithms is reasonably organized, and the description of histogram-based split evaluation in MapReduce—local histograms in map, merged in reduce, controller decides splits—gives a reader the basic idea. The observation that delaying split evaluation and running it in parallel can reduce wall-clock time is plausible and worth studying. The tables do show a consistent pattern: larger batch sizes reduce runtime, and accuracy stays flat. That is a real behavior if the implementation exists.\n\nWhere it falls apart: the exact-equality claim is load-bearing and unsupported. Section 3.1 says only that the selected feature \"often closely aligns\" and that trees are \"nearly identical\"; Section 4 escalates to \"exact same decision trees.\" No argument shows that deferring split evaluation to batch boundaries preserves every split point deterministically. The threshold Θ is described as \"appropriate\" and \"high-confidence\" but no value, schedule, or sensitivity analysis is given. The tables report rounded accuracies (83.11%, 84.94%, 77%) that are compatible with different trees, so they cannot verify structural equality. The D4 example says batches of 22 require \"around 20,000\" split computations, but 4 million divided by 22 is about 182,000 (4M/800=5,000 is at least consistent with the other number); that arithmetic error suggests the numbers were not carefully checked. Also, pdsCART itself is never cited, so a reader cannot audit whether \"pdsCART\" here matches anything in the literature. The citation pattern is heavy on introductory/big-data references and thinner on the streaming decision-tree literature that matters.\n\nNet: if the implementation exists and the equivalence can be shown, there may be a useful engineering note for people deploying dsCART at scale. As it stands, the paper is not ready for peer review: the central claim is unverifiable and the experiments are internally inconsistent. I would desk-reject it and tell the author to cite the original pdsCART, provide code or a detailed derivation for exact equivalence, fix the experimental reporting, and either publish a software/benchmark paper or revise as a tutorial.","headline":"A student-style exposition of pdsCART whose central claim—exact tree equivalence with dsCART—is asserted rather than demonstrated, and the experiments as reported do not support it.","tokens_in":7595,"tokens_out":3260,"would_cite":false,"duration_ms":29061,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A parallel decision-tree learner built on dsCART is claimed to produce identical trees and accuracy at lower runtime.","keywords":["parallel decision trees","data streams","MapReduce","CART","dsCART","histograms","Gini index","single-pass learning"],"falsifier":"Run pdsCART and dsCART on a stream constructed so that early batches misrepresent the eventual class-feature distribution, such as features that become informative only after many records, and compare the trees node by node under the paper's parameter settings; any node where the chosen split differs refutes the exact-same-trees claim.","tokens_in":6632,"feed_emoji":"🌲","tokens_out":6751,"duration_ms":66123,"temperature":0.7,"pith_summary":"This paper presents and analyzes pdsCART, a parallel version of the streaming decision-tree algorithm dsCART. The central claim is that pdsCART, by horizontally partitioning stream records across MapReduce mappers and merging per-leaf histograms, generates exactly the same decision trees as dsCART with identical accuracy, while reducing execution time. The paper argues this shows existing single-pass stream learners can be adapted to distributed processing without sacrificing predictive behavior. If true, pdsCART is a drop-in parallel replacement for dsCART.","feed_headline":"Parallel decision-tree learner matches sequential trees exactly","feed_subtitle":"It partitions streams across mappers and reports the same accuracy as dsCART in less time.","key_machinery":"The machinery is a horizontal partitioning of the stream combined with per-leaf histograms aggregated through MapReduce. Mappers route each incoming record to its current leaf and update local histograms of feature and class-label counts; reducers merge these into a global histogram per leaf; and a controller reads the merged histograms to estimate the top two candidate splitting features for each leaf and applies the threshold Θ to decide whether to split. These histograms let the algorithm compute Gini-index improvements without sorting or repeatedly scanning the full dataset, which is what makes single-pass parallel construction possible.","core_discovery":"The paper's central claim is that pdsCART, a horizontal parallelization of the streaming CART algorithm dsCART implemented under MapReduce, produces exactly the same decision trees as dsCART whenever the split-confidence threshold Θ is tuned appropriately. The authors report that across all test cases, aside from the number of records processed, the two implementations generated the same trees with identical accuracy levels, so the comparison then focuses on runtime rather than predictive performance. The argument rests on the observation that a feature chosen from a limited stream segment often closely aligns with the feature chosen from the full dataset, which motivates estimating splits only after batches of records and still obtaining the same structure. They treat the result as a proof of concept that an existing single-pass stream learner can be adapted to distributed processing without changing its output, rather than as a new learning model.","pith_inferences":["The paper's identical-tree claim is empirical and dataset-specific; a principled extension would be a probabilistic bound on how often a batch-limited split matches the full-data split, which the paper leaves unquantified.","The histogram-partition observation suggests a natural further parallelization in which each candidate split's Gini computation runs independently, an extension the author flags as future work.","The same batch-limited split-estimation strategy could in principle be applied to other stream tree learners such as Hoeffding trees, but the paper tests only the CART-based dsCART.","For streams with concept drift, 'identical trees' is unlikely to hold unless the threshold or histograms are adapted, since the motivating alignment property assumes a stationary relationship between the stream segment and the full data."],"forward_implications":["If the result holds, pdsCART can replace dsCART as a drop-in parallel implementation, preserving the same tree structure and accuracy while reducing training time.","Larger record batches per split evaluation decrease runtime because fewer split computations are needed, though they add latency, so stream applications can trade responsiveness for throughput.","Runtime grows with the number of features and histogram bins because every candidate partition must be evaluated, pointing to per-partition parallel split evaluation as a next step.","Because the tree is built in a single pass, the method is suited to data streams that cannot be revisited, unlike earlier parallel tree methods that reprocess the full dataset at each level."],"supporting_citations":[{"why":"Supplies dsCART, the streaming CART algorithm whose serial behavior is the reference pdsCART must reproduce.","marker":"[28]"},{"why":"Provides the MapReduce programming model that structures pdsCART's map/reduce split-evaluation pipeline.","marker":"[15]"},{"why":"Defines the Hoeffding-tree/VFDT family of stream learners that pdsCART positions itself against.","marker":"[27]"},{"why":"Introduces histogram-based split estimation that pdsCART's local and global histograms rely on.","marker":"[23]"},{"why":"Demonstrates prior MapReduce parallelization of tree ensembles and serves as the comparative parallel baseline.","marker":"[24]"},{"why":"Supplies a parallel decision-tree algorithm that pdsCART distinguishes itself from.","marker":"[22]"}],"fun_headline_variants":["Parallel stream learner matches CART trees exactly","MapReduce CART gives identical trees, quicker","pdsCART replicates dsCART trees in parallel","Exact tree match from parallel data streams","Same accuracy, faster: pdsCART on MapReduce"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that a split chosen after seeing only a batch of stream records will match the split that would be chosen from the full record set; if that alignment fails, the parallel tree stops being identical to the sequential one.","fun_headline_variants_meta":{"raw":{"variants":["Parallel stream learner matches CART trees exactly","MapReduce CART gives identical trees, quicker","pdsCART replicates dsCART trees in parallel","Exact tree match from parallel data streams","Same accuracy, faster: pdsCART on MapReduce"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1326,"prompt_tokens":788,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":404,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":404,"tokens_out":538,"duration_ms":5358,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:47:54.542169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run pdsCART and dsCART on a stream constructed so that early batches misrepresent the eventual class-feature distribution, such as features that become informative only after many records, and compare the trees node by node under the paper's parameter settings; any node where the chosen split differs refutes the exact-same-trees claim.","supporting_citations":[{"cited_title":"The cart decision tree for mining data streams","cited_arxiv_id":null,"evidence_quote":"Supplies dsCART, the streaming CART algorithm whose serial behavior is the reference pdsCART must reproduce."},{"cited_title":"Mapreduce: simplifie d data processing on large clusters","cited_arxiv_id":null,"evidence_quote":"Provides the MapReduce programming model that structures pdsCART's map/reduce split-evaluation pipeline."},{"cited_title":"A vfdt algorithm optimization and appl ication thereof in data stream classification","cited_arxiv_id":null,"evidence_quote":"Defines the Hoeffding-tree/VFDT family of stream learners that pdsCART positions itself against."},{"cited_title":"Mapreduce in the clouds for science","cited_arxiv_id":null,"evidence_quote":"Introduces histogram-based split estimation that pdsCART's local and global histograms rely on."},{"cited_title":"Planet: massively parallel learning of tree ensembles with mapreduce","cited_arxiv_id":null,"evidence_quote":"Demonstrates prior MapReduce parallelization of tree ensembles and serves as the comparative parallel baseline."},{"cited_title":"Scalpa rc: A new scalable and efficient parallel classifica- tion algorithm for mining large datasets","cited_arxiv_id":null,"evidence_quote":"Supplies a parallel decision-tree algorithm that pdsCART distinguishes itself from."}],"review_version":1}