ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation Dataset and Cold-Start Benchmark

Minh Hoang Nguyen

ViHoRec is a public Vietnamese hotel recommendation dataset and cold-start benchmark where short-history users sharply degrade learned models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-15 02:06 UTC pith:EUTJF6YD

load-bearing objection Useful low-resource hotel recsys dataset paper; entity resolution is the load-bearing uncheckable claim on abstract alone, but the resource gap and cold-start framing are real. the 3 major comments →

arxiv 2607.12946 v1 pith:EUTJF6YD submitted 2026-07-14 cs.IR cs.AI

ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation Dataset and Cold-Start Benchmark

Minh Hoang Nguyen This is my paper

classification cs.IR cs.AI

keywords hotel recommendationVietnamese recommender systemscold-startentity resolutionquality-controlled datasettemporal leave-last-one-outlow-resource recommendationprivacy-preserving release

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vietnamese recommender research has lacked a public, carefully documented hotel interaction resource. The paper introduces ViHoRec: 18,267 interactions among 6,832 users and 560 hotels, assembled from Booking.com, Traveloka, and Ivivu through a reproducible pipeline that reconciles hotel names across platforms, audits quality with quantitative metrics, and releases privacy-preserving HMAC pseudonyms. It also supplies a public cold-start benchmark based on a temporal leave-last-one-out split, data-centric ablations, and dependency-free baselines. On that split, learned models such as BPR-MF drop sharply for users with short histories (Recall@10 0.065 versus 0.120), while UserKNN remains strongest overall. The result is a sparse, cold-start-dominated public testbed for low-resource recommendation.

Core claim

ViHoRec is a quality-controlled public Vietnamese hotel recommendation resource of 18,267 interactions, 6,832 users, and 560 hotels that, under a temporal leave-last-one-out cold-start split, shows learned models degrade sharply for short-history users while UserKNN remains strongest overall, establishing a sparse cold-start-dominated testbed for low-resource recommendation.

What carries the argument

A reproducible construction pipeline that performs cross-platform hotel entity resolution, quantitative quality control, HMAC-based privacy-preserving release, and a temporal leave-last-one-out cold-start benchmark with dependency-free baselines.

Load-bearing premise

Cross-platform entity resolution correctly matches the same hotels across Booking.com, Traveloka, and Ivivu so that the interaction graph and reported cold-start numbers are not artifacts of mismatched entities.

What would settle it

Re-running the entity-resolution step with a higher-precision matching method or independent manual audit and checking whether the cold-start Recall@10 gap (BPR-MF 0.065 vs 0.120) and UserKNN ranking remain essentially unchanged.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Low-resource hotel recommendation research can now train and evaluate on a public Vietnamese interaction graph instead of private crawls.
Cold-start methods must be stress-tested on short-history users, where learned collaborative filters lose roughly half their Recall@10 relative to longer-history users.
Simple neighborhood methods such as UserKNN remain competitive baselines on this sparse regime and should be reported alongside neural models.
Data-centric ablations on the released split can quantify how quality-control choices affect ranking metrics without re-crawling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same entity-resolution-plus-quality-control template could be reused to build public recommendation datasets for other low-resource languages or verticals (restaurants, attractions) that also suffer from fragmented platform names.
Because the benchmark is deliberately cold-start dominated, methods that fuse limited interactions with side information (hotel attributes, reviews, geography) are likely to show larger relative gains here than on denser English datasets.
HMAC pseudonymization suggests a practical pattern for releasing multi-platform crawls while remaining GDPR-compatible, which other public IR datasets could adopt.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Useful low-resource hotel recsys dataset paper; entity resolution is the load-bearing uncheckable claim on abstract alone, but the resource gap and cold-start framing are real.

read the letter

This is a resource paper that fills a genuine hole: there is no public, documented Vietnamese hotel interaction set with a cold-start split. The numbers are modest (18k interactions, 6.8k users, 560 hotels) but that is exactly the sparse regime people in low-resource booking markets actually face. What is new is the combination of multi-platform crawl, claimed entity resolution + quantitative QC, HMAC release, and a temporal leave-last-one-out cold-start benchmark with simple baselines. The reported pattern—BPR-MF dropping from 0.120 to 0.065 Recall@10 on short-history users while UserKNN stays strongest—is the useful takeaway if the graph is clean.

Credit where it is due: the abstract is clear about the three practical problems (name reconciliation, reproducible QC, privacy-preserving release) and ships a public GitHub link. Dataset papers that actually release data and dependency-free baselines are more useful than most method papers that do not. Circularity risk is low; the interactions are external crawls, not fitted to produce the headline metrics.

The soft spot is real and load-bearing: cross-platform entity resolution across Booking.com, Traveloka, and Ivivu. Without measured match accuracy, blocking strategy, or audit sample, residual mismatches would silently corrupt both the QC numbers and the cold-start split. That is the single claim that has to hold for the benchmark to be trustworthy. Everything else (HMAC, temporal split, UserKNN vs BPR-MF contrast) is standard and inspectable once the artifacts are up. On abstract alone we cannot verify it, so confidence stays low until the repo and full methods are checked.

Who this is for: people working on recsys for under-resourced languages, sparse travel markets, or cold-start evaluation who need a shared Vietnamese hotel testbed. Not for anyone looking for a new algorithm. It deserves a serious referee rather than desk rejection—resource contributions in low-resource IR are scarce, and the construction claims are falsifiable once the code and data are examined. I would send it out; if entity resolution is documented with numbers, accept with minor revision. If not, major revision or reject. Worth a look at the GitHub before deciding how much time to spend.

Referee Report

3 major / 3 minor

Summary. The manuscript introduces ViHoRec, a quality-controlled Vietnamese hotel recommendation dataset of 18,267 interactions among 6,832 users and 560 hotels, crawled from Booking.com, Traveloka, and Ivivu. It claims three contributions: (i) a reproducible construction pipeline with cross-platform entity resolution and quantitative quality control; (ii) a privacy-preserving public release via HMAC pseudonyms; and (iii) a public cold-start benchmark using a temporal leave-last-one-out split, data-centric ablations, and dependency-free baselines. On the public split, learned models are reported to degrade for short-history users (BPR-MF Recall@10 0.065 vs. 0.120) while UserKNN remains strongest overall, positioning ViHoRec as a sparse, cold-start-dominated testbed for low-resource recommendation. Data are released at a public GitHub repository.

Significance. If the pipeline and entity resolution hold, ViHoRec would address a real gap: the lack of a public, documented hotel interaction resource for Vietnamese recommender research. A privacy-preserving release, fixed temporal cold-start protocol, and simple baselines would support reproducible work in a low-resource, sparse setting. The reported BPR-MF vs. UserKNN contrast under short histories is a concrete, falsifiable characterization of the testbed. Public data, a claimed reproducible pipeline, and a cold-start-focused benchmark are material contributions for IR/recsys work on non-English and low-resource domains, provided the construction claims are fully specified and auditable.

major comments (3)

[Abstract, contribution (i)] The central claim that the 18,267 interactions form a coherent multi-platform graph over 560 hotels rests on cross-platform entity resolution across Booking.com, Traveloka, and Ivivu. The abstract asserts a 'reproducible construction pipeline with cross-platform entity resolution and quantitative quality control' but reports no match precision/recall, blocking strategy, human-audit sample size, or sensitivity of the Recall@10 figures to residual mismatches. Residual false positives/negatives would silently corrupt both QC numbers and the temporal leave-last-one-out split; this is load-bearing for benchmark validity and must be quantified in the full manuscript.
[Abstract, reported metrics (BPR-MF / UserKNN)] The headline cold-start contrast (BPR-MF Recall@10 0.065 vs. 0.120 for short- vs. longer-history users; UserKNN strongest overall) is stated without error bars, confidence intervals, definition of the short-history cohort, ranking candidate set, or number of runs. For a benchmark paper these protocol details are load-bearing: without them the degradation claim and the 'UserKNN remains strongest' ranking are not inspectable or reproducible from the abstract alone.
[Abstract, contribution (i) quality control] Quantitative quality control is listed as a core contribution, yet the abstract does not name the QC metrics, thresholds, or how many interactions/hotels were removed under them. The claim of a 'quality-controlled' resource is therefore not yet falsifiable; the full text must define and report these metrics so that the 18,267 / 6,832 / 560 figures can be audited.

minor comments (3)

[Abstract] Number formatting in the abstract uses LaTeX-style escaping (18{,}267); prefer plain consistent numerals for readability outside TeX pipelines.
[Abstract, contribution (iii)] Clarify what 'dependency-free baselines' means in practice (e.g., pure Python/NumPy vs. no external recsys libraries) so that reimplementation expectations are clear.
[Abstract, data availability] Ensure the public GitHub release includes the exact temporal split, QC scripts, entity-resolution artifacts (or match logs), and baseline code needed to reproduce the stated Recall@10 numbers.

Circularity Check

0 steps flagged

No circularity: abstract-only dataset/benchmark paper with external crawled interactions and independent baseline comparisons.

full rationale

This is an abstract-only review of a dataset and cold-start benchmark paper. The claimed resource (18,267 interactions, 6,832 users, 560 hotels) is constructed from external crawls of Booking.com, Traveloka, and Ivivu, not fitted to produce the headline metrics by construction. The reported results (BPR-MF Recall@10 0.065 vs 0.120 for short-history users; UserKNN strongest overall) are empirical outcomes of standard models on a fixed temporal leave-last-one-out split; they are not definitional renamings of fitted inputs. Contributions (i)–(iii) describe a construction pipeline, privacy release, and public benchmark; none of these reduce by equation or self-citation to the target numbers. Entity-resolution accuracy is a correctness/validity risk (uninspectable match quality could corrupt the graph), but it is not circularity under the stated criteria: no self-definitional loop, no fitted parameter renamed as prediction, no load-bearing uniqueness theorem from the same authors, and no ansatz smuggled via self-citation. With only the abstract available and no equations or self-citations to inspect, the honest finding is score 0 and empty steps. The derivation chain is self-contained against external data once the split is fixed.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

Resource paper: central claim rests on crawl fidelity, entity-resolution correctness, QC metric design, temporal split validity, and HMAC privacy—not on fitted physical constants or new particles. No free parameters are fitted to force the headline Recall numbers; domain assumptions about multi-platform hotel identity and cold-start evaluation are the load-bearing premises.

axioms (4)

domain assumption Hotel entities can be reliably resolved across Booking.com, Traveloka, and Ivivu so interactions are comparable.
Contribution (i); without correct matching the interaction graph is invalid.
domain assumption Quantitative quality-control metrics (unspecified in abstract) adequately audit crawl noise and duplicates.
Abstract claims QC with reproducible metrics rather than ad hoc cleaning.
domain assumption Temporal leave-last-one-out is a valid cold-start evaluation protocol for this sparse hotel setting.
Public benchmark design; standard in recsys but still an evaluation choice.
domain assumption HMAC pseudonyms preserve privacy while keeping the set benchmarkable.
Contribution (ii); privacy–utility tradeoff assumed sufficient for public release.

pith-pipeline@v1.1.0-grok45 · 6142 in / 2495 out tokens · 25712 ms · 2026-07-15T02:06:35.947332+00:00 · methodology

0 comments

read the original abstract

Recommender-system research for Vietnamese remains limited by the absence of a public, well-documented hotel interaction resource. Building such a resource is challenging for three reasons: cross-platform hotel names must be reconciled before interactions are comparable; quality must be audited with reproducible metrics rather than ad hoc cleaning; and public release must preserve privacy while remaining benchmarkable under realistic cold-start conditions. We introduce ViHoRec, a quality-controlled Vietnamese hotel recommendation dataset of 18{,}267 interactions between 6{,}832 users and 560 hotels, crawled from Booking.com, Traveloka, and Ivivu. Our contributions are: (i) a reproducible construction pipeline with cross-platform entity resolution and quantitative quality control; (ii) a privacy-preserving release with HMAC pseudonyms; and (iii) a public cold-start benchmark with temporal leave-last-one-out split, data-centric ablations, and dependency-free baselines. On the public split, learned models degrade sharply for users with short histories (BPR-MF Recall@10: 0.065 vs. 0.120), while UserKNN remains strongest overall, establishing ViHoRec as a sparse, cold-start-dominated testbed for low-resource recommendation. All data are publicly available at https://github.com/MinhNguyenDS/ViHoRec.

ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation Dataset and Cold-Start Benchmark

Core claim

What carries the argument

Load-bearing premise

What would settle it

If this is right

Where Pith is reading between the lines

discussion (0)