Research & Papers

ViHoRec: Vietnamese hotel recommendation dataset tackles cold-start challenge

New public dataset with 18K interactions addresses sparse recommendation for Vietnamese users

Deep Dive

Minh Hoang Nguyen has released ViHoRec, the first public quality-controlled Vietnamese hotel recommendation dataset. It comprises 18,267 interactions between 6,832 users and 560 hotels, crawled from three major Vietnamese platforms: Booking.com, Traveloka, and Ivivu. The construction pipeline includes cross-platform entity resolution to reconcile hotel names and reproducible quality audits using quantitative metrics rather than ad-hoc cleaning. Privacy is preserved via HMAC pseudonyms, enabling public release without compromising user data.

ViHoRec establishes a cold-start benchmark using a temporal leave-last-one-out split. Experiments show that learned models like BPR-MF degrade sharply for users with short interaction histories — Recall@10 falls from 0.120 to 0.065 — while the UserKNN baseline remains the strongest overall. This confirms ViHoRec as a challenging sparse, cold-start-dominated testbed. The dataset is publicly available, filling a critical gap for Vietnamese-language and low-resource recommendation research.

Key Points
  • Dataset: 18,267 interactions from 6,832 users across 560 hotels from three platforms
  • Cold-start benchmark: BPR-MF Recall@10 drops 46% (0.120 to 0.065) for short-history users
  • UserKNN outperforms all learned models, highlighting sparse-data challenges

Why It Matters

Fills missing public resource for Vietnamese hotel recommendations and advances cold-start research in low-resource NLP.

📬 Get the top 10 AI stories daily