JTH: A Dataset for Evaluating Cold-Start and Temporal Dynamics in Job Recommendation

Abstract

Job-matching platforms operate under brief and asymmetric lifetimes: vacancies close in weeks, while candidates are active for only a few days. Consequently, real-world recommenders must cope with pervasive cold start, extreme sparsity, and shifting temporal overlap. Existing public benchmarks lack the rich two-sided semantics and fine-grained timestamps needed to study these challenges. We present Job Tracking History (JTH), a seven-year corpus curated by professional recruiters. JTH pairs structured, LLM-enriched profiles for 38k candidates and 6k vacancies with day-level traces for 42k application trajectories, covering every stage from shortlist to offer. All data are de-identified via k-anonymity, noise injection on numeric fields, and coarse geocoding. We detail the collection and cleaning pipeline and report statistics highlighting JTH’s heavy-tailed lifetimes. Finally, we release the anonymized tables, a time-respecting evaluation framework for cold-start scenarios, and standard baseline models, establishing a new, accessible testbed for research on highly temporal job recommendation.