JTH v1.1 – Job Tracking History

A time-resolved, two-sided dataset for research on cold-start and temporal dynamics in job recommendation

---

1 Overview

JTH contains seven years (2018-03 → 2025-04) of real-world recruitment data curated by professional head-hunters in France. It links 37,554 candidates, 6,011 vacancies and 42,288 multi-stage application traces, encompassing 78,195 specific recruiting events each time-stamped at day granularity.

This dataset serves as the official artifact for the RecSys 2026 paper: JTH: A Dataset for Evaluating Cold-Start and Temporal Dynamics in Job Recommendation.

All records are rigorously pseudonymised (incorporating k=5 anonymity for quasi-identifiers, clipping extreme numerical outliers at the 95th percentile, and applying a 5-day Laplace noise to timestamps) and released for non-commercial research under CC BY-NC 4.0.

• Dataset Download (Zenodo): https://zenodo.org/records/21390581
• Code Repository: https://github.com/Aunsiels/JTH
• Paper DOI: https://doi.org/10.1145/3773078.3831846

---

2 File list

file	rows	size	description
candidates.csv	37,554	19 MB	static profile of each applicant
jobs.csv	6,011	2.3 MB	static profile of each vacancy
history.csv	42,288	2.7 MB	full, ordered timeline of every candidate–job process

All files are UTF-8, comma-separated, with one header row and Unix line endings.

---

3 Schema

3.1 candidates.csv

column	type	notes
candidate_id	str	deterministic hash
create_date	YYYY-MM-DD	profile creation date
skills	str	semicolon list of manual skills; _rare_skill_ sentinel for infrequent ones
expertise_area	str	manual area(s) of expertise; semicolon list
job_category	str	manual job category(ies); semicolon list
years_experience	str or NaN	binned: 0-2 years, 3-6 years, +6 years
actual_salary	float €	manually entered; µ-aggregated + Laplace noise
actual_daily_salary	float €	same as above
contract_type	str	semicolon list (e.g. Permanent;Freelance)
zipcode	str	French département code (2-digit, coarsened)
source	str	acquisition channel, anonymized

LLM-Inferred Fields (Extracted using Llama 3.3 80B):

column	type	notes
llm_sex	str	inferred sex (Male, Female)
llm_nationality	str	inferred nationality; _rare_nationality_ for infrequent items
llm_languages_spoken	str	inferred list
llm_highest_diploma	str	e.g. Bachelor, Master, PhD
llm_from_top_university	bool	top institution indicator
llm_multiple_degrees	bool	whether has multiple diplomas
llm_ongoing_education	bool	whether still studying
llm_years_of_work_experience	str	binned years: e.g. 2-5, 10-15
llm_number_of_previous_positions	int	count of prior jobs
llm_number_of_unique_employers	int	count of distinct employers
llm_management_experience	bool	management role experience
llm_years_of_management_roles	str	binned, e.g. 0-2, 10-15
llm_industry_domains	str	semicolon list of industry sectors
llm_entrepreneurial_experience	bool	entrepreneurship indicator
llm_startup_experience	bool	startup background indicator
llm_large_company_experience	bool	large company background indicator
llm_freelance_experience	bool	freelance experience
llm_contract_experience	bool	contract job experience
llm_international_work_experience	bool	worked internationally
llm_remote_work_experience	bool	remote work history
llm_average_job_duration_months	str	binned duration (e.g. 12-24)
llm_number_of_career_gaps	int	number of gaps
llm_leadership_experience	bool	leadership role experience
llm_hard_skills	str	inferred hard skills list
llm_soft_skills	str	inferred soft skills list
llm_programming_languages	str	inferred programming languages
llm_tools_technologies	str	inferred tools used
llm_certifications	str	inferred certifications list
llm_volunteer_experience	bool	volunteer activity
llm_has_publications	bool	publication record
llm_has_patents	bool	patent record
llm_has_portfolio_or_github_or_website	bool	visible portfolio presence
llm_has_linkedin	bool	contains LinkedIn link
llm_participated_in_competitions	bool	e.g. Kaggle, hackathons
llm_hobbies	str	inferred hobbies list
llm_expertise_area	str	inferred area(s) of expertise
llm_job_category	str	inferred job category(ies)
llm_estimated_salary_eur_per_year	float €	estimated annual salary
llm_client_facing_role	bool	customer-facing role experience
llm_values	str	inferred personal values
llm_seniority_level	str	Intern, Junior, Mid, Senior
llm_age_bucket	str	e.g. 30-40, 40-50, >=70
llm_graduation_recency	str	time since last graduation, binned

3.2 jobs.csv

column	type	notes
job_id	str	deterministic hash
create_date	YYYY-MM-DD	job creation date
expertise_area	str	manual expertise area(s); semicolon list
job_category	str	manual job category(ies); semicolon list
skills	str	manual skills list; _rare_skill_ sentinel
contract_type	str	e.g. Permanent, Freelance
years_experience	str or NaN	binned: 0-2, 3-6, +6 years
zipcode	str	département code; may be empty
salary	float €	manually entered
daily_rate	float €	manually entered
source	str	anonymized source code

LLM-Inferred Fields (Extracted using Llama 3.3 80B):

column	type	notes
llm_remote_possible	bool	remote option
llm_industry_domains	str	sector list
llm_company_values	str	inferred company values
llm_seniority_level	str	e.g. Mid, Senior
llm_salary_estimation_eur_per_year	float €	inferred
llm_required_languages_spoken	str	required spoken languages
llm_required_lowest_diploma	str	e.g. Bachelor, PhD
llm_required_years_of_work_experience	int	required years
llm_required_management_experience	bool	boolean
llm_is_startup	bool	startup flag
llm_is_large_company	bool	large company flag
llm_required_freelance_experience	bool	boolean
llm_required_contract_experience	bool	boolean
llm_required_international_work_experience	bool	boolean
llm_required_leadership_experience	bool	boolean
llm_hard_skills	str	required hard skills
llm_soft_skills	str	required soft skills
llm_programming_languages	str	required programming languages
llm_tools_technologies	str	required tools
llm_certifications	str	required certifications
llm_expertise_area	str	inferred areas of expertise
llm_job_category	str	inferred job category(ies)
llm_client_facing_role	bool	client-facing role indicator

3.3 history.csv

Each row = one candidate–job pair.

column	type	meaning
candidate_id / job_id	str	foreign keys
spontaneous_application_date … 4th_interview_date	YYYY-MM-DD | NaN	chronological funnel steps (11 stages)
job_offer_proposed_date / job_offer_accepted_date	YYYY-MM-DD | NaN	final outcome
end_of_process_date	YYYY-MM-DD	last recorded action
last_stage_reached	str	label of deepest non-null stage

All dates are in UTC, anonymised with a Laplace noise scale of 5 days (monotonic sorting ensures order is preserved for recruitment funnels).

---

4 Regression features

Computed features for each candidate–job pair generated for baseline classifiers.

column	type	notes
skill_coverage	float	manual skill match ratio
hard_skill_coverage	float	LLM-inferred hard skills match
soft_skill_coverage	float	LLM-inferred soft skills match
prog_language_coverage	float	LLM-inferred programming language match
tool_coverage	float	tool/tech match
certification_coverage	float	match on certifications
expertise_area_coverage	float	manual coverage match
job_category_coverage	float	manual coverage match
expertise_area_llm_coverage	float	LLM-based coverage
job_category_llm_coverage	float	LLM-based coverage
industry_coverage	float	match in industry domains
values_coverage	float	personal/company values match
language_coverage	float	spoken language match
contract_match	bool	exact contract compatibility
remote_match	bool	remote condition match
freelance_match	bool	freelance compatibility
with_contract_match	bool	any contract match
years_bucket_match	bool	manual experience years match
management_match	bool	experience vs requirement
leadership_match	bool	leadership alignment
startup_match	bool	startup compatibility
large_company_match	bool	large company experience match
international_exp_match	bool	international work experience
client_role_match	bool	client-facing role match
seniority_gap	int	numeric difference in seniority levels
diploma_match	bool	diploma requirement met
salary_gap	float €	difference between entered salaries
daily_gap	float €	difference between daily rates
est_salary_gap	float €	inferred salary gap
same_zip	bool	same département
work_years_match	bool	match on inferred years of work
top_university	bool	candidate from elite school
multi_degree	bool	candidate has several degrees
ongoing_study	bool	still studying
entrepreneurial_exp	bool	entrepreneurial experience
volunteer	bool	volunteer activities present
has_publications	bool	candidate has published
has_patents	bool	candidate holds patents
visible_portfolio	bool	portfolio or GitHub link present
has_linkedin	bool	LinkedIn link in CV
entered_competitions	bool	participated in competitions
n_prev_positions	int	number of prior roles
n_unique_employers	int	count of unique employers
years_mgmt_roles	int	years in management
avg_job_duration	float	months
n_career_gaps	int	number of gaps
entity_delta_days	int	job - candidate creation
rec_delta_days	int	recommendation date - job
existed_before	bool	candidate existed before recommendation

---

5 Missing-value rates

High missingness in human-entered fields underscores the sparsity in real-world platforms and the need for robust cold-start handling and LLM derivation.

       candidates   jobs

creation_date        7 %     0 %
skills               5 %     0 %
expertise_area       8 %    30 %
job_category        27 %    30 %
years_experience    83 %    89 %
salary              94 %    34 %
daily_rate          98 %    68 %
contract_type       15 %     0 %
zipcode             41 %    68 %

Empty cells in CSVs are rendered as empty strings → NaN when read with pandas.

---

6 Quick start

6.1 Quick start (code)

# 1. Generate time-aware train / test splits (binary labels)
python code/utils/binary_splitter.py data/history.csv data/binary

# 2. Run every baseline (models live in code/baseline/)
bash scripts/run_all_baselines.sh data results

# 3. Evaluate all saved predictions → results.csv
bash scripts/evaluate_all.sh data results

After step 3 you will find results.csv in the result directory, containing MRR, NDCG@K, P@K, etc. for both directions.

6.2 Quick start (example Python join)

Download the dataset and save it in the data directory.

import pandas as pd

cand = pd.read_csv("data/candidates.csv", parse_dates=["create_date"])
jobs = pd.read_csv("data/jobs.csv", parse_dates=["create_date"])
hist = pd.read_csv("data/history.csv", parse_dates=[
    "spontaneous_application_date", "shortlist_date",
    "qualification_date", "resume_sent_to_company_date",
    "1st_interview_date", "2nd_interview_date",
    "3rd_interview_date", "4th_interview_date",
    "job_offer_proposed_date", "job_offer_accepted_date",
    "end_of_process_date"
])

# join candidate → job interactions
df = hist.merge(cand, on="candidate_id", how="left") \
         .merge(jobs, on="job_id", suffixes=("_cand", "_job"))

Reproducible baselines and evaluation scripts reside in /code. Run ./scripts/run_all_baselines.sh and ./scripts/evaluate_all.sh to reproduce paper tables.

---

7 Results

The evaluation highlights the directional asymmetry between jobs and candidates, the primacy of fine-grained time features, and the necessary synergy of signals. See the paper for ablation discussions.

7.1 Candidates to Jobs

Baseline	MRR	P@1	P@5	P@10	R@1	R@5	R@10	F1@1	F1@5	F1@10	NDCG@1	NDCG@5	NDCG@10	MAP@1	MAP@5	MAP@10
Random	0.001	0.000	0.000	0.000	0.000	0.001	0.002	0.000	0.000	0.000	0.000	0.000	0.001	0.000	0.000	0.000
Jaccard	0.019	0.006	0.005	0.004	0.006	0.024	0.038	0.006	0.008	0.007	0.006	0.015	0.019	0.006	0.012	0.013
-human feat.	0.010	0.004	0.002	0.002	0.004	0.011	0.018	0.004	0.004	0.003	0.004	0.007	0.009	0.004	0.006	0.007
-LLM feat.	0.032	0.012	0.009	0.007	0.012	0.045	0.075	0.012	0.015	0.014	0.012	0.028	0.038	0.012	0.023	0.027
Popularity	0.004	0.000	0.000	0.000	0.000	0.001	0.003	0.000	0.000	0.000	0.000	0.001	0.001	0.000	0.000	0.001
Pop. + recency	0.086	0.031	0.022	0.018	0.031	0.112	0.182	0.031	0.037	0.033	0.031	0.071	0.094	0.031	0.058	0.067
Temporal	0.036	0.014	0.008	0.006	0.014	0.039	0.062	0.014	0.013	0.011	0.014	0.027	0.034	0.014	0.023	0.026
Past temporal	0.057	0.008	0.013	0.012	0.008	0.063	0.117	0.008	0.021	0.021	0.008	0.035	0.052	0.008	0.026	0.033
User-based CF	0.058	0.035	0.016	0.010	0.035	0.080	0.096	0.035	0.027	0.018	0.035	0.058	0.064	0.035	0.051	0.053
MF CF	0.033	0.021	0.009	0.005	0.021	0.043	0.053	0.021	0.014	0.010	0.021	0.032	0.036	0.021	0.029	0.030
CF + recency	0.119	0.076	0.028	0.017	0.076	0.140	0.170	0.076	0.047	0.031	0.076	0.110	0.120	0.076	0.100	0.104
Logistic reg.	0.283	0.177	0.079	0.050	0.177	0.393	0.503	0.177	0.131	0.092	0.177	0.289	0.325	0.177	0.255	0.269
-LLM feat.	0.276	0.176	0.076	0.048	0.176	0.379	0.477	0.176	0.126	0.087	0.176	0.281	0.313	0.176	0.249	0.262
-LLM/time feat.	0.025	0.008	0.006	0.006	0.008	0.031	0.057	0.008	0.010	0.010	0.008	0.019	0.027	0.008	0.015	0.019
-time feat.	0.031	0.010	0.008	0.007	0.010	0.042	0.067	0.010	0.014	0.012	0.010	0.026	0.034	0.010	0.021	0.024
-human/time feat.	0.009	0.002	0.002	0.002	0.002	0.009	0.018	0.002	0.003	0.003	0.002	0.005	0.008	0.002	0.004	0.005
-human feat.	0.146	0.061	0.042	0.032	0.061	0.209	0.316	0.061	0.070	0.057	0.061	0.137	0.171	0.061	0.113	0.127
Random forest	0.256	0.144	0.076	0.049	0.144	0.378	0.491	0.144	0.126	0.089	0.144	0.265	0.301	0.144	0.227	0.242
-time feat.	0.026	0.007	0.006	0.006	0.007	0.030	0.055	0.007	0.010	0.010	0.007	0.019	0.027	0.007	0.015	0.018
Past temporal (raw data)	0.129	0.039	0.039	0.034	0.039	0.196	0.342	0.039	0.065	0.062	0.039	0.117	0.164	0.039	0.091	0.110
CF + recency (raw data)	0.073	0.034	0.018	0.012	0.034	0.092	0.121	0.034	0.031	0.022	0.034	0.063	0.073	0.034	0.054	0.058
Logistic reg. (raw data)	0.323	0.221	0.087	0.053	0.221	0.436	0.532	0.221	0.145	0.097	0.221	0.333	0.363	0.221	0.298	0.311

7.2 Jobs to Candidates

Baseline	MRR	P@1	P@5	P@10	R@1	R@5	R@10	F1@1	F1@5	F1@10	NDCG@1	NDCG@5	NDCG@10	MAP@1	MAP@5	MAP@10
Random	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
Jaccard	0.005	0.002	0.001	0.001	0.002	0.005	0.008	0.002	0.002	0.002	0.002	0.003	0.004	0.002	0.003	0.003
-human feat.	0.003	0.001	0.000	0.000	0.001	0.002	0.003	0.001	0.001	0.001	0.001	0.002	0.002	0.001	0.001	0.002
-LLM feat.	0.010	0.002	0.002	0.002	0.002	0.012	0.022	0.002	0.004	0.004	0.002	0.007	0.010	0.002	0.005	0.007
Popularity	0.006	0.003	0.002	0.001	0.003	0.009	0.012	0.003	0.003	0.002	0.003	0.006	0.007	0.003	0.005	0.006
Pop. + recency	0.015	0.008	0.004	0.003	0.008	0.019	0.027	0.008	0.006	0.005	0.008	0.014	0.016	0.008	0.012	0.013
Temporal	0.010	0.003	0.003	0.002	0.003	0.013	0.019	0.003	0.004	0.003	0.003	0.008	0.010	0.003	0.006	0.007
Past temporal	0.019	0.000	0.001	0.002	0.000	0.003	0.023	0.000	0.001	0.004	0.000	0.001	0.007	0.000	0.001	0.003
User-based CF	0.011	0.004	0.003	0.003	0.004	0.014	0.027	0.004	0.005	0.005	0.004	0.009	0.013	0.004	0.007	0.009
MF CF	0.024	0.017	0.006	0.003	0.017	0.029	0.034	0.017	0.010	0.006	0.017	0.023	0.025	0.017	0.022	0.022
CF + recency	0.009	0.002	0.002	0.002	0.002	0.009	0.018	0.002	0.003	0.003	0.002	0.006	0.009	0.002	0.004	0.006
Logistic reg.	0.038	0.016	0.010	0.008	0.016	0.051	0.081	0.016	0.017	0.015	0.016	0.034	0.044	0.016	0.029	0.032
-LLM feat.	0.064	0.034	0.017	0.012	0.034	0.086	0.124	0.034	0.029	0.022	0.034	0.061	0.073	0.034	0.052	0.057
-LLM/time feat.	0.012	0.004	0.002	0.002	0.004	0.012	0.025	0.004	0.004	0.004	0.004	0.008	0.012	0.004	0.007	0.008
-time feat.	0.010	0.003	0.003	0.002	0.003	0.013	0.023	0.003	0.004	0.004	0.003	0.008	0.011	0.003	0.006	0.007
-human/time feat.	0.003	0.000	0.000	0.001	0.000	0.002	0.005	0.000	0.001	0.001	0.000	0.001	0.002	0.000	0.001	0.001
-human feat.	0.021	0.009	0.005	0.004	0.009	0.027	0.041	0.009	0.009	0.007	0.009	0.018	0.022	0.009	0.015	0.016
Random Forest	0.078	0.033	0.023	0.017	0.033	0.114	0.169	0.033	0.038	0.031	0.033	0.074	0.091	0.033	0.060	0.068
-time feat.	0.008	0.002	0.001	0.002	0.002	0.007	0.015	0.002	0.002	0.003	0.002	0.005	0.007	0.002	0.004	0.005
Past temporal -Anon	0.124	0.043	0.037	0.032	0.043	0.184	0.320	0.043	0.061	0.058	0.043	0.114	0.157	0.043	0.091	0.108
CF + recency -Anon.	0.008	0.002	0.001	0.001	0.002	0.007	0.011	0.002	0.002	0.002	0.002	0.005	0.006	0.002	0.004	0.004
Logistic reg. -Anon.	0.085	0.054	0.023	0.014	0.054	0.113	0.143	0.054	0.038	0.026	0.054	0.085	0.095	0.054	0.076	0.080

---

8 Recommended Future Tasks

JTH establishes a foundation for several advanced methodological paradigms in Recommender Systems:

1. Funnel-Aware and Multi-Task Learning: Leverage detailed, timestamped application trajectories to optimize for deep-funnel success.
2. Heterogeneous Graph Networks and Explainability: Connect bipartite candidate-job interactions via shared attribute nodes (e.g., skills, locations, contract types).
3. LLM Integration and Semantic Matching: Evaluate models that rely on granular semantic alignment utilizing both human-curated and Llama-extracted features.
4. Survival Modeling and Temporal Dynamics: Predict the likelihood of a vacancy expiring unfilled and address temporal biases from bursty recruiter behavior.

---

9 Licensing & permitted use

CC BY-NC 4.0 — non-commercial research and teaching only. Commercial exploitation or any attempt to re-identify individuals or companies is strictly forbidden.

---

10 Citation

If you use this dataset in your research, please cite our RecSys '26 paper:

@inproceedings{romero2026jth,
  title = {JTH: A Dataset for Evaluating Cold-Start and Temporal Dynamics in Job Recommendation},
  author = {Romero, Julien and Millet, Yann and Behar, Eric},
  booktitle = {Proceedings of the 20th ACM Conference on Recommender Systems (RecSys '26)},
  year = {2026},
  publisher = {ACM},
  doi = {10.1145/3773078.3831846},
  url = {[https://doi.org/10.1145/3773078.3831846](https://doi.org/10.1145/3773078.3831846)}
}

---

11 Change log

version	date	notes
1.1	2026-07	Camera-ready release for RecSys '26. Updated metadata/citations.
1.0	2025-08	initial public release; negative salaries clipped, README added

Contributions and issues are welcome via GitHub pull requests or the Issues tracker.