Research map / Time series / Foundation models for time series
Job-SDF: A Multi-Granularity Dataset for Job Skill Demand Forecasting and Benchmarking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Job-SDF, a multi-granularity dataset of monthly skill demand time series built from millions of Chinese online job ads, and benchmarks 17 time-series forecasters with MAE/RMSE, SMAPE/RRMSE, and structural-break analyses.
Problem
Job skill demand forecasting lacks comprehensive, publicly available datasets, and existing work covers mainly occupation-level demand and uses MSE/RMSE, which poorly reflects low-demand skills.
Approach
Construct a monthly skill-demand time series dataset at occupation, company, and regional granularities (and their combinations) from job advertisements, and benchmark a broad range of time-series models with scale-independent metrics and structural-break splits.
Key results
- FiLM achieves the best MAE/RMSE in most settings (e.g., L1-Occupation MAE 12.95, RMSE 117.17), while PatchTST attains the best SMAPE in most cases (L1-Occupation SMAPE 34.70%).
- Koopa and NStransformer perform well on low-demand skill prediction measured by SMAPE, while Prophet and Autoformer/FEDformer perform poorly.
- Performance under structural breaks is substantially worse than on the full dataset (e.g., L1-Occupation MAE rises from 12.95 for FiLM to 62.86 on break series).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use patch-based time-series input instead of point-wise Transformer modeling | PatchTST on Job-SDF · Job-skill demand forecasting | MAE and RMSE | Helped |
| Use segment-wise iterations rather than conventional recurrent prediction | SegRNN on Job-SDF · Job-skill demand forecasting | MAE and RMSE | Helped |
| Use FiLM for forecasting | Job-SDF benchmark models · Most granularities | MAE and RMSE | Helped |
| Evaluate on time series with structural breaks rather than the full dataset | Job-SDF benchmark models · Skill-demand forecasting | Predictive performance | Hurt |
| Forecast at finer granularity rather than coarser granularity | Job-SDF benchmark models · Different granularities | RRMSE | Hurt |
Surprising findings
- Simple linear model DLinear outperforms most Transformer-based models, while the more complex TSMixer performs poorly, suggesting overfitting.
- Prophet, a classic statistical model, performs relatively poorly, and the Transformer variants designed for long-range dependencies underperform on these short series.
Limitations
- Dataset covers only Chinese online recruitment platforms and 7 regions in China.
- Skill graph is built from co-occurrence in training data rather than expert-curated skill relationships.
- Several graph-based baselines (CHGH, Pre-DyGAE) perform poorly due to mismatch with the dataset's lack of skill supply data.
Experimental setup
Datasets: Job-SDF
Benchmarks: Job-SDF
Metrics: MAE RMSE SMAPE RRMSE
Baselines: ARIMA Prophet LSTM SegRNN CHGH Pre-DyGAE Transformer Autoformer Informer Reformer FEDformer NStransformer
Builds on: ARIMA PatchTST FiLM Koopa SegRNN
Closest papers by meaning
- A Cross-View Hierarchical Graph Learning Hypernetwork for Skill Demand-Supply Joint Prediction (AAAI 2024 · similarity 0.87)
- Unearthing Skill-level Insights for Understanding Trade-offs of Foundation Models (ICLR 2025 · similarity 0.81)
- Probing Social Bias in Labor Market Text Generation by ChatGPT: A Masked Language Model Approach (NeurIPS 2024 · similarity 0.80)
- Cognitive Bias and Reassignment: Who Can Contribute High Quality LLM Data (AAAI 2025 · similarity 0.80)
- Exploring Large Language Model for Graph Data Understanding in Online Job Recommendations (AAAI 2024 · similarity 0.80)
- Your Career Path Matters in Person-Job Fit (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.