Generating Synthetic Data for Better Prediction Modeling in Skill Demand Forecasting

The changing world of work brings with it massive changes. However, it is largely unclear what exactly these changes will look like in terms of the skills required. In order to close part of this knowledge gap, we collected, classified, and analyzed 12.66 million job advertisements published online in Germany from 2014 to 2020 in the fields of administration & customer care, human resources, marketing & sales, and information technology. We specifically focus on the ability to evaluate skills in higher-level clusters and how these develop in different industries and company sizes over time. We find major differences in the speed of adjustment between the professional fields. In addition, we cluster industries according to their demand and development over time for specific skills. This allows industries to be identified in which the demand for skills is much alike. Companies that are a part of a particular cluster might therefore specifically look for employees from other companies in the same cluster, as their overall skill requirements are similar. In addition, we train a Long Short-Term Memory (LSTM) predictor with synthetically generated data using a Time-Generative-Adversarial-Network (TimeGAN) on real validation and test data, which outperforms the model trained only with real data. For the training, time series of the demand of entire industries is aggregated into single time series, which further emphasizes the benefit of generating synthetic data in applications with a small subset of data.