SCODIS: Job advert-derived time series for high-demand skillset discovery and prediction
In this paper, we consider a dataset compiled from online job adverts for consecutive fixed periods, to identify whether repeated and automated observation of skills requested in the job market can be used to predict the relevance of skillsets and the predominance of skills in the near future. The data, consisting of co-occurring skills observed in job adverts, is used to generate a skills graph whose nodes are skills and whose edges denote the co-occurrence appearance. To better observe and interpret the evolution of this graph over a period of time, we investigate two clustering methods that can reduce the complexity of the graph. The best performing method, evaluated according to its modularity value (0.72 for the best method followed by 0.41), is then used as a basis for the SCODIS framework, which enables the discovery of in-demand skillsets based on the observation of skills clusters in a time series. The framework is used to conduct a time series forecasting experiment, resulting in the F-measures observed at 72%, which confirms that to an extent, and with enough previous observations, it is indeed possible to identify which skillsets will dominate demand for a specific sector in the short-term.