Implementing AIRM: a new AI recruiting model for the Saudi Arabia labour market

Abstract

AbstractOne of the goals of Saudi Vision 2030 is to keep the unemployment rate at the lowest level to empower the economy. Prior research has shown that an increase in unemployment has a negative effect on a country’s Gross Domestic Product (GDP). This paper aims to utilise cutting-edge technology such as Data Lake (DL), Machine Learning (ML) and Artificial Intelligence (AI) to assist the Saudi labour market by matching job seekers with vacant positions. Currently, human experts carry out this process; however, this is time-consuming and labour-intensive. Moreover, in the Saudi labour market, this process does not use a cohesive data centre to monitor, integrate or analyse labour-market data, resulting in several inefficiencies, such as bias and latency. These inefficiencies arise from a lack of technologies and, more importantly, from having an open labour-market without a national data centre. This paper proposes a new AI Recruiting Model (AIRM) architecture that exploits DLs, ML and AI to rapidly and efficiently match job seekers to vacant positions in the Saudi labour market. A Minimum Viable Product (MVP) is employed to test the proposed AIRM architecture using a labour market dataset simulation corpus for training purposes; the architecture is further evaluated against three research collaborators who are all professionals in Human Resources (HR). As this research is data-driven in nature, it requires collaboration from domain experts. The first layer of the AIRM architecture uses balanced iterative reducing and clustering using hierarchies (BIRCH) as a clustering algorithm for the initial screening layer. The mapping layer uses sentence transformers with a robustly optimised BERT pre-training approach (RoBERTa) as the base model, and ranking is carried out using the Facebook AI Similarity Search (FAISS). Finally, the preferences layer takes the user’s preferences as a list and sorts the results using the pre-trained cross-encoders model, considering the weight of the more important words. This new AIRM has yielded favourable outcomes: This research considered accepting an AIRM selection ratified by at least one HR expert to account for the subjective character of the selection process when exclusively handled by human HR experts. The research evaluated the AIRM using two metrics: accuracy and time. The AIRM had an overall matching accuracy of 84%, with at least one expert agreeing with the system’s output. Furthermore, it completed the task in 2.4 min, whereas human experts took more than 6 days on average. Overall, the AIRM outperforms humans in task execution, making it useful in pre-selecting a group of applicants and positions. The AIRM is not limited to government services. It can also help any commercial business that uses Big Data.