Exploring automated methods for supporting worker re-skilling

Abstract

There are plenty job portals (e.g., linkedin.com, indeed.com, ziprecruiter.com etc), that leverage machine learning models to connect employers and job seekers via job or candidate recommendations. However, much less attention is paid to recommending specific skills that would help workers reskill or employers identify how to retrain their employees. This thesis seeks to build a system that recommends skills with the following two properties: 1) recommended skills are similar to a worker's existing skills so they are more likely to try and acquire them; 2) recommended skills increase chances of income enhancement. Existing research has largely focused on building models with employee data such as resumes and LinkedIn profiles. We instead explore the value of much-less-used employer data, i.e. language contained in job postings. The last few years have seen tremendous advances in natural language processing (NLP), including the rise of dense vector representations for text (i.e. "text embeddings") to help solve a plethora of prediction, classification, and other tasks. In building our system, we compare the performance of several language embedding models and skill valuation models to identify and recommend opportunities for re-skilling.