Personalized Career Recommendation Using FT-Transformer, Lambda MART Ranking, and Knowledge-Graph Skill Path Optimization

Abstract— Learners are different in terms of their academic preparation, technical skills, practical evidence, preferences and readiness for further training. There is no one-career label approach to career coaching. In this research, we introduce a novel AI-based system with an FT-Transformer based classifier, LambdaMART learning-to-rank model, and a skill-path extension based on the knowledge-graph. The Feature Tokeniser turns the numerical and categorical learner qualities into tokens, and the Transformer learns the interactions between the academic, skill, experience and preference variables. LambdaMART gets the most likely career families and re-ranks career families by profile-career criteria including skill alignment, work-style fit, practical evidence, learning effort and bounded labor-demand information. A* search is used to find an optimal set of missing competencies for the given vocation for the given cost. The roles, skills, prerequisites, courses and projects are represented by a separate knowledge graph. We show the computational workflow through a controlled synthetic evaluation. The FT-Transformer can achieve 94.6% accuracy and 93.9% macro-F1, and LambdaMART can improve the NDCG at 5 from 0.881 to 0.927. This creates a dashboard showing a ranked list of careers with a clear linear path of skill development, as opposed to an incomprehensible recommendation.