Dynamic Job Recommendation by Profiling Undergraduates Academic Performances

Job-seeking tasks are always challenging. Often, job recommendation systems require human intervention in the job-seeking process. Therefore, the study focuses on recommendation of most relevant job sectors and prioritizing companies based on a student’s profile. The objectives of this study are: (i) to identify important features that optimize job recommendation, (ii) to construct a predictive model that recommends most relevant job sectors, and (iii) to recommend companies by computing the similarity between student and job profiles. In this study, the dataset was collected from Graduate Tracer Study from a university. Additionally, a job dataset was collected to extend the training dataset. As a result, both students and job profiles are used in this study. To enhance the accuracy, several models have been utilized for classifying job sector. This includes both hierarchical and single level classification. In hierarchical classification, Random Forest and Categorical Boosting were utilized; while in single level classification, a total of 9 different machine learning models were utilized. To assess the model’s performance, the metrics such as accuracy, weighted precision, weighted recall, and weighted f1-socre, were utilized. The finding shows that Hierarchical Classification outperforms Single Level Classification, with evaluation metrics ranging from approximately 72% to 76%, whereas Single Level Classification achieved around 58% to 62%. In conclusion, the integration of BorutaShap with Bidirectional Encoder Representation Transformers with 12 transformed layers enhances the performance of Hierarchical Classification, with the highest evaluation metrics around 75%. To recommend companies, a predefined rule is utilized to filter relevant companies, then, the similarity of the companies is measured using Cosine Similarity after transforming both student and company information using Bidirectional Encoder Representation Transformers with 12 transformed layers.