Skills prediction based on multi-label resume classification using CNN with model predictions explanation

Skills extraction is a critical task when creating job recommender systems. It is also useful for building skills profiles and skills knowledge bases for organizations. The aim of skills extraction is to identify the skills expressed in documents such as resumes or job postings. Several methods have been proposed to tackle this problem. These methods already perform well when it comes to extracting explicitly mentioned skills from resumes. But skills have different levels of abstraction: high-level skills can be determined by low-level ones. Instead of just extracting skill-related terms, we propose a multi-label classification architecture model based on convolutional neural networks to predict high-level skills from resumes even if they are not explicitly mentioned in these resumes. Experiments carried out on a set of anonymous IT resumes collected from the Internet have shown the effectiveness of our method reaching 98.79% of recall and 91.34% of precision. In addition, features (terms) detected by convolutional filters are projected on the input resumes in order to present to the user, the terms which contributed to the model decision.