Improving Next-Application Prediction with Deep Personalized-Attention Neural Network

Recently, due to the ubiquity and supremacy of E-recruitment platforms, job\nrecommender systems have been largely studied. In this paper, we tackle the\nnext job application problem, which has many practical applications. In\nparticular, we propose to leverage next-item recommendation approaches to\nconsider better the job seeker's career preference to discover the next\nrelevant job postings (referred to jobs for short) they might apply for. Our\nproposed model, named Personalized-Attention Next-Application Prediction\n(PANAP), is composed of three modules. The first module learns job\nrepresentations from textual content and metadata attributes in an unsupervised\nway. The second module learns job seeker representations. It includes a\npersonalized-attention mechanism that can adapt the importance of each job in\nthe learned career preference representation to the specific job seeker's\nprofile. The attention mechanism also brings some interpretability to learned\nrepresentations. Then, the third module models the Next-Application Prediction\ntask as a top-K search process based on the similarity of representations. In\naddition, the geographic location is an essential factor that affects the\npreferences of job seekers in the recruitment domain. Therefore, we explore the\ninfluence of geographic location on the model performance from the perspective\nof negative sampling strategies. Experiments on the public CareerBuilder12\ndataset show the interest in our approach.\n