Using chained machine learning models for scientific articles recommendation

Abstract

Recommender systems are commonly used when it comes to online multimedia service providers or worldwide retail companies. Although, regarding educational resources, scientific papers and books, or other items with extensive textual content and description, recommendation systems are only in early development. In this paper, we propose a new approach entirely based on chained machine learning model store present and rank scientific papers. The first model a word embeddings model supported on a shallow neural network - is trained using a synthesized paper unit - a composition of the title, the abstract, the publishing conference or journal, and the year - that accurately captures paper’s semantic information. Later we train pairwise learning to a rank model based on a support vector machine (SVM) using relevant and irrelevant papers. We show that our approach achieves state-of-art results and does not rely on any language dependent or domain knowledge. It only uses available on-line data and proves to be efficient in either user-dependent and user independent modeling.