Science or Nature? A recommender system for choosing a journal for a paper

29 Oct 2022

Problem

A researcher who wants to publish in a strong journal sends a paper, waits for the editor and reviewers, and may go through several revision rounds before a final rejection. The journal then has to be chosen again and the cycle restarts. Journals on one topic differ in impact factor and quartile (Q1–Q4), but those numbers do not say what a given journal actually accepts. A hint at the first step ("do not send here, try there") saves months.

Example from the article: A department employee sent a paper to Tribology Letters and was rejected. The model gave her abstract about a 7% chance there and recommended two other journals.

Solution in brief

Train a multiclass classifier that predicts the journal from the abstract alone. Use TF-IDF word features plus abstract length, an inverse hyperbolic sine transform and gradient boosting. Explain each prediction with SHAP to see which words and style features push the abstract toward or away from a journal.

Questions this case answers
  • How to recommend the best-fitting venue for a text from its short summary when the candidate classes overlap in topic and differ in vocabulary and style
  • How to explain a text classifier's decision at the level of individual words and style features
  • How to choose the journal most likely to accept a scientific paper before submitting it?
  • How to find out why a model assigns an abstract to one journal and not another?

Tools and methods

  • TF-IDF on lemmatized abstracts — word features after removing stop words, symbols and Greek letters
  • Gradient boosting — multiclass journal classifier with cross-validation and parameter search
  • SHAP — per-word and per-feature contribution to belonging to each journal
Continue reading — it's free

Google or a one-time email code. No password.