Leveraging User History with Transformers for News Clicking: The DArgk Approach

Code

• github.com/dkw-aau/RecSys2024Challenge
  In paperpytorch★ 0

Abstract

This paper provides an overview of the approach we used as team DArgk1 for the ACM RecSys Challenge 2024. The competition was organized by Ekstra Bladet and focused on addressing both technical and normative challenges in designing an effective and responsible online news recommender system. Our proposed method aims to model user preferences based on implicit behavior while considering the news agenda’s dynamic influence and the news items’ rapid decay. We employed deep learning models to estimate the likelihood of a user clicking on a list of articles seen during a specific timeframe. To this end, we proposed a transformer-based model capable of encoding user reading history to rank articles according to the user preferences with a focus on beyond accuracy performance for users with different preferences than the average user. Our submission achieved the 2nd rank and overall score of 0.7709 in the competition academia-track final results. We release our source code at: https://github.com/dkw-aau/RecSys2024Challenge.

Tasks

Reproductions