What is it about?
The ubiquity of personalized news recommendation has given rise to increasingly complex neural recommender architectures that aim to tailor suggestions to users' preferences. In this work, we introduce a unified framework for neural news recommendation which facilitates systematic and fair comparisons of models across three crucial dimensions. We additionally propose replacing complex user encoders with simple pooling of dot-product scores between candidate and clicked news embeddings. Extensive evaluation of a wide range of models shows that neural news recommendation can be drastically simplified.
Featured Image
Photo by Brett Jordan on Unsplash
Why is it important?
Our findings show that late fusion - replacing user encoders with simple pooling of dot-product scores between candidate and clicked news embeddings - drastically simplifies neural news recommendation, simultaneously improving performance and reducing model complexity. Additionally, we show that contrastive learning represents a viable training alternative to standard point-wise classification objectives.
Perspectives
I hope that this article will inspire more transparent evaluations in the field of neural news recommendation, as well as simpler yet effective model architectures.
Andreea Iana
Universitat Mannheim
Read the Original
This page is a summary of: Simplifying Content-Based Neural News Recommendation: On User Modeling and Training Objectives, July 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3539618.3592062.
You can read the full text:
Resources
Contributors
The following have contributed to this page
