Rethinking negative sampling in content-based news recommendation

Miguel Ângelo Rebelo, João Vinagre, Ivo Pereira, Álvaro Figueira·November 13, 2024

Summary

The study emphasizes negative sampling's importance in neural models for content-based news recommendation. It impacts model outcomes, aids decentralization, and competes with state-of-the-art accuracy. The technique reduces complexity, accelerates training, and supports privacy and scalability in decentralized models. Personalized news retrieval enhances user satisfaction, addressing information overload. The paper introduces a decentralized neural news recommendation system (DNNR) that employs a novel negative sampling technique to provide better implicit negative examples, tackling the cold-start problem and improving relevance. The system uses content-rich news articles enabled by recent natural language processing advances. The decentralized nature of the proposed method addresses issues in an information-centric age, such as cost, latency, and privacy, by hosting computation tasks closer to data sources and users. The text discusses advancements in content-based news recommendation, focusing on negative sampling, user profiling, and neural network models. It explores the trade-offs between negative sample size, predictive performance, and training/prediction times. The paper reviews related work, highlighting Collaborative Filtering's common use in recommendation systems and content-based approaches' prevalence in news recommendation due to text content's analytical potential.

Key findings

15

Tables

1

Advanced features