Multifaceted neural news recommendation : common components, bias mitigation, and multilinguality

Iana, Andreea

Preview

PDF
IANA_DOKTORARBEIT.pdf - Published
Download (23MB)

URN:	urn:nbn:de:bsz:180-madoc-721287
Document Type:	Doctoral dissertation
Year of publication:	2026
Place of publication:	Mannheim
University:	Universität Mannheim
Evaluator:	Paulheim, Heiko
Date of oral examination:	2026
Publication language:	English
Institution:	School of Business Informatics and Mathematics > Data Science (Paulheim 2018-)
License:	Creative Commons Attribution 4.0 International (CC BY 4.0)
Subject:	004 Computer science, internet
Keywords (English):	recommender systems , neural news recommendation , bias mitigation , diversity , multilinguality , cross-lingual recommendation , responsible AI
Abstract:	Personalized neural news recommender systems (NNRs) have become central to how users access information on digital news platforms. In this thesis, we advance responsible and inclusive neural news recommendation by addressing the problem from three complementary angles. First, we systematically analyze common NNR components, showing that simpler architectures, such as late click behavior fusion, can match or outperform complex models when paired with high-quality news representations. Second, we examine algorithmic biases in recommendation, and introduce MANNeR, a modular framework for multi-aspect news recommendation enabling dynamic control over relevance and aspect-based diversity trade-offs without retraining. Third, we address linguistic inclusivity by constructing xMIND, the first multilingual news recommendation benchmark spanning 14 diverse languages, and developing a news-specialized multilingual encoder that significantly improves cross-lingual transfer, especially in cold-start scenarios. Our contributions pave the way for work on more efficient, responsible, and inclusive recommendation technology that equitably serves users across languages and contexts.