PaperlayerView exact revision on arXiv

Source-linked AI summary

News Recommender System: A review of recent progress, challenges, and opportunities

Shaina Raza, Chen Ding

arXiv:2009.04964v4cs.IR

TL;DR

News recommender systems must help readers navigate large, dynamic news collections while addressing evolving interests, content quality, and behavioral effects. This survey synthesizes conventional and deep learning solutions, datasets, evaluation beyond accuracy, platforms, and user-behavior research. It reports that deep learning-based NRS research commonly addresses user modeling and timeliness, while real-world evaluation and user-experience benchmarks remain limited.

  • Problem

    News recommendation must handle dynamic content, evolving user interests, quality concerns, and behavioral effects that general recommendation methods do not fully address.

  • Method

    The survey reviews conventional and deep learning NRS algorithms, datasets, evaluation criteria, recommendation platforms, user-behavior effects, and possible remedies.

  • Results

    Deep learning-based NRS research most widely addresses user modeling and also addresses timeliness, while content quality receives comparatively little attention.

  • Takeaways & Limitations

    NRS research should consider beyond-accuracy aspects and the effects of recommendations on user behavior alongside recommendation performance.

  • Takeaways & Limitations

    Online evaluation and user studies remain limited because real-time, large-scale news evaluation is costly, and conventional recommenders can become obsolete before collecting sufficient interaction data.

Abstract

from arXiv · show

Nowadays, more and more news readers tend to read news online where they have access to millions of news articles from multiple sources. In order to help users to find the right and relevant content, news recommender systems (NRS) are developed to relieve the information overload problem and suggest news items that users might be interested in. In this paper, we highlight the major challenges faced by the news recommendation domain and identify the possible solutions from the state-of-the-art. Due to the rapid growth of building recommender systems using deep learning models, we divide our discussion in two parts. In the first part, we present an overview of the conventional recommendation solutions, datasets, evaluation criteria beyond accuracy and recommendation platforms being used in NRS. In the second part, we explain the deep learning-based recommendation solutions applied in NRS. Different from previous surveys, we also study the effects of news recommendations on user behavior and try to suggest the possible remedies to mitigate these effects. By providing the state-of-the-art knowledge, this survey can help researchers and practical professionals in their understanding of developments in news recommendation algorithms. It also sheds light on potential new directions

1 Introduction

News recommender systems address the need to help readers navigate abundant online news, but must handle timeliness, evolving preferences, content quality, personalization trade-offs, and user-behavior effects. This survey reviews conventional and deep learning approaches while extending prior surveys with behavioral effects and possible remedies.

  • Motivation: 93% of US adults read news online, increasing the need for systems that tailor timely updates to readers’ information needs.Online news is available through digital newspapers, social media, and news apps, but existing media do not prescribe procedures for offering varied news or modeling behavior effectively.
  • Challenges: Timeliness, dynamic user behavior, and manipulated content make news recommendation distinct from general recommendation.News stories have short lifetimes and arrive continuously, while readers’ preferences can evolve gradually or abruptly.
  • Challenges: Over-personalization can restrict exposure to diverse news and contribute to denial of opposing viewpoints.The paper frames this as a user-level exposure problem with societal implications for democracy.
  • Evaluation: Accuracy-centric evaluation may overlook choice satisfaction, perceived effectiveness, recommendation quality, and exposure to different viewpoints.The survey therefore emphasizes beyond-accuracy aspects when assessing news recommendation quality.
  • Survey scope: Unlike previous surveys, this review emphasizes deep learning models and examines how recommendations affect user behaviors, including possible remedies.It also identifies content quality as an emerging challenge alongside established issues such as timeliness and user modeling.
  • Survey scope: The review uses a defined search strategy across bibliographic collections and scholarly search engines, covering journal and conference literature under stated inclusion criteria.The search includes ACM Digital Library, SpringerLink, IEEE Xplore, Elsevier, and multiple scholarly search engines.

2 Characteristics of News Domain

News recommendation differs from other domains because news is rapidly changing, context-dependent, socially consequential, and consumed through evolving user behaviors. These characteristics create demands for timeliness, diversity, privacy awareness, and richer handling of content and context.

  • News-specific characteristics: News stories have short lifespans, arrive in large volumes, and require near-real-time delivery.Requests can exceed 100 per second, with preferred response times within 100ms.
  • News-specific characteristics: Readers consume news sequentially and expect different or ongoing stories rather than repeated items.This distinguishes news consumption from repeated consumption patterns common for music.
  • News-specific characteristics: Diversity is important for engagement, exposure to counter-attitudinal views, and democratic news media.The paper presents diversity as both a user-engagement concern and a societal principle.
  • Context and behavior: Anonymous consumption and imperfect implicit signals complicate modeling user interests and appreciation.Reading time, clicks, browsing, and navigation can be used as proxies, but may be misinterpreted.
  • Broader challenges: The domain also raises privacy, emotional, and bias concerns, while multimodal and multilingual news remain under-researched.The survey therefore limits its review to text-based news in one language but identifies broader coverage as needed.
  • Context and behavior: News recommendation must account for evolving reading context, including time, location, social influences, events, trends, and weather.Reading context is described as highly evolving, time-ordered, social, and specific to news.

3 Overview of Research in News Recommender Systems

The survey organizes news recommender research around conventional algorithmic approaches and the domain-specific challenges that limit their effectiveness. It emphasizes that news dynamics, context dependence, sparse interactions, and evolving interests require approaches beyond straightforward traditional methods.

  • Survey scope: The survey presents NRS research alongside major news-domain challenges and conventional solutions, followed by deep learning-based approaches.Its structure separates the research overview, challenges and conventional solutions, and deep learning-based NRS.
  • General algorithmic solutions: Traditional recommender algorithms include collaborative filtering, content-based filtering, and hybrid approaches.These approaches rely on user and item content together with their underlying interactions.
  • General algorithmic solutions: Traditional methods may perform poorly because news is dynamic, context-dependent, rapidly decaying, and often consumed by temporary or anonymous users.CF needs sufficient interaction histories, while CBF may miss semantics and struggle with anonymous users.

ALGORITHMS

The survey covers conventional and neural approaches used in news recommendation, with content-based filtering the most commonly used conventional algorithm. It also documents growing adoption of deep learning and reviews factorization-based methods for modeling sparse interactions and contextual information.

  • Conventional algorithms: Content-based filtering is the most used algorithm in NRS, followed by hybrid methods and then collaborative filtering.The paper attributes CBF’s popularity partly to the relative ease of developing systems from content metadata.
  • Factorization methods: Factorization methods decompose user-item interaction matrices into lower-dimensional representations for recommendation.Matrix factorization can discover latent features from interactions between users and items.
  • Factorization methods: NMF constrains the interaction and factor matrices to nonnegative values and can perform better than MF on sparse NRS matrices.NMF has been applied to news clustering, topic discovery, and joint news-content and user-interaction embeddings.
  • Factorization methods: Tensor factorization incorporates additional dimensions such as time, location, and social interactions, but too many dimensions increase computational cost.It is used when recommendations need richer contextual information.
  • Factorization methods: PMF scales linearly with observations and suits large, sparse, imbalanced news datasets, while BPR optimizes personalized item rankings.Related NRS work combines content, interactions, and social information in PMF models.
  • Deep learning-based solutions: Deep learning-based NRS research expanded after 2016, with more than 30 papers published since 2017 using deep neural networks.The survey reports increasing employment of deep learning each year; the 2021 count is incomplete because the paper was written mid-year.

3.3 Evaluating the Quality of Recommendations

The survey evaluates NRS quality through accuracy, beyond-accuracy, and subjective user-satisfaction measures, while noting that accuracy remains dominant and broader user experience is underrepresented.

  • NRS evaluation measures are categorized as objective accuracy, objective beyond-accuracy, and subjective user-satisfaction measures.The survey reviews how these categories are used and defines their metrics in Table 2.
  • Accuracy-centric evaluation measures algorithmic predictions against known user ratings but cannot determine whether users are satisfied with recommendations.
  • Beyond-accuracy measures: Diversity measures dissimilarity among recommended items, commonly through re-ranking, including Intra-List Similarity and temporal diversity.
  • Beyond-accuracy measures: Coverage measures the percentage of distinct items, users, ratings, or recommended user-item pairs that a system can recommend.
  • Beyond-accuracy measures: Novelty captures how different or unknown recommendations are, while serendipity combines relevance, novelty, and unexpectedness.The survey notes that novelty in NRS has mainly been addressed at the item level and should also cover content, events, and story uniqueness.
  • Evaluation evidence: Accuracy is the most widely used NRS evaluation measure, while work on novelty, coverage, and user experience remains very limited.The survey also reports 58 offline evaluations, 13 online evaluations, and 9 user studies.

3.4 Research Datasets

NRS research uses public, proprietary, crawled, and synthetic datasets, but private datasets dominate because public resources are limited and research needs vary.

  • News datasets mainly contain textual data and include publicly available, proprietary, crawled, and synthetic collections.
  • Dataset examples: Plista contains logs from 13 German news portals, including millions of article impressions and time-related information.It is accessible upon request for research purposes.
  • Dataset examples: Adressa is a public Norwegian benchmark dataset with reading time and reading counts but no explicit ratings.
  • Dataset examples: Yahoo Webscope R6A and R6B provide timestamps, explicit ratings, and news categories, but anonymize the actual story content.The missing content can make recommendation difficult when story information is needed.
  • Dataset examples: MIND contains about 160k English news articles, more than 15 million impression logs, and data from 1 million users with rich article content and click histories.
  • Dataset usage: Figure 7 reports 62 papers using private datasets and 16 using public datasets.Researchers often build proprietary datasets because public datasets are scarce or do not meet their specific requirements.

3.5 Open News Recommendation Platforms

Open platforms and libraries support reproducible, comparative, and real-time NRS research, spanning general recommendation frameworks and systems designed specifically for news.

  • Apache Mahout provides distributed collaborative-filtering algorithms for academic and commercial use with real-world news data.
  • Idomaar enables efficient, reproducible evaluation of recommendation algorithms through a programming-language-flexible web service.
  • StreamingRec offers pre-built news recommendation algorithms and simulates real-world news recommendation scenarios.
  • CLEF NEWSREEL used the Plista dataset and ORP to support real-time and replay-based evaluation through standardized interactions.
  • Platform scope: Idomaar and Apache Mahout target general recommender systems, whereas CLEF NEWSREEL, StreamingRec, and MIND were designed specifically for NRS.CLEF NEWSREEL is obsolete, while the MIND platform remains active.

4 Major Challenges in News Recommender Systems and Conventional Solutions

The survey revisits timeliness and user modeling, introduces content quality as an emerging NRS challenge, and separates conventional non-neural solutions from deep-learning approaches.

  • The survey discusses timeliness and user modeling with new insights while identifying news content quality as an emerging challenge.
  • Cold-start and data sparsity are omitted because prior surveys already covered them and they are common to general recommender systems.
  • The conventional-solutions category refers to non-neural solutions, while deep-learning-based solutions are reserved for Section 5.

4.1 Challenge 1: Timeliness

Timeliness is central to news recommendation because news value changes quickly and systems must respond rapidly. The survey reviews time-sensitive, sequential, and popularity-based approaches, while noting credibility and dynamic-behavior limitations.

  • 100ms response time is identified as a requirement for well-formed news recommendations in large-scale, real-time news environments.News recommendations also depend on popularity, recency, freshness, trends, uniqueness, and low latency.
  • Time-decay models weight recent news more heavily to capture the declining value of data over time.They require short-term preference models that predict readers’ recent news interests.
  • Sliding or timing windows retain recent news or interaction data while discarding or downweighting older information.Prior work reports different window sizes and weights, with some recommending non-fixed windows.
  • Graph-based models represent user–item links and sequential dependencies to predict the next news item.Examples include Context Trees, Browse-Graphs, and Markov decision processes.
  • Popularity-based models use article visits, clickthrough rates, or social ties, but top-N selection can amplify popularity or temporal bias.Popularity does not guarantee that news is credible or genuinely popular; false stories may be 70% more likely to be retweeted than true stories.
  • Traditional timeliness models may be limited in addressing dynamic user behaviors in news recommendation systems.

4.2 Challenge 2: User Modeling

User modeling in news recommendation must handle incomplete, implicit, volatile, anonymous, and sequential preferences. The survey reviews stereotype-, feature-, collaborative-, knowledge-, and microblog-based approaches alongside their limitations.

  • News recommendation user modeling must address anonymous readers, registered-user profiles, passive consumption, negative implicit feedback, and context-dependent reading intents.
  • Stereotypical modeling assigns users to classes and infers preferences from class-level information when individual background knowledge is incomplete.NRS stereotypes include geolocation and user habits.
  • Stereotyping cannot learn completely new stereotypes and may create segregated groups or filter bubbles among like-minded users.
  • Feature-based models recommend articles using similarity between extracted news features and content previously read by the user.Common representations include BoW, TFIDF, Hashing, and Word2vec.
  • Traditional statistical profiles struggle with volatile preferences, anonymous users, and time-ordered dependencies in reading behavior.Long-term interests can be inferred from prior-reading keywords or implicit feedback, but complete profiles are difficult to obtain.
  • Collaborative filtering requires temporal distinctions in user preferences to predict the next news article effectively.Without preserved temporal information, similar-user histories may not capture changing interests.
  • Knowledge-based models apply semantics, ontologies, and contextual information, while microblogging models use social-media interactions for personalized and trending news.Microblog content requires additional quality assessment because it is generally less trustworthy than curated news stories.
  • The survey concludes that user modeling should include short-term, seasonal, diversified, and sequential interests in addition to user histories.

4.3 Challenge 3: Quality Control of the News Content

News recommender systems must process large, unstructured, real-time content while controlling duplication, semantic gaps, spam, and publisher bias. Existing work addresses these issues unevenly, with content quality receiving limited attention.

  • Online news creates a massive, largely textual information stream that must be handled and evaluated efficiently in real time.Big-data technologies and modern NLP have partially addressed efficiency, scalability, and representation challenges.
  • News-content quality problems include duplication, missing semantics, spamming, and biases.
  • Duplicate stories from multiple URLs and sources can distort rankings and bore readers with repeated recommendations.
  • Jargon and slang with missing semantics are difficult for available NLP libraries to interpret.
  • Clickbait uses catchy headlines to manipulate clicks, while changing tactics make its hidden meanings difficult to extract consistently.
  • Publisher, author, and media-group writing styles and tones can introduce bias, including hyper-partisan bias.
  • Traditional content-feature methods recommend similar articles but can repeat the same story across different publishers.
  • Ontology- and taxonomy-based methods improve semantic representation but may not fully represent semantics across news bodies and titles.

5 Deep Learning Models for News Recommender Systems

Deep learning models address news recommendation through richer representations, interaction modeling, sequential modeling, and improved handling of cold-start and sparse data. The survey finds user modeling and timeliness receive attention, while content quality remains comparatively underdeveloped.

  • Deep learning is applied to news recommendation because it can address user modeling, timeliness, and other domain-specific challenges more flexibly than conventional methods.
  • Deep neural methods such as CNNs, RNNs, and BERT support representation learning for massive multimodal news data.
  • Deep learning supports sequential modeling of changing user behavior and session-based recommendation, unlike static conventional datasets.
  • Deep learning extracts features from news and user data to improve user and item profile estimation under cold-start and data-sparsity conditions.
  • Multi-Layer Perceptron (MLP): MLPs add nonlinear transformations to matrix-factorization models and learn rich user–item interactions.They are described as simple and efficient neural extensions of matrix-factorization-based models.
  • Autoencoders (AE): Autoencoders learn hidden patterns and low-dimensional representations from news content and historical user feedback.
  • Convolutional Neural Networks (CNN): CNNs extract local features from headlines and news bodies, supporting similarity-based recommendations without manual feature engineering.They can represent multimodal text, audio, and video features.
  • Recurrent Neural Networks (RNN): RNNs model variable-length sequences and session-based recommendations, while GRUs learn short-term preferences and slightly outperform LSTMs in reported work.

6 Effects of News Recommendation Algorithms on Readers’ Behavior

The survey examines how news recommendations can shape readers’ behavior, including selective exposure, polarization, and manipulation, and reviews mitigation strategies. It argues that accuracy-oriented design should be complemented by diversity and other beyond-accuracy considerations.

  • News recommendation algorithms may affect readers’ psychology, consumption behavior, and overall experience, while also influencing news production and content quality.
  • Selective exposure, filter bubbles, and echo chambers can limit exposure to counter-attitudinal news, although studies also find that online platforms may expose users to opposing viewpoints.
  • Research links recommendation environments with polarization, fragmentation, biased assimilation, dehumanization, and reinforced digital gate-keeping.
  • The survey concludes that more research is needed to include selective exposure in NRS design and develop stronger ways to prevent, detect, and reduce harmful effects.
  • Nudges can steer readers toward fake or sensational news, but transparent and ethical designs may guide users toward more politically balanced consumption.
  • Accuracy combined with beyond-accuracy measures such as coverage, diversity, and serendipity improves user behavior, while research balancing these factors remains limited.

7 Discussion on Research Implications and Future Work

The discussion identifies growing opportunities for deep-learning-based NRS research while emphasizing that conventional methods require substantial adaptation. It highlights gaps in beyond-accuracy evaluation, diversity, real-world testing, implicit feedback, and behavioral effects, and calls for interdisciplinary work.

  • NRS research is gaining attention alongside online news adoption and the rapid evolution of deep-learning methods.
  • Traditional recommendation algorithms only partially address news-specific challenges, whereas latent-factor and deep-learning models are major successful model classes.
  • The survey classifies deep-learning models to help researchers choose suitable NRS models, including restricted Boltzmann machines for feature extraction from large news datasets.
  • Deep-learning methods require more data and parameter tuning than standard methods and provide limited interpretability and explainability.
  • Beyond-accuracy research remains limited, while online evaluation and user studies are uncommon compared with offline protocols.
  • NRS diversity research is sparse, despite diversity’s role in engagement and avoiding filter bubbles; future designs should consider nudging, selective exposure, and fake-news detection.
  • User-experience benchmarking is unavailable, existing evaluation frameworks may be too expensive for real-time news settings, and implicit feedback can be ambiguous.
  • The effects of news recommendations on user behavior remain overlooked, and existing mitigation approaches are small-scale and insufficient for detecting or preventing those effects.

8 Conclusion

The conclusion synthesizes the distinctive challenges of news recommendation and the need to extend conventional algorithms. It emphasizes deep learning, beyond-accuracy evaluation, and attention to datasets, platforms, protocols, and user behavior.

  • Prominent NRS challenges include timeliness, evolving reader preferences, news-content quality, and effects of recommendations on user behavior.
  • General recommendation algorithms are insufficient for news and require substantial modification, variation, or extension.
  • Deep-learning solutions have addressed many limitations of conventional recommenders in the news domain.
  • Beyond accuracy, diversity, coverage, novelty, and serendipity are important aspects of recommendation quality and user experience.
  • Datasets, open recommendation platforms, and evaluation protocols all contribute to developing news recommendation solutions.
Loading 2009.04964v4…