PaperlayerView exact revision on arXiv

Source-linked AI summary

Personalized News Recommendation: Methods and Challenges

Chuhan Wu, Fangzhao Wu, Yongfeng Huang, Xing Xie

arXiv:2106.08934v3cs.IR

TL;DR

Personalized news recommendation helps users find relevant news amid information overload, but important problems and challenges remain. This survey organizes the field around core recommender-system problems, reviews associated techniques, datasets, evaluation methods, and responsible recommendation, and identifies future research directions.

  • Problem

    News platforms need personalized recommendation to help users find relevant information because the volume of published news makes browsing all available articles impossible.

  • Method

    The survey reviews personalized news recommendation through core problems and associated techniques and challenges, covering modeling, ranking, training, datasets, evaluation, and responsible systems.

  • Results

    The paper provides a comprehensive overview of technologies, datasets, evaluation metrics, responsible recommendation, and future directions in personalized news recommendation.

  • Takeaways & Limitations

    The core-problem perspective offers a broad view for understanding advances and open challenges in personalized news recommendation.

  • Takeaways & Limitations

    News ranking still faces unresolved challenges in accurately and efficiently evaluating relevance, mitigating filter bubbles, and exploring potential interests without harming user experience.

Abstract

from arXiv · show

Personalized news recommendation is an important technique to help users find their interested news information and alleviate their information overload. It has been extensively studied over decades and has achieved notable success in improving users' news reading experience. However, there are still many unsolved problems and challenges that need to be further studied. To help researchers master the advances in personalized news recommendation over the past years, in this paper we present a comprehensive overview of personalized news recommendation. Instead of following the conventional taxonomy of news recommendation methods, in this paper we propose a novel perspective to understand personalized news recommendation based on its core problems and the associated techniques and challenges. We first review the techniques for tackling each core problem in a personalized news recommender system and the challenges they face. Next, we introduce the public datasets and evaluation methods for personalized news recommendation. We then discuss the key points on improving the responsibility of personalized news recommender systems. Finally, we raise several research directions that are worth investigating in the future. This paper can provide up-to-date and comprehensive views to help readers understand the personalized news recommendation field. We hope this paper can facilitate research on personalized news recommendation as well as related fields in natural language processing and data mining.

1 INTRODUCTION

Personalized news recommendation helps users find relevant information amid the volume of online news. This survey reviews the field through core recommender-system problems rather than conventional method categories.

  • Online platforms publish too many articles for users to browse all available news, creating information overload.
  • A personalized news recommender recalls candidate articles, ranks them using inferred user interests, displays the top K, and updates user profiles from behavior.
  • The survey organizes research around core problems, associated techniques, and challenges instead of collaborative filtering, content-based, and hybrid categories.
  • Its coverage includes news modeling, user modeling, personalized ranking, model training, datasets, benchmarks, evaluation, responsible recommendation, and future directions.

2 FRAMEWORK OF PERSONALIZED NEWS RECOMMENDATION

The framework decomposes personalized news recommendation into modeling news and users, ranking candidate articles, training models, and evaluating results. It also highlights challenges involving cold starts, relevance, exploration, feedback, metrics, and responsible recommendation.

  • 2 FRAMEWORK OF PERSONALIZED NEWS RECOMMENDATION: The framework treats news modeling, user modeling, personalized ranking, model training, and evaluation as key problems in recommender development.
  • 2.1 News Modeling: ID-based news representations suffer severe cold-start problems because new articles appear continuously and old articles quickly disappear.
  • 2.1 News Modeling: Deep learning-based news models automatically learn informative representations and can better understand news content than traditional feature-based methods.
  • 2.7 Responsible News Recommendation: Responsible recommendation extends accuracy goals to privacy, diversity, fairness, and protection against harmful or low-quality content.

3 NEWS MODELING

News modeling represents articles using handcrafted features or learned deep representations to capture their content and characteristics. The survey covers collaborative, textual, property, contextual, hybrid, neural, and knowledge-enhanced approaches, alongside cold-start and semantic-understanding challenges.

  • News modeling approaches: News modeling uses feature-based methods with handcrafted representations or deep learning methods that learn hidden representations from raw inputs.The survey distinguishes categories by whether news representations are manually designed or learned from scratch.
  • Feature-based News Modeling: ID-based collaborative filtering representations are vulnerable to cold start because newly published news quickly lack collaborative signals.Content-agnostic ID representations are difficult to apply to novel articles on fast-changing news websites.
  • Feature-based News Modeling: Feature-based systems combine textual, semantic, topic, emotion, vision, property, and context features to represent news content and its publishing environment.Property features include categories, locations, and publishers; context features include popularity, recency, and impression timestamps.
  • Feature-based News Modeling: Hybrid methods combine news IDs with content or topic features, mitigating drawbacks of purely ID-based modeling while retaining collaborative signals.Examples combine IDs with bag-of-words, keywords, categories, topics, click frequency, or category weights.
  • Deep Learning-based News Modeling: Neural news models learn representations from news text using architectures including autoencoders, CNNs, self-attention, and pre-trained language models.Knowledge-aware methods further incorporate entities, knowledge-graph embeddings, and complementary commonsense information.
  • Deep Learning-based News Modeling: CNN and related neural architectures improve text modeling but may remain insufficient for capturing deep semantic information in news texts.The survey identifies shallow semantic modeling as a limitation of these approaches.

4 USER MODELING

User modeling infers personal interests primarily from user behaviors and supplementary features, using handcrafted or neural representations. The survey covers temporal, sequential, attention-based, and graph-based techniques, while highlighting sparsity, efficiency, and representation trade-offs.

  • User Modeling: User modeling infers personal news interests from behavior-based news representations and may incorporate additional user features for personalization.The framework bases user modeling on news users interact with, supplemented by user-specific information.
  • Feature-based User Modeling: Feature-based methods represent users with IDs, clicked-news features, segments, demographics, feedback, graphs, or combinations of IDs and other features.These methods address user representation through handcrafted behavioral and supplementary information.
  • Feature-based User Modeling: Sparse click behavior limits the accuracy of feature-based user models, while ID-based representations additionally suffer from data sparsity.Averaging LDA features from clicked news is difficult when users have few click behaviors.
  • Feature-based User Modeling: Modeling both long-term and short-term interests captures the evolution of user preferences more effectively than representing only one timescale.Long-term interests may use tag frequencies or category-level word statistics, while short-term interests use recent clicks.
  • Deep Learning-based User Modeling: Deep learning methods learn user representations from click histories using aggregation, recurrent networks, CNNs, self-attention, co-attention, and graph neural networks.RNNs model sequential dependencies and interest dynamics, whereas graph methods capture high-order relations among behaviors.
  • Deep Learning-based User Modeling: Candidate-aware attention dynamically constructs user representations for each candidate but requires retaining all clicked-news representations during testing, reducing efficiency.Global or user-specific attention is more efficient but cannot fully model candidate-clicked-news relatedness.

5 PERSONALIZED RANKING

Personalized ranking methods seek to match candidate news with user interests while balancing relevance, diversity, efficiency, and long-term user experience. The section highlights fine-grained matching and reinforcement learning as responses to limitations of unified relevance matching and greedy ranking.

  • Relevance-based ranking: Relevance-based ranking methods score candidate news by their personalized relevance to user interests, often using similarities between user and news representations.Two-tower architectures support efficient inference by precomputing user and news features.
  • Relevance-based ranking: Fine-grained matching models relatedness between candidate news and clicked news to address the limits of matching candidates against a unified user-interest representation.FIM uses word-level interactions and 3-D CNNs to compute relevance scores.
  • Diversity and serendipity: Relevance-based methods may create filter bubbles by favoring news similar to previous clicks, motivating ranking strategies that incorporate novelty, popularity, recency, diversity, or serendipity.Newsjunkie ranks by novelty, while SCENE refines relevance-based rankings using popularity and recency.
  • Reinforcement learning-based ranking: Reinforcement learning-based ranking optimizes long-term reward rather than only current-candidate objectives, with approaches including LinUCB and deep reinforcement learning.DRN uses a DQN to estimate a reward combining click labels and user activeness.
  • Challenges: News ranking remains challenged by accurately and efficiently modeling candidate–interest relevance, mitigating filter bubbles, and exploring potential interests without harming user experience.Fine-grained matching can improve relevance modeling but is more time-consuming because user representations depend on candidate news.

6 MODEL TRAINING

Model training for personalized news recommendation relies mainly on abundant implicit feedback and increasingly incorporates auxiliary tasks and richer behavioral signals. The section identifies unresolved challenges in sample construction, scalability, offline-to-online mismatch, and privacy-preserving training environments.

  • Training objectives: Most news recommendation models use abundant click feedback as the prediction target because explicit ratings are usually too sparse for accurate training.Click prediction may classify whether a candidate news article will be clicked, while contrastive methods exploit clicked and nonclicked examples.
  • Training objectives: Additional training tasks use reading satisfaction, dwell time, finish behavior, topics, sentiment, or other news information to enrich model learning.CPRS jointly predicts clicks and reading satisfaction, while FeedRec uses click, dwell-time, and finish-prediction tasks.
  • Training environment: Offline training on centralized data and computing resources can face resource limitations, offline-to-online gaps, and privacy concerns in real-world recommender systems.Large models may require distributed learning and acceleration mechanisms for industrial practice.
  • Training samples: Randomly sampled negative examples may be too easy to distinguish, motivating more effective strategies for constructing representative labeled training samples.Negative sampling is identified as an important model-training problem beyond selecting positive samples.
  • Future improvements: Future model-training improvements include better tasks and samples, adaptive multi-task loss tuning, and more effective, efficient, and privacy-preserving environments.The survey also provides a public repository containing a joint table of recent methods’ modeling, ranking, and training details.

7 EVALUATION METRICS

Evaluation metrics assess personalized news recommenders primarily through ranking relevance, while also covering regression, classification, engagement, novelty, diversity, fairness, and privacy.

  • Ranking and classification: AUC is widely used when news recommendation is formulated as a classification task.Precision, recall, and F1 are additional classification metrics.
  • Regression: MAE, MSE, RMSE, and PCC evaluate recommendation methods formulated as regression tasks.These metrics compare real and predicted ratings or their statistical correlation.
  • Ranking and classification: AP, HR, MRR, and nDCG evaluate ranking tasks, including top-K variants such as HR@K and nDCG@K.For these metrics, relevance is based on whether ranked news was clicked.
  • Beyond accuracy: Novelty, dwell time, finish ratio, and other engagement metrics evaluate recommendation qualities beyond ranking accuracy.Novelty may be judged by human ratings, while engagement metrics can include average dwell time and finish ratio.
  • Beyond accuracy: Diversity metrics compare recommended items or sentiment with previously clicked or candidate news, while fairness metrics assess user- and provider-side equity.EAGF and SPD measure fairness across author groups, with higher EAGF and lower SPD indicating better fairness.
  • Beyond accuracy: Privacy protection can be evaluated through model-gradient privacy budgets and membership inference attacks on user behavior histories.These measures target the degree of privacy protection in federated or privacy-preserving recommendation.

8 DATASET, COMPETITION AND BENCHMARK

Personalized news recommendation research relies on a small set of public datasets and competition benchmarks. MIND provides standardized splits, consistent metrics, and a public leaderboard that support head-to-head evaluation.

  • Public datasets: Only a few public datasets exist, while many studies use proprietary data collected from major news platforms.The public datasets include plista, Adressa, Globo, Yahoo!, and MIND.
  • Public datasets: MIND contains logs from 1 million users over 6 weeks, covering 161,013 articles, 15,777,377 impressions, and 24,155,470 clicks.It is a large-scale English dataset released by MSN News.
  • Public datasets: MIND supports standardized comparison because its training, validation, and test samples and evaluation metrics are consistent.Different preprocessing methods on datasets such as Adressa make cross-paper comparisons difficult.
  • Competitions and benchmarks: NEWSREEL evaluates recommendation algorithms in a living lab on an operating service, with high click-through rate as the goal for one task.The challenge ran from 2013 to 2017.
  • Competitions and benchmarks: 73.04% AUC is the current top MIND leaderboard result, achieved by UniUM-Fastformer-Pretrain.The top submission during the competition achieved 71.33% AUC, while the current leaderboard result is higher.

9 RESPONSIBLE PERSONALIZED NEWS RECOMMENDATION

Responsible personalized news recommendation must address privacy, bias and fairness, diversity, and content quality alongside user-interest targeting. The survey describes existing approaches and their remaining deployment and measurement challenges.

  • Privacy protection: Centralized storage of user behavior data creates privacy concerns and risks of data leakage.Only a few works study privacy preservation in news recommendation.
  • Privacy protection: Uni-FedRec combines server-side candidate recall with local ranking and decomposes user embeddings into basis vectors for privacy-preserving serving.It improves on FedRec by considering both privacy-preserving training and serving.
  • Privacy protection: Differential privacy can cause large performance sacrifices, while federated systems face difficulties involving context and collaborative information and deploying in practice.These remain challenges for privacy-preserving news recommendation.
  • Debiasing and fairness: Presentation, popularity, and sensitive-attribute biases can distort user modeling and produce unfair recommendations.FairRec separates bias-aware and bias-free user embeddings through decomposed adversarial learning.
  • Diversity: Most methods optimize accuracy while insufficiently diversifying recommendations, although some diversify against clicked news or within the recommendation list.The survey identifies diversity as important for information variety, user experience, and engagement.
  • Content quality: Content moderation remains rarely studied despite clickbait, misinformation, adversarial clues, and harmful content in online news.Short dwell times and social-media comments or sharing can provide signals for detecting problematic content.

10 FUTURE DIRECTION AND CONCLUSION

The survey concludes that personalized news recommendation has progressed substantially but still contains unresolved problems. It highlights deeper news understanding and more comprehensive user-interest modeling as future directions.

  • Conclusion: Personalized news recommendation has made substantial progress, but many challenges and unresolved problems remain.The survey therefore raises several directions for future investigation.
  • News modeling: Advanced NLP techniques such as knowledge-aware PLMs may improve news modeling by better understanding textual content.The survey also points to multimodal information as relevant to news modeling.
  • User modeling: A universal user modeling framework is needed to represent dynamic and diverse user interests accurately and comprehensively.The proposed direction concerns inferring interest from multiple kinds of user behavior.

10.3 Effective and Efficient Personalized Ranking

Effective personalized ranking must improve relevance while remaining computationally efficient, diverse, and oriented toward long-term user engagement. The paper also identifies hyperbolic representation learning as a possible direction beyond Euclidean modeling.

  • Ranking effectiveness and efficiency: Fine-grained user–news relatedness can improve targeting but may be inefficient under limited computation and latency constraints.The paper calls for ranking methods that jointly achieve effectiveness and efficiency.
  • Accuracy and diversity: Ranking based only on relevance can cause filter bubbles, motivating strategies that trade off recommendation accuracy and diversity.
  • Long-term engagement: Greedy ranking may not maximize long-term user engagement because it considers only the current ranking list.The paper suggests optimizing long-term rewards through more suitable ranking policies.
  • Representation space: Hyperbolic representation learning may address drawbacks of Euclidean user and news modeling and personalized ranking.Existing news recommendation commonly uses Euclidean representations with inner product or cosine similarity for relevance scoring.

10.5 Unified Model Training

The section highlights the need for richer and more reliable training signals and for models that incorporate social context while protecting user privacy. Social interactions can inform user preferences, news quality, authenticity, and breaking-news detection, but are often neglected.

  • Training signals: Click signals can be noisy and biased, while task-specific supervision may be insufficient for training personalized news recommendation models.The paper identifies unified incorporation of multiple supervised signals as a future direction.
  • Social context: Comments, replies, and sharing reveal users’ opinions, preferences, and satisfaction with recommended news.
  • Social context: Users’ discussions and dissemination behaviors can help assess news content, quality, and authenticity and recognize breaking news.These signals may also support adjustment of recommendation results.
  • Open challenges: Social contexts are important but usually neglected, leaving their effects on personalization accuracy, timeliness, and quality open for study.
  • Privacy: Privacy protection is a responsibility goal for news recommenders, with federated learning identified as one approach for privacy-preserving model training.The passage notes that only a few works explore this direction.

10.8 Secure and Robust News Recommendation

Secure and robust news recommendation must address malicious users and platforms as well as adversarial inputs. The section also connects recommendation quality with diversity and information variety.

  • Security threats: Real-world news recommenders face threats from malicious users and platforms, unlike systems evaluated only in trusted environments.
  • Security threats: Poisoning attacks can promote selected items, trigger backdoors, or degrade recommender-system performance.
  • Robustness: Adversarial samples are another source of vulnerability for news recommender systems.
  • Recommendation diversity: Temporal-spatial diversity-aware recommendation seeks news that differ from one another and from users’ historically clicked news.This is intended to better satisfy preferences for information variety.

10.10 Bias-free News Recommendation

Bias-free news recommendation addresses behavioral bias, fairness across user groups, and fair exposure across news providers. The paper emphasizes that these goals require methods that manage bias propagation and provider differences.

  • Debiasing: Behavioral biases can propagate into recommendation models and be amplified through recommendation loops.Reducing their influence is important for high-quality news recommendations.
  • Fairness: Fair recommenders should provide equitable services across user groups and fair recommendation opportunities across news providers.
  • Consumer-side fairness: Consumer-side fairness requires reducing bias associated with sensitive user attributes.The paper notes that adversarial learning is mature but brittle and difficult to tune for this problem.
  • Provider-side fairness: Provider-side fairness is difficult because publishers differ in topic preferences, reputation, and other characteristics.Balancing recommendation chances across providers and publishers is therefore non-trivial.

10.12 Content Moderation in News Recommendation

Content moderation is an underexplored challenge in news recommendation because low-quality or harmful news can affect users and society. Research should examine harmful-news dynamics, impacts, and mitigation.

  • Content moderation helps platforms avoid recommending low-quality or harmful news and mitigate its effects on users and society.
  • Personalized news recommenders can imperceptibly influence users’ opinions and views when displaying personalized content.
  • Future research should analyze recommendation algorithms’ societal effects on political events, economic activities, and psychological health.

Conclusion

The survey organizes personalized news recommendation around core problems, associated techniques, datasets, evaluation, responsibility, and future directions. It also covers traditional and up-to-date deep learning methods to provide insights for extending research.

  • The survey reviews technologies across different core modules of personalized news recommender systems.
  • It introduces datasets, evaluation metrics, responsible-recommendation considerations, and future research directions.
  • Unlike conventional method taxonomies, the survey uses key problems and their associated techniques and challenges as its organizing perspective.
  • The survey comprehensively covers traditional and up-to-date deep learning techniques for personalized news recommendation.
  • The authors aim to facilitate research on personalized news recommendation and related fields in natural language processing and data mining.
Loading 2106.08934v3…