News recommender systems (NRS) play a key role in delivering personalised content in fast-paced, high-volume environments. However, models optimised solely for accuracy often overlook important societal objectives such as fairness and diversity, leading to over-personalisation, biased exposure, and narrow content consumption. In this paper, we propose a contrastive learning framework for improving user representations in neural news recommendation.1 We build upon a bi‑encoder architecture and introduce self-supervised objectives that group semantically related news items by theme, encouraging the model to bring similar items closer in the embedding space while pushing dissimilar ones apart. This strategy mitigates embedding collapse and guides the model toward producing recommendations with broader topical coverage.
Recommender systems (RSs) are designed to provide personalised suggestions for items that users
are most likely to find relevant or appealing. By analysing users’ preferences and behaviours, these
systems extract valuable insights that help deliver suitable products and services [
Despite impressive performance in accuracy-based metrics, many NRSs sufer from
overpersonalisation and embedding bias, leading to narrow content exposure and poor representation
of user interests [
To address these challenges, researchers have increasingly turned to “beyond-accuracy” objectives,
including diversity, fairness, as complementary goals in recommendation [
In this paper, we focus on news recommender systems (NRS). We focus on improving user representations in NRSs through the lens of contrastive learning (CL). Contrastive learning (CL) is a self-supervised 1Our code is publicly available at: https://github.com/tan9zj/xnrs-CL/tree/main These authors contributed equally.
CEUR
ceur-ws.org
learning approach that pulls semantically similar instances closer in the embedding space and pushes
dissimilar ones apart [
This consideration leads us to the research focus of our study. Our main research question is:
• Main RQ: How can we optimise user embeddings to improve recommendations accuracy and
diversity? We have the following sub-research questions:
• RQ0 Can we reproduce research [
In this section, we review existing research on news recommendation systems and methods involving contrastive learning.
News recommender systems aim to provide users with personalised news content based on their preferences and browsing behaviour. A typical workflow involves three key stages: candidate news retrieval, personalised ranking, and feedback-based profile updating [ 14]. When a user visits a platform, a small set of candidate articles is recalled and then ranked based on inferred interests from historical interactions. Top-ranked articles are shown to the user, and their click behaviour is used to update profiles [ 15]. However, user interests are diverse, context-dependent, and dynamic, making accurate modelling challenging [16].
Classical recommender systems are typically divided into content-based, collaborative filtering, and
hybrid methods [17]. Content-based systems analyse item features to recommend similar content [17],
while collaborative filtering uses the preferences of similar users [ 18]. Recent work uses deep learning
to improve NRS performance. Models like NRMS [19], NAML [20], LSTUR [21], and NPA [22] adopt
diferent neural architectures (e.g., CNNs [ 23], LSTMs [24], attention mechanisms [25]) to learn
highquality representations of news content and user behaviour. These methods improve performance by
encoding contextual semantics and user interest dynamics. More recent approaches further explore
graph-based models [26] or user-news co-embedding strategies [
Contrastive learning (CL) improves representation by contrasting positive and negative samples from
diferent data views [ 27]. In the domain of news recommendation, contrastive learning is used to
improve the robustness and expressiveness of user and item embeddings, especially under sparse or
implicit feedback [28, 29, 30]. For instance, SFCNR [
With respect to diversity, the goal is to provide a varied selection of recommendations that reduce redundancy, enrich user experience, and prevent users from being confined to a narrow range of content, often referred to as the filter bubble [44]. In recommender systems, diversity generally refers to the recommended items difering from one another, either in terms of content, category, or user intent [45, 46]. Promoting diversity in recommendations is important in news domains, where repeated exposure to similar viewpoints may lead to confirmation bias and reduced information plurality [ 47, 48]. Diverse recommendations can help users discover novel content, improve user satisfaction, and even enhance long-term engagement [49]. Several strategies have been proposed to improve diversity in recommendation outputs. These include pre-processing methods that modify user profiles by adding or removing items [37]; re-ranking approaches based on pairwise item dissimilarity [45]; topic modelling techniques to promote topical variety [49]; and multi-objective optimisation approaches that balance accuracy and diversity [50]. Some approaches explicitly include diversity-aware regularisation terms in the model’s objective function, while others operate in a post-processing stage, adjusting the output list after ranking. Despite its benefits, improving diversity often comes at the cost of recommendation accuracy, leading to a trade-of that must be carefully managed depending on the application context [ 51].
This section describes our proposed news recommendation approach, which integrates a content-based bi-encoder architecture with a contrastive learning objective to optimize user representation quality. Then, we describe the training procedure, including the joint loss formulation and optimisation strategy.
MSE loss s1 dot 1 Multi-Layer Perceptron 1
Attention 2 3
InfoNCE loss Positive sample
Negative sample 1
1 Multi-Layer Perceptron
Attention ··· ′ ′1 Multi-Layer Perceptron 1
Attention 2 3 ···
Multi-Layer Perceptron 1
Attention 2 3 Pre-trained languagemodel
Pre-trained languagemodel
Pre-trained languagemodel Title w1 ··· w2 ··· w3
Title w1 ··· w2 ··· w3
Title w1 ··· w2 ··· w3 CandidateNews
Clicked News
UserEncoder News Encoder Click Prediction ContrastiveLearning
The overall structure of our model is shown in Figure 1, which consists of a news encoder, a user
encoder, and a scoring module. We build upon the work of [
We use the Siamese sentence-transformer (S-BERT), which shows the best
performance among RoBERTa, NewsBERT, and FastText [
ℎ( + )) , where are the attention weights for the individual token embeddings . They are predicted from the respective embeddings as shown in Equation 3, in which , , and are learnable parameters. The softmax function ensures that all weights sum to one. After the attention mechanism, the model further transforms the embeddings through a small multi-layer perception (MLP).
For each user, we use the same additive attention mechanism over the user’s history news embeddings for the most recently read news: = ( ∑ ).
Here are the attention weights gained similarly to through Equation 3. (1) (2) (3) (4) Contrastive learning
In our approach, we extend the standard news recommendation pipeline by integrating contrastive learning to improve user representation learning. While the candidate and historical news are encoded as usual via the news encoder and aggregated into a user embedding using the user encoder, we add an auxiliary contrastive loss on the user embeddings. This loss encourages user embeddings with similar interests (e.g., sharing the same category or theme) to be closer in the embedding space, while pushing dissimilar ones apart.
Given a batch of user embeddings u1, u2, ..., u and their corresponding class labels 1, 2, ..., , the similarity matrix is: where is a temperature scaling hyperparameter, and the embeddings are ℓ2 normalized before computing dot-product similarity. For each anchor user , the contrastive loss is defined as: where () denotes the set of indices in the batch that share the same label as (excluding itself). The ifnal loss is averaged over all users in the batch: (5) (6) (7) (8) (9) (10)
The proposed model is trained with a joint loss function that combines the standard click prediction objective and the contrastive learning objective described earlier.
The primary training objective is to predict whether a user will click on a candidate news article. Given a user embedding and a candidate news embedding , the predicted relevance score is computed as in Equation 1 and the click probability is computed by applying a sigmoid function to their dot product score:
=̂ ( ⊤ ), where (⋅) denotes the sigmoid function. Our approach employs mean squared error (MSE) loss to model user-news interaction scores as continuous relevance values.
Let ∈ [
( ∑≠ ( )
, ) = 1 || ∑ () .
=1 MSE = 1 ∑ ( − ̂ ) ,
2 = + ⋅ , where is the number of training samples in the batch.
To leverage both supervised click signals and self-supervised contrastive signals, we combine the click prediction loss with the contrastive learning loss: where is a hyperparameter that controls the relative weight of the contrastive loss. In practice, is chosen based on validation performance to balance accuracy and representation robustness.
This section presents the experimental design used to evaluate the efectiveness of our proposed approach. We describe the datasets employed, detail the experimental settings, and report key metrics to analyse model performance. For reproducibility, our code is publicly available at: https://github.com/ tan9zj/xnrs-CL/tree/main.
We run experiments on a real-world news recommendation dataset Microsoft News Dataset (MIND) [15]. MIND is a large-scale English news recommendation dataset constructed from anonymised user behaviour logs on the Microsoft News platform. It contains user click histories, impression logs, and detailed information about the news articles, including titles, abstracts, categories, and entities of the title and abstract. For our experiments, we use the MIND-small subset, which is a randomly sampled portion of the full dataset. The detailed statistics are summarised in Table 1.
To better align with the objectives of our method, we reorganise the original category labels into a set of high-level thematic groups specifically designed for our contrastive learning approach. The mapping between themes and their corresponding categories is presented in Table 2. This reorganisation serves more than one purpose: first, it facilitates the learning of more generalised user representations by grouping semantically similar categories; second, it encourages the model to consider a broader range of content types during training; third, we note that contrastive learning setups on category implicitly force each user embedding to align strongly with a single content category, which may reinforce narrow interest profiles and cause the filter bubble efect. We aim to promote diversity by encouraging the inclusion of multiple distinct categories in the final ranked list. By grouping categories into broader themes, we aim to guide the contrastive learning process to be sensitive to semantic-level distinctions. This allows user embeddings to be aligned with higher-level semantic themes rather than overly specific labels, increasing the likelihood that users are exposed to a diverse set of categories. As a result, the model generates recommendations that are both relevant and semantically varied, contributing to improved fairness and diversity.
We follow the work of Möller et al. [
There are two key hyperparameters in our contrastive learning setup: the temperature scaling factor ( ) and the contrastive loss weight ( ). To assess the sensitivity of model performance to these parameters, we perform a grid search over a predefined range.
For the temperature , we evaluate values in {0.08, 0.1, 0.9}, and for the contrastive loss weight , we explore values in {0.005, 0.01, 0.012, 0.02}. Based on the experimental results, we find that setting = 0.08 and = 0.01 yields the best overall performance, and we use these values for all subsequent experiments.
Click-Prediction Evaluation For accuracy metrics, we use the well-established ranking metrics normalised discounted cumulative gain (nDCG), mean reciprocal rank (MRR), click-through rate (CTR), and area under the receiver operating characteristic curve (AUC). These metrics evaluate the quality of ranked recommendations from multiple perspectives: nDCG assesses ranking quality by considering both relevance and position of recommended items [54]. The highly relevant items are more useful when appearing earlier in the ranking than the less relevant items. MRR evaluates the ranking quality based on the position of the first relevant item in the recommendation list. A higher MRR indicates that relevant items appear earlier in the ranking list. CTR measures user engagement by calculating the ratio of clicked recommendations to the total number of displayed recommendations. AUC measures the probability that a randomly chosen relevant item is ranked higher than a randomly chosen irrelevant item. Higher value indicates the efectiveness of a recommender system in distinguishing relevant items from irrelevant ones.
Diversity Evaluation To evaluate the diversity of the recommended results, we adopt distributionbased metrics that compare the statistical diference between the category distributions of the recommended items and a reference distribution, typically the user’s historical reading distribution. We use Kullback–Leibler (KL) divergence and Jensen–Shannon (JS) divergence to quantify this dissimilarity [55]. Fairness Evaluation To evaluate the relevance and category-level fairness of recommendations, particularly toward minority categories, we adopt a modified version of the normalised Discounted Cumulative Gain (nDCG) called fair-nDCG [56].
In this section, we analyse our model performance across multiple aspects, including recommendation accuracy, representation geometry, fairness and diversity of recommendations, and robustness.
We start by evaluating the click-prediction performance of our proposed model against several strong
neural baselines. Our method, XNRS+CL, extends the original XNRS model [
Overall, our results in Table 3, show that both contrastive learning variants of XNRS (XNRS+CL(theme) and XNRS+CL(category)) consistently outperform the original XNRS model across all metrics, nDCG, MRR, CTR, and AUC. This could be explained by the fact that contrastive learning contributes to better user-item matching, likely by improving the quality of user embeddings.
Notably, XNRS+CL(category) shows slightly stronger performance than the theme-based variant on several metrics, including nDCG@5, MRR, and CTR@1. Although the diferences are modest, this suggests that the type of semantic grouping used to define positive contrastive pairs may afect specific aspects of model behaviour. Further exploration of grouping strategies could provide additional insight into the relationship between content semantics and recommendation efectiveness.
To gain insight into how contrastive learning shapes the user representation space, we visualise the
user embeddings before and after applying our theme-based contrastive learning objective. Specifically,
we project the high-dimensional user embeddings onto a polar coordinate plot to enable qualitative
comparison, following the approach of Möller et al. [
As illustrated in Figure 2, the contrastive learning objective encourages the user embeddings to expand and become more geometrically structured. In the original embedding space (Figure 5.2), user embeddings are relatively concentrated and show limited separation. After contrastive learning (Figure 5.2), the embeddings are more widely dispersed. This geometric expansion suggests that contrastive learning promotes better separation among users. Figure 3 provides an additional t-SNE visualisation of user embeddings. After applying contrastive learning, the embeddings show clearer clustering structures and are more directionally aligned.
t-SNE of User Embeddings (Before CL) t-SNE of User Embeddings (After CL) 2 iEdm 0 N S -t 60 40 20 20 40 60
In addition to accuracy, we evaluate our model on fairness and diversity criteria to assess its ability to provide balanced exposure across categories, especially minority ones.
Fairness To evaluate the fairness of the recommendation lists, we adopt the fair-nDCG metric, a modification of the traditional nDCG that assigns relevance only to items belonging to predefined underrepresented or minority categories. In our experiments, we define the minority categories as: tv, entertainment, music, kids, movies, middleeast, games, weather, and autos.
As shown in Figure 4, we compare the fairness performance across top- recommendations between diferent model variants. The green line represents the ground-truth distribution from the test set, while the orange and blue lines correspond to models trained with contrastive learning (CL) based on theme-level and category-level grouping, respectively.
We observe that both contrastive variants outperform the baseline in terms of minority category coverage, especially as the value of increases. The curves for CL+Theme and CL+Category almost entirely overlap, indicating that both variants contribute similarly to improved fairness, despite using diferent semantic groupings during training. This trend is also reflected in Table 3, where the NDCG scores of these two variants are nearly identical. This consistency may suggest that the introduction of contrastive signals, rather than the specific grouping schema, is the primary factor driving improved fairness in recommendation ranking.
Diversity We quantify the diversity of the recommendations using statistical divergence measures. To evaluate the topical diversity of the recommendation lists, we compute the average Kullback–Leibler (KL) divergence and Jensen–Shannon (JS) divergence between the category distributions of the recommended items and a reference distribution, either derived from the user’s historical reading behaviour or the global news category distribution. Here, we consider the user’s reading history distribution as the reference one. We observe that both contrastive learning variants, whether based on category or theme, produce remarkably similar diversity outcomes. Specifically, the average KL divergence across all users is 2.2192, and the mean JS divergence is 0.4941 for both variants. Figure 5 illustrates the distribution of divergence scores alongside the user count. These findings suggest that, despite diferences in how user representations are constructed, the overall category diversity of the generated recommendation lists remains comparable.
However, when we restrict the evaluation to the top-5 or top-10 recommended items, the themebased variant shows better diversity. Table 4 presents a comparison between XNRS+CL(theme) and XNRS+CL(category) under Top-5 and Top-10 cutofs.
These results suggest that our theme-based methods lead to a more diverse set of recommended news articles, particularly in the top-ranked positions where user attention is highest. Although absolute diferences are relatively small, the consistent advantage of the theme-based variant at the top position indicates that higher semantic granularity in contrastive learning, such as grouping by theme rather than a single category, can encourage more varied and semantically rich recommendation lists.
To further understand the impact of diferent model components, we conduct an ablation study
comparing multiple architectural and training variations, including diferent user representation strategies (e.g.,
scoring function) and contrastive grouping granularity (e.g., themes vs. categories). We name these
variants following the setup from previous work [
First, we remove the MLP inside the user encoder (Equation 4), obtaining the base model. Then,
we replace the attention mechanisms in both the news and user encoders (Equations 2 and 4) with a
simple average pooling over the input embeddings, naming this variant Mean. For the final matching
layer between the user representation and candidate news items , we experiment with two types of
scoring functions:
• Dot Product (dot): This baseline computes the interaction score as the dot product between the
user and item embeddings:
= ⊤
(11)
Optionally, the embeddings can be ℓ2-normalised before computing the similarity. This is eficient
and widely used in many recommendation models.
• Bilinear (bilin): Inspired by prior work [
Our results are summarised in Table 5. As shown, we observe that base+dot+theme consistently
outperforms base+bilin+theme. This indicates that the dot-product scoring function is more efective
than the bilinear alternative in our contrastive learning setup. Among the model variants, the standard
configuration outperforms the mean variant across most metrics, consistent with prior findings [
We note that when contrastive learning is applied to the lightweight base model, its performance improves substantially, outperforming even NAML on several metrics. In the case of the base model, contrastive learning is applied directly to the user embeddings produced by the encoder, without additional architectural complexity. Surprisingly, this straightforward setup yields strong performance. However, the reasons behind its efectiveness remain unclear and warrant further investigation.
Our results demonstrate that incorporating contrastive learning into a bi-encoder news recommendation architecture improves user representation quality and overall recommendation accuracy. Both themebased and category-based grouping strategies outperform the base model, with particularly strong performance from the theme-based variant.
Beyond accuracy, we observe measurable gains in diversity and fairness metrics, especially at deeper ranks (e.g., top-30). However, these improvements are less pronounced at top-5 and top-10, where real-world user attention is typically concentrated. This discrepancy highlights a key challenge in designing recommendation models that balance fairness and diversity with real-world usability.
Another important observation is the strong performance of the lightweight base model when combined with contrastive learning. Despite its architectural simplicity, it competes closely with, and sometimes outperforms, more complex models like NAML. This suggests that contrastive learning can be a powerful technique even in low-complexity setups and motivates further exploration into contrastive signals and pairing strategies.
In this paper, we propose a contrastive learning framework for content-based news recommendation using a bi-encoder architecture.
Outlook Our method builds upon the XNRS model [
In addition to improvements in accuracy, our method contributes positively to diversity and fairness. While gains are more prominent at deeper ranks, the theme-based contrastive setup shows promise for encouraging a more balanced exposure of content.
Future Work One key limitation of the current model is its reliance solely on news titles for content encoding. Other textual fields, such as subtitles, abstracts, and metadata (e.g., keywords, subcategories), ofer rich semantic information that could improve representation learning. Incorporating these modalities may help capture finer-grained topical signals and disambiguate semantically similar content.
Future work should also address fairness in the top-ranked positions, where user engagement is highest. Although our method improves fairness at top-30 positions, the impact is limited at top5 or top-3 ranks. Fairness-aware re-ranking techniques or adaptive contrastive loss functions that prioritise exposure fairness at higher ranks may help bridge this gap. Another promising direction is the exploration of alternative contrastive learning strategies. These may include hierarchical or dynamic positive pair generation, temporal-aware contrastive signals, or dual user-item contrastive frameworks. Such extensions could further enhance both the performance and generalisability of contrastive learning in news recommendation systems.
This publication is part of the AI, Media & Democracy Lab. For more information about the lab and its further activities, visit https://www.aim4dem.nl/. The authors thank Bo-Chan Jack, Sven Lankester, and Natalie Halaskova for their valuable advice.
The authors declare that GenAI is only used to identify and correct grammatical errors, typos, and other writing mistakes. This helps improve the clarity and professionalism of the writing. [14] C. Wu, F. Wu, Y. Huang, X. Xie, Personalized news recommendation: Methods and challenges,
ACM Transactions on Information Systems 41 (2023) 1–50. [15] F. Wu, Y. Qiao, J.-H. Chen, C. Wu, T. Qi, J. Lian, D. Liu, X. Xie, J. Gao, W. Wu, et al., Mind: A large-scale dataset for news recommendation, in: Proceedings of the 58th annual meeting of the association for computational linguistics, 2020, pp. 3597–3606. [16] C. Feng, M. Khan, A. U. Rahman, A. Ahmad, News recommendation systems-accomplishments, challenges & future directions, IEEE Access 8 (2020) 16702–16725. [17] G. Adomavicius, A. Tuzhilin, Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE transactions on knowledge and data engineering 17 (2005) 734–749. [18] J. B. Schafer, D. Frankowski, J. Herlocker, S. Sen, Collaborative filtering recommender systems, in:
The adaptive web: methods and strategies of web personalization, Springer, 2007, pp. 291–324. [19] C. Wu, F. Wu, S. Ge, T. Qi, Y. Huang, X. Xie, Neural news recommendation with multi-head self-attention, in: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLPIJCNLP), 2019, pp. 6389–6394. [20] C. Wu, F. Wu, M. An, J. Huang, Y. Huang, X. Xie, Neural news recommendation with attentive multi-view learning, arXiv preprint arXiv:1907.05576 (2019). [21] M. An, F. Wu, C. Wu, K. Zhang, Z. Liu, X. Xie, Neural news recommendation with long-and short-term user representations, in: Proceedings of the 57th annual meeting of the association for computational linguistics, 2019, pp. 336–345. [22] C. Wu, F. Wu, M. An, J. Huang, Y. Huang, X. Xie, Npa: neural news recommendation with personalized attention, in: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 2576–2584. [23] R. Yamashita, M. Nishio, R. K. G. Do, K. Togashi, Convolutional neural networks: an overview and application in radiology, Insights into imaging 9 (2018) 611–629. [24] A. Graves, A. Graves, Long short-term memory, Supervised sequence labelling with recurrent neural networks (2012) 37–45. [25] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin,
Attention is all you need, Advances in neural information processing systems 30 (2017). [26] S. Ge, C. Wu, F. Wu, T. Qi, Y. Huang, Graph enhanced representation learning for news recommendation, in: Proceedings of the web conference 2020, 2020, pp. 2863–2869. [27] J. Giorgi, O. Nitski, B. Wang, G. Bader, DeCLUTR: Deep contrastive learning for unsupervised textual representations, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Association for Computational Linguistics, 2021, pp. 879–895. doi:10.18653/v1/2021.acl- long.72. [28] L. Wu, H. Lin, C. Tan, Z. Gao, S. Z. Li, Self-supervised learning on graphs: Contrastive, generative, or predictive, IEEE Transactions on Knowledge and Data Engineering (2021). [29] A. Iana, G. Glavaš, H. Paulheim, Train once, use flexibly: A modular framework for multi-aspect neural news recommendation, arXiv preprint arXiv:2307.16089 (2023). [30] X. Ren, W. Wei, L. Xia, C. Huang, A comprehensive survey on self-supervised learning for recommendation, ACM Computing Surveys (2024). [31] W. He, G. Sun, J. Lu, X. S. Fang, Candidate-aware graph contrastive learning for recommendation, in: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR, 2023, p. 1670–1679. doi:10.1145/3539618.3591647. [32] W. Jin, T. Derr, H. Liu, Y. Wang, S. Wang, Z. Liu, J. Tang, Self-supervised learning on graphs: Deep insights and new direction, arXiv preprint arXiv:2006.10141 (2020). [33] E. Pitoura, K. Stefanidis, G. Koutrika, Fairness in rankings and recommendations: an overview,
The VLDB Journal (2022) 1–28. [34] F. Calmon, D. Wei, B. Vinzamuri, K. Natesan Ramamurthy, K. R. Varshney, Optimized preprocessing for discrimination prevention, Advances in neural information processing systems 30 (2017). [35] L. Chen, L. Wu, K. Zhang, R. Hong, D. Lian, Z. Zhang, J. Zhou, M. Wang, Improving recommendation fairness via data augmentation, in: Proceedings of the ACM Web Conference 2023, 2023, pp. 1012–1020. [36] K. Balasubramanian, A. Alshabanah, E. Markowitz, G. Ver Steeg, M. Annavaram, Biased user history synthesis for personalized long-tail item recommendation, in: Proceedings of the 18th ACM Conference on Recommender Systems, 2024, pp. 189–199. [37] M. Slokom, S. Daniil, L. Hollink, How to diversify any personalized recommender?, in: Advances in Information Retrieval: 47th European Conference on Information Retrieval, ECIR 2025, Lucca, Italy, April 6–10, 2025, Proceedings, Part IV, 2025, p. 307–323. URL: https://doi.org/10.1007/ 978-3-031-88717-8_23. doi:10.1007/978-3-031-88717-8_23. [38] R. Burke, N. Sonboli, A. Ordonez-Gauger, Balanced neighborhoods for multi-sided fairness in recommendation, in: Conference on fairness, accountability and transparency, PMLR, 2018, pp. 202–214. [39] B. Yang, D. Liu, T. Suzumura, R. Dong, I. Li, Going beyond local: Global graph-enhanced personalized news recommendations, in: Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 24–34. [40] L. Boratto, F. Fabbri, G. Fenu, M. Marras, G. Medda, Fair augmentation for graph collaborative ifltering, in: Proceedings of the 18th ACM Conference on Recommender Systems, 2024, pp. 158–168. [41] M. Zehlike, F. Bonchi, C. Castillo, S. Hajian, M. Megahed, R. Baeza-Yates, Fa* ir: A fair top-k ranking algorithm, in: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, 2017, pp. 1569–1578. [42] S. C. Geyik, S. Ambler, K. Kenthapadi, Fairness-aware ranking in search & recommendation systems with application to linkedin talent search, in: Proceedings of the 25th acm sigkdd international conference on knowledge discovery & data mining, 2019, pp. 2221–2231. [43] M. Jiang, K. Bao, J. Zhang, W. Wang, Z. Yang, F. Feng, X. He, Item-side fairness of large language model-based recommendation system, in: Proceedings of the ACM on Web Conference 2024, 2024, pp. 4717–4726. [44] T. T. Nguyen, P.-M. Hui, F. M. Harper, L. Terveen, J. A. Konstan, Exploring the filter bubble: the efect of using recommender systems on content diversity, in: Proceedings of the 23rd ACM international conference on World wide web, 2014, pp. 677–686. [45] S. Vargas, P. Castells, Rank and relevance in novelty and diversity metrics for recommender systems, in: Proceedings of the 5th ACM conference on Recommender systems, 2011, pp. 109–116. [46] M. Kunaver, T. Požrl, Diversity in recommender systems–a survey, Knowledge-Based Systems 123 (2017) 154–162. [47] E. Bozdag, Bias in algorithmic filtering and personalization, Ethics and Information Technology 15 (2013) 209–227. [48] N. Helberger, K. Karppinen, L. D’Acunto, Exposure diversity as a design principle for recommender systems, Information, Communication & Society 21 (2018) 191–207. [49] C.-N. Ziegler, S. M. McNee, J. A. Konstan, G. Lausen, Improving recommendation lists through topic diversification, in: Proceedings of the 14th international conference on World Wide Web, 2005, pp. 22–32. [50] M. Zhang, N. Hurley, J. Peng, Avoiding monotony: improving the diversity of recommendation lists, Proceedings of the 2008 ACM conference on Recommender systems (2008) 123–130. [51] P. Castells, S. Vargas, J. Wang, Novelty and diversity in recommender systems, Recommender
Systems Handbook (2021) 845–884. [52] A. Iana, G. Glavas, H. Paulheim, Simplifying content-based neural news recommendation: On user modeling and training objectives, in: Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval, 2023, pp. 2384–2388. [53] T. Qi, F. Wu, C. Wu, Y. Huang, News recommendation with candidate-aware user modeling, in: Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 1917–1921. [54] K. Järvelin, J. Kekäläinen, Ir evaluation methods for retrieving highly relevant documents, in:
ACM SIGIR Forum, volume 51, 2017, pp. 243–250. [55] H. Steck, Calibrated recommendations, in: Proceedings of the 12th ACM Conference on Recommender Systems, RecSys ’18, 2018, p. 154–162. [56] E. Pitoura, K. Stefanidis, G. Koutrika, Fairness in rankings and recommendations: an overview,
The VLDB Journal 31 (2021) 431–458. [57] L. Möller, S. Padó, Understanding the relation of user and news representations in content-based neural news recommendation, The 10th International Workshop on News Recommendation and Analytics (INRA) (2022).