pith. machine review for the scientific record. sign in

arxiv: 2604.15573 · v1 · submitted 2026-04-16 · 💻 cs.IR · cs.LG

Recognition: unknown

Collaborative Filtering Through Weighted Similarities of User and Item Embeddings

Pedro R. Pires , Rafael T. Sereicikas , Gregorio F. Azevedo , Tiago A. Almeida
Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:40 UTC · model grok-4.3

classification 💻 cs.IR cs.LG
keywords collaborative filteringmatrix factorizationuser-item similarityitem-item similarityweighted ensembletop-N recommendationsshared embeddingshybrid recommender
0
0 comments X

The pith

Shared embeddings from matrix factorization let a weighted mix of user-item and item-item similarities deliver competitive top-N recommendations without extra tuning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how a single set of user and item embeddings learned by ordinary matrix factorization can be reused to compute both user-to-item and item-to-item similarities, then combined with weights to generate top-N lists. This unifies two classic recommendation strategies into one lightweight ensemble that reuses the base algorithm's hyperparameters and skips any embedding fine-tuning. The resulting method stays competitive on standard datasets and holds up whether the data favors user-centric or item-centric patterns. A reader would care because it revives simple, efficient collaborative filtering as a practical alternative to heavy neural models.

Core claim

A single matrix-factorization embedding space can supply both user-item and item-item similarities; weighting those two similarity matrices produces top-N recommendations that match or approach more complex models, remain stable across user-item versus item-item dominated regimes, and require no additional embedding training or hyperparameter search beyond what the base algorithm already used.

What carries the argument

Weighted similarity framework that fuses user-item and item-item scores computed from the same embedding vectors.

If this is right

  • Existing matrix-factorization code and hyperparameters transfer directly to the hybrid without retraining.
  • Model size and inference cost stay close to a single matrix-factorization run rather than growing with separate user and item models.
  • The ensemble remains effective even on data where one of the two similarity types is clearly dominant.
  • Implementation reduces to computing two similarity matrices from the same embeddings and a single weighted sum.
  • No embedding-specific validation set is needed once the base algorithm has been trained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same shared-embedding idea could be tested as a lightweight baseline against newer neural recommenders to measure how much complexity is truly required.
  • If the optimal weight between the two similarities can be estimated from a small held-out set, the method would become even easier to deploy on new domains.
  • Applying the weighting trick to other embedding sources such as autoencoders or graph embeddings would check whether the benefit is specific to matrix factorization.

Load-bearing premise

Embeddings learned once by matrix factorization already contain enough structure for both user-item and item-item similarities to be useful when simply added with weights.

What would settle it

On a dataset where the best-performing weight between the two similarity types produces lists noticeably worse than either pure user-item or pure item-item lists, or where performance collapses unless the embeddings are re-tuned, the claim would fail.

Figures

Figures reproduced from arXiv: 2604.15573 by Gregorio F. Azevedo, Pedro R. Pires, Rafael T. Sereicikas, Tiago A. Almeida.

Figure 1
Figure 1. Figure 1: Visual representation of using the weighted simi [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework of the recommendation through weighted similarities of user and item embeddings. 3.1 Embedding Generation The first step generates user and item embeddings via an external embedding model. We evaluated several models, including alter￾nating least squares (ALS) [14], Bayesian personalized ranking (BPR) [27], and RecVAE, a state-of-the-art variational autoencoder recommender [31]. These models are … view at source ↗
Figure 3
Figure 3. Figure 3: NDCG@𝑁 with 𝑁 ranging from 1 to 20. Recommender UI corresponds to user–item, II to item–item, and WS to the weighted similarities ensemble. For ease of comparison, each embedding-based algorithm follows a similar color palette [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual representation of the Nemenyi test con [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

In recent years, neural networks and other complex models have dominated recommender systems, often setting new benchmarks for state-of-the-art performance. Yet, despite these advancements, award-winning research has demonstrated that traditional matrix factorization methods can remain competitive, offering simplicity and reduced computational overhead. Hybrid models, which combine matrix factorization with newer techniques, are increasingly employed to harness the strengths of multiple approaches. This paper proposes a novel ensemble method that unifies user-item and item-item recommendations through a weighted similarity framework to deliver top-N recommendations. Our approach is distinctive in its use of shared user and item embeddings for both recommendation strategies, simplifying the architecture and enhancing computational efficiency. Extensive experiments across multiple datasets show that our method achieves competitive performance and is robust in varying scenarios that favor either user-item or item-item recommendations. Additionally, by eliminating the need for embedding-specific fine-tuning, our model allows for the seamless reuse of hyperparameters from the base algorithm without sacrificing performance. This results in a method that is both efficient and easy to implement. Our open-source implementation is available at https://github.com/UFSCar-LaSID/weighted-sims-recommender.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a hybrid collaborative filtering approach for top-N recommendations that combines user-item and item-item strategies via a weighted similarity framework applied to shared user and item embeddings learned from a base matrix factorization model. The method is presented as distinctive for reusing the same embeddings for both recommendation types, thereby simplifying the architecture and improving efficiency. The authors claim that extensive experiments on multiple datasets demonstrate competitive performance, robustness across scenarios favoring either user-item or item-item approaches, and the ability to reuse hyperparameters from the base MF algorithm without any embedding-specific fine-tuning. An open-source implementation is provided.

Significance. If the empirical results hold and substantiate the robustness and performance claims without per-embedding adjustments, this would constitute a modest but practical contribution by showing how traditional MF embeddings can be directly repurposed in a simple ensemble to bridge user-item prediction with similarity-based retrieval. The emphasis on computational efficiency and hyperparameter reuse aligns with efforts to keep recommender systems lightweight, and the open-source code supports reproducibility.

major comments (2)
  1. [Experiments / Results section] The central empirical claim (that a single set of MF embeddings trained only for user-item reconstruction can be directly reused to compute item-item similarities whose weighted combination yields competitive and robust top-N results) is load-bearing but under-supported. The MF objective minimizes user-item prediction error and does not constrain the geometry of the item embedding space for nearest-neighbor similarity retrieval; nothing guarantees that cosine (or other) similarities on item vectors will capture useful collaborative signals. The manuscript must include ablation results (e.g., performance of pure user-item, pure item-item, and weighted versions) and explicit reporting of how the similarity weighting factor is selected or tuned across datasets to substantiate the 'no embedding-specific fine-tuning' and 'reuse hyperparameters' assertions.
  2. [Abstract and Results] The abstract asserts 'competitive performance' and 'robustness in varying scenarios' yet supplies no quantitative metrics, baseline comparisons (e.g., against standard MF, item-item CF, or neural hybrids), statistical significance tests, or details on the weighting procedure. Without these in the results tables or text, the soundness of the central claim cannot be evaluated.
minor comments (2)
  1. [Abstract] The abstract refers to 'award-winning research' showing MF competitiveness; adding a specific citation would clarify the reference.
  2. [Method] Notation for the weighting factor and similarity functions (user-item dot product vs. item-item cosine) should be introduced with explicit equations early in the method section to improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on strengthening the empirical support for our claims. We address each major point below and have revised the manuscript to incorporate additional ablation studies and clarifications on the weighting procedure.

read point-by-point responses

  1. Referee: [Experiments / Results section] The central empirical claim (that a single set of MF embeddings trained only for user-item reconstruction can be directly reused to compute item-item similarities whose weighted combination yields competitive and robust top-N results) is load-bearing but under-supported. The MF objective minimizes user-item prediction error and does not constrain the geometry of the item embedding space for nearest-neighbor similarity retrieval; nothing guarantees that cosine (or other) similarities on item vectors will capture useful collaborative signals. The manuscript must include ablation results (e.g., performance of pure user-item, pure item-item, and weighted versions) and explicit reporting of how the similarity weighting factor is selected or tuned across datasets to substantiate the 'no embedding-specific fine-tuning' and 'reuse hyperparameters' assertions.

    Authors: We agree that explicit ablations are necessary to substantiate the empirical claims, as the approach relies on observed performance rather than a theoretical guarantee on embedding geometry. In the revised manuscript, we have added a dedicated ablation subsection in the Experiments section. This includes side-by-side results for (i) pure user-item MF predictions, (ii) pure item-item nearest-neighbor retrieval using cosine similarity on the same embeddings, and (iii) the weighted ensemble, evaluated on all datasets with Recall@10 and NDCG@10. The results show the weighted version is competitive or superior in scenarios favoring either component. For the weighting factor, we have added explicit reporting in Section 4.2: the scalar weight w is selected via a grid search over {0.0, 0.1, ..., 1.0} on a validation split, using the identical hyperparameter search budget and procedure as the base MF model. No further embedding optimization or fine-tuning occurs, directly supporting the reuse claim. This procedure is applied uniformly across datasets. revision: yes

  2. Referee: [Abstract and Results] The abstract asserts 'competitive performance' and 'robustness in varying scenarios' yet supplies no quantitative metrics, baseline comparisons (e.g., against standard MF, item-item CF, or neural hybrids), statistical significance tests, or details on the weighting procedure. Without these in the results tables or text, the soundness of the central claim cannot be evaluated.

    Authors: The Results section already contains the requested quantitative elements: tables report Recall@10 and NDCG@10 for our method versus standard MF, item-item CF, and neural baselines, with statistical significance assessed via paired t-tests (p < 0.05) over five random seeds. The weighting procedure is now detailed in the Experimental Setup as described above. To improve accessibility, we have added one sentence to the abstract summarizing that the method matches or exceeds baselines on multiple datasets without embedding-specific tuning. We retain the high-level tone conventional for abstracts while ensuring all supporting metrics and procedures appear in the main text and tables. revision: partial

Circularity Check

0 steps flagged

No circularity: embeddings treated as external inputs; weighting is an added post-processing step

full rationale

The paper takes embeddings from a base matrix factorization algorithm as given inputs and proposes computing user-item scores via dot products and item-item similarities (e.g., cosine) on the same vectors, then forming a weighted combination for top-N lists. No equation or step defines the embeddings in terms of the final weighted output or vice versa; the weighting parameter is introduced as a tunable hyperparameter rather than derived by construction from the MF objective. Experiments on multiple datasets serve as external validation rather than tautological confirmation. No self-citation chain, uniqueness theorem, or ansatz smuggling is present in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that matrix-factorization embeddings already encode useful similarities for both users and items, plus the existence of a weighting parameter whose value is not derived from first principles.

free parameters (1)
  • similarity weighting factor
    A scalar that balances the contribution of user-item versus item-item recommendations; its selection or fitting procedure is not specified in the abstract.
axioms (1)
  • domain assumption Embeddings produced by a standard matrix-factorization algorithm capture both user-user and item-item similarities sufficiently well for top-N ranking.
    Invoked when the paper states that the same embeddings can be reused for both recommendation strategies without further adaptation.

pith-pipeline@v0.9.0 · 5509 in / 1351 out tokens · 34498 ms · 2026-05-10T09:40:03.459918+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    Mohammed Fadhel Aljunid and Manjaiah Doddaghatta Huchaiah. 2021. An efficient hybrid recommendation model based on collaborative filtering recom- mender systems.CAAI Transactions on Intelligence Technology6, 4 (2021), 480–492. doi:10.1049/cit2.12048

    work page doi:10.1049/cit2.12048 2021

  2. [2]

    Oren Barkan and Noam Koenigstein. 2016. Item2Vec: Neural Item Embedding For Collaborative Filtering. InIEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP 2016). IEEE, Vietri sul Mare, Italy, 1–6. doi:10.1109/MLSP.2016.7738886

    work page doi:10.1109/mlsp.2016.7738886 2016

  3. [3]

    Ballester, X

    J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez. 2013. Recommender systems survey.Knowledge-Based Systems46 (2013), 109–132. doi:10.1016/j. knosys.2013.03.012

    work page doi:10.1016/j 2013

  4. [4]

    Maurizio Ferrari Dacrema, Paolo Cremonesi, and Dietmar Jannach. 2019. Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Rec- ommendation Approaches. InProceedings of the 13th ACM Conference on Recom- mender Systems (RecSys‘19). Association for Computing Machinery, Copenhagen, Denmark, 101–109. doi:10.1145/3298689.3347058

    work page doi:10.1145/3298689.3347058 2019

  5. [5]

    Janez Demšar. 2006. Statistical Comparisons of Classifiers over Multiple Data Sets. The Journal of Machine Learning Research7 (2006), 1–30. doi:10.5555/1248547. 1248548

    work page doi:10.5555/1248547 2006

  6. [6]

    Mukund Deshpande and George Karypis. 2004. Item-based top-N recommenda- tion algorithms.ACM Transactions on Information Systems22, 1 (2004), 143–177. doi:10.1145/963770.963776

    work page doi:10.1145/963770.963776 2004

  7. [7]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2020. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.arXiv 2010.11929 (2020), 1–22. doi:10.48550/arXiv.2010.11929

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2020

  8. [8]

    Travis Ebesu, Bin Shen, and Yi Fang. 2018. Collaborative Memory Network for Recommendation Systems. InThe 41st International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR‘18). Association for Computing Machinery, Ann Arbor, MI, USA, 515–524. doi:10.1145/3209978. 3209991

    work page doi:10.1145/3209978 2018

  9. [9]

    Simon Funk. 2016. Netflix Update: Try This at Home. https://sifter.org/simon/ journal/20061211.html

    work page arXiv 2016

  10. [10]

    Geetha, M

    G. Geetha, M. Safa, C. Fancy, and D. Saranya. 2018. A Hybrid Approach using Collaborative filtering and Content based Filtering for Recommender System. InProceedings of the National Conference on Mathematical Techniques and its Applications (NCMTA 18). IOP Publishing, Kattankulathur, India, 1–7. doi:10. 1088/1742-6596/1000/1/012101

    2018

  11. [11]

    Blake Hallinan and Ted Striphas. 2016. Recommended for you: The Netflix Prize and the production of algorithmic culture.New Media and Society18, 1 (2016), 117–137. doi:10.1177/1461444814538646

    work page doi:10.1177/1461444814538646 2016

  12. [12]

    Konstan, and John Riedl

    Jon Herlocker, Joseph A. Konstan, and John Riedl. 2002. An Empirical Analysis of Design Choices in Neighborhood-Based Collaborative Filtering Algorithms. Information Retrieval5 (2002), 287–310. doi:10.1023/A:1020443909834

    work page doi:10.1023/a:1020443909834 2002

  13. [13]

    Binbin Hu, Chuan Shi, Wayne Xin Zhao, and Philip S. Yu. 2018. Leveraging Meta- path based Context for Top- N Recommendation with A Neural Co-Attention Model. InProceedings of the 24th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining (KDD‘18). Association for Computing Machinery, London, UK, 1531–1540. doi:10.1145/3219819.3219965

    work page doi:10.1145/3219819.3219965 2018

  14. [14]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative Filtering for Implicit Feedback Datasets. InProceedings of the 8th IEEE International Conference on Data Mining (ICDM ‘08). IEEE Computer Society, Pisa, Italy, 263–272. doi:10. 1109/ICDM.2008.22

    2008

  15. [15]

    Shan Khsuro, Zafar Ali, and Irfan Ullah. 2016. Recommender Systems: Issues, Challenges, and Research Opportunities. InProceedings of the 7th International Conference on Information Science and Applications (ICISA 2016). Springer Sci- ence+Business Media, Ho Chi Minh, Vietnam, 1179–1189. doi:10.1007/978-981- 10-0557-2_112

    work page doi:10.1007/978-981- 2016

  16. [16]

    Yehuda Koren. 2009. Collaborative filtering with temporal dynamics. InProceed- ings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining (KDD ‘09). Association for Computing Machinery, Paris, France, 447–456

    2009

  17. [17]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Tech- niques For Recommender Systems.Computer42, 8 (2009), 30–37. doi:10.1109/ MC.2009.263

    2009

  18. [18]

    Raúl Lara-Cabrera, Ángel GonzÁlez-Prieto, and Fernando Ortega. 2020. Deep Matrix Factorization Approach for Collaborative Filtering Recommender Systems. Applied Sciences10, 14 (2020), 1–14. doi:10.3390/app10144926

    work page doi:10.3390/app10144926 2020

  19. [19]

    Xiaopeng Li and James She. 2017. Collaborative Variational Autoencoder for Recommender Systems. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD‘17). Association for Computing Machinery, Halifax, Canada, 305–314. doi:10.1145/3097983.3098077

    work page doi:10.1145/3097983.3098077 2017

  20. [20]

    Pasquale Lops, Elio Musacchio, Cataldo Musto, Marco Polignano, Antonio Silletti, and Giovanni Semeraro. 2023. Reproducibility Analysis of Recommender Systems relying on Visual Features: traps, pitfalls, and countermeasures. InProceedings of the 17th ACM Conference on Recommender Systems (RecSys‘23). Association for Computing Machinery, Singapore, Singapor...

    work page doi:10.1145/3604915 2023

  21. [21]

    Xin Luo, Mengchu Zhou, Yunni Xia, and Qingsheng Zhu. 2014. An Efficient Non-Negative Matrix-Factorization-Based Approach to Collaborative Filtering for Recommender Systems.IEEE Transactions on Industrial Informatics10, 2 (2014), 1273–1284. doi:10.1109/TII.2014.2308433

    work page doi:10.1109/tii.2014.2308433 2014

  22. [22]

    Pawel Matuszyk, Jo ao Vinagre, Myra Spiliopoulou, Alípio Mário Jorge, and Jo ao Gama. 2015. Forgetting Methods for Incremental Matrix Factorization in Recommender Systems. InProceedings of the 30th Annual ACM Symposium on Applied Computing (SAC ‘15). Association for Computing Machinery, Salamanca, Spain, 947–953. doi:10.1145/2695664.2695820

    work page doi:10.1145/2695664.2695820 2015

  23. [23]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Conrado, and Jeffrey Dan. 2013. Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of the 26th International Conference on Neural Information Processing Systems (NIPS 2013). Curran Associates Inc., Stateline, NV, USA, 3111–3119. doi:10. 5555/2999792.2999959

    work page arXiv 2013

  24. [24]

    Govindarajan Parthasarathy and Shanmugam Sathiya Devi. 2022. Hybrid Rec- ommendation System Based on Collaborative and Content-Based Filtering.Cy- bernetics and Systems54, 4 (2022), 432–453. doi:10.1080/01969722.2022.2062544

    work page doi:10.1080/01969722.2022.2062544 2022

  25. [25]

    Pires and Tiago A

    Pedro R. Pires and Tiago A. Almeida. 2022. Interact2Vec: Neural Item and User Embedding for Collaborative Filtering. InProceedings of the 11th Brazilian Conference on Intelligent Systems (BRACIS 2022). Springer Nature, Campinas, Brazil, 494–509. doi:10.1007/978-3-031-21689-3_35

    work page doi:10.1007/978-3-031-21689-3_35 2022

  26. [26]

    Stefen Rendle. 2010. Factorization Machines. InProceedings of the 10th IEEE International Conference on Data Mining (ICDM 2010). IEEE, Sydney, Australia, 14–17. doi:10.1109/ICDM.2010.127

    work page doi:10.1109/icdm.2010.127 2010

  27. [27]

    Stefen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  28. [28]

    InProceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI ‘09)

    BPR: Bayesian Personalized Ranking from Implicit Feedback. InProceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI ‘09). AUAI Press, Montreal, Canada, 452–461. doi:10.5555/1795114.1795167

    work page doi:10.5555/1795114.1795167

  29. [29]

    Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl. 1994. GroupLens: An Open Architecture for Collaborative Filtering of Netnews. InProceedings of ACM 1994 Conference on Computer Supported Coopera- tive Work (CSCW‘94). Association for Computing Machinery, Chapel Hill, North Carolina, USA, 175–186. doi:10.1145/192844.192905

    work page doi:10.1145/192844.192905 1994

  30. [30]

    Francesco Ricci, Lior Rokach, and Bracha Shapira. 2011. Introduction to Recom- mender Systems Handbook. InRecommender Systems Handbook, Francesco Ricci, Lior Rokach, Bracha Shapira, and Paul B. Kantor (Eds.). Springer US, New York, NY, USA, Chapter 1, 1–38. doi:10.1007/978-0-387-85820-3

    work page doi:10.1007/978-0-387-85820-3 2011

  31. [31]

    Ruslan Salakhutdinov and Andriy Mnih. 2007. Probabilistic Matrix Factorization. InProceedings of the 20th International Conference on Neural Information Processing Systems (NIPS’07). Curran Associates Inc., Vancouver, Canada, 1257–1264

    2007

  32. [32]

    Nikolenko

    Ilya Shenbin, Anton Alekseev, Elena Tutubalina, Valentin Malykh, and Sergey I. Nikolenko. 2020. RecVAE: A New Variational Autoencoder for Top-N Recommen- dations with Implicit Feedback. InProceedings of the 13th International Conference on Web Search and Data Mining(Houston, TX, USA)(WSDM ’20). Association for Computing Machinery, New York, NY, USA, 528–5...

    work page doi:10.1145/3336191 2020

  33. [33]

    Suriati, M

    S. Suriati, M. Dwiastuti, and T. Tulus. 2017. Weighted hybrid technique for recom- mender system. InProceedings of the International Conference on Information and Communication Technology (IconICT 2017). IOP Publishing, Medan, Indonesia, 1–7. doi:10.1088/1742-6596/930/1/012050

    work page doi:10.1088/1742-6596/930/1/012050 2017

  34. [34]

    Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep Learning Based Recommender System: A Survey and New Perspectives.Comput. Surveys52, 1 (2019), 1–35. doi:10.1145/3285029

    work page doi:10.1145/3285029 2019

  35. [35]

    Zhuangzhuang Zhang and Yunquan Dong. 2022. Weighted Similarity and Core- User-Core-Item Based Recommendations.Entropy24, 5 (2022), 1–28. doi:10.3390/ e24050609

    2022

  36. [36]

    Erion Çano and Maurizio Morisio. 2017. Hybrid Recommender Systems: A Systematic Literature Review.Intelligent Data Analysis21, 6 (2017), 1487–1524. doi:10.3233/IDA-163209

    work page doi:10.3233/ida-163209 2017