CoChat
Sign Up
Cited by 0Recommender Systems and TechniquesMulti-Criteria Decision MakingInformation Retrieval and Search Behavior

Neuro-Fuzzy Aggregation for Multi-Criteria Recommender Systems: Comparative Study with Fixed-Weight Averaging

Open access
Singh, Reetu, Pinheiro, Greetta, Ajagalla, Mahadev, Patel, Tanya·2026·Procedia Computer Science
DOI 10.1016/j.procs.2026.06.453
Open this paper in CoChat — free accountOpen PDF

Abstract

Recommender systems have proved effective in handling information overload by providing personalized predictions. However, most of these systems are single-criterion recommenders that rely on a single rating. Multi-criteria recommender systems have gained popularity recently because they provide more effective recommendations than single-criterion models. This paper presents a comparative analysis of two aggregation techniques: standard fixed-weight averaging and a novel neuro-fuzzy aggregation framework. The proposed method combines fuzzy linguistic representations with a neural-network-based aggregator to model complex, nonlinear relationships among criteria. To generate criterion-level predictions, we use various collaborative filtering techniques, including User-based Collaborative Filtering (UserCF), Item-based Collaborative Filtering (ItemCF), Singular Value Decomposition (SVD), SVD++ (SVD with implicit-feedback extension), Probabilistic Matrix Factorization (PMF), and Deep Matrix Factorization (DeepMF). These predictions are then processed through both aggregation strategies. Evaluation on two benchmark datasets—TripAdvisor and Yahoo! Movies—demonstrates that the neuro-fuzzy method consistently achieves lower RMSE and MAE, underscoring its capability to enhance predictive performance in multi-criteria recommendation settings.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper compares fixed-weight averaging with a neuro-fuzzy aggregation framework for multi-criteria recommender systems, using six collaborative filtering models to predict per-criterion ratings before aggregation. Across TripAdvisor and Yahoo!Movies, the neuro-fuzzy method generally achieves lower RMSE and MAE, with the largest gains on TripAdvisor.

Key findings

  • On TripAdvisor, neuro-fuzzy aggregation improved RMSE and MAE for all six CF models. User-based CF dropped from RMSE 2.1223 to 1.6387 and MAE 1.7276 to 1.3702; Item-based CF dropped from RMSE 2.2634 to 1.6406 and MAE 1.7768 to 1.3594; DeepMF dropped from RMSE 2.1767 to 1.6910 and MAE 1.9104 to 1.4133.
  • The paper reports that User-based CF and Item-based CF showed the largest TripAdvisor gains, with RMSE improvements up to 0.62 and MAE reductions over 0.40.
  • On Yahoo!Movies, gains were less dramatic but neuro-fuzzy aggregation remained superior in most scenarios: SVD++ improved from RMSE 2.9205 to 2.7841 and MAE 2.2520 to 2.1468; PMF improved from RMSE 3.1864 to 2.9856 and MAE 2.6459 to 2.3351.
  • DeepMF was slightly worse on Yahoo!Movies under neuro-fuzzy aggregation: RMSE 3.1389 vs. 3.1369 fixed, and MAE 2.5149 vs. 2.4964 fixed.

Methods

  • Six collaborative filtering models—UserCF, ItemCF, SVD, SVD++, PMF, and DeepMF—were trained to predict per-criterion ratings for each user-item pair. Predicted criterion scores were min-max scaled to [0,1], converted into Low/Medium/High fuzzy membership degrees using triangular membership functions, and concatenated into a fuzzy feature vector.
  • A multilayer perceptron aggregator with two hidden layers (64 units, ReLU, dropout 0.2) was trained to predict the overall rating by minimizing MSE, with early stopping on validation RMSE.
  • Experiments used TripAdvisor (7,504 ratings, 7,313 users, 75 items, 4 criteria, 98.63% sparsity) and Yahoo!Movies (62,156 ratings, 6,078 users, 976 items, 4 criteria, 98.95% sparsity), with an 80/20 train-test split repeated five times and evaluation by RMSE and MAE.

Limitations

  • The paper does not include a dedicated limitations section.
  • Stated in the discussion: Yahoo!Movies gains were less dramatic due to greater scale and sparsity, and DeepMF was slightly worse there, which the authors suggest could indicate overfitting or diminished generalizability in the movie-based multi-criteria domain.
  • Evident from the text: the comparative evaluation is limited to two datasets and six CF models, and no statistical significance testing is reported for the observed differences.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.procs.2026.06.453

Cites

15 of 15 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

  • Findingreasoning

    On the TripAdvisor dataset, DeepMF exhibits a significant performance boost with neuro-fuzzy aggregation.

    “DeepMF also exhibits a significant performance boost, indicating that fuzzy inference complements deep latent feature learning.”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    On the Yahoo!Movies dataset, DeepMF is slightly worse with neuro-fuzzy aggregation.

    “DeepMF is slightly worse in this scenario, which could be an indication of overfitting or diminished generalizability in the movie-based multi-criteria domain.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The neuro-fuzzy method consistently achieves lower RMSE and MAE than fixed weighting across multiple models and datasets.

    “Evaluation on two benchmark datasets-TripAdvisor and Yahoo!Movies-demonstrates that the neuro-fuzzy method consistently achieves lower RMSE and MAE, underscoring its capability to enhance predictive performance in multi-criteria recommendation settings.”
    • Also stated in: § 1. Introduction · boosted
      “Experimental results verify that our suggested method consistently achieves lower RMSE and MAE than fixed weighting across multiple models and datasets, validating its effectiveness and generalizability.”
    • Also stated in: § 6. Conclusion and Future Scope · boosted
      “Large-scale experiments on two real-world datasets-TripAdvisor and Yahoo!Movies demonstrate that the introduced neuro-fuzzy aggregation uniformly improves RMSE and MAE over fixed-weight baselines for all models.”
    Quote in the paperSubstantial claim
  • Findingown data

    On the Yahoo!Movies dataset, fuzzy aggregation remains superior in most scenarios.

    “While gains are less dramatic here-due to greater scale and sparsity-fuzzy aggregation remains superior in most scenarios.”
    Quote in the paperSubstantial claim
  • Findingown data

    Across both datasets and models, the neuro-fuzzy aggregation approach matches or outperforms fixed-weighting performance.

    “Across both data sets and models, the neuro-fuzzy aggregation approach matches or outperforms fixed-weighting performance.”
    Quote in the paperSubstantial claim
  • May combine two claims. 5
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 5

  • May combine two claims.

    Findingown data

    On the Yahoo!Movies dataset, SVD++, PMF, and Item-based CF perform better with MAE reduction under neuro-fuzzy aggregation.

    “SVD++, PMF, and Item-based CF perform better with MAE reduction.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    On the TripAdvisor dataset, the neuro-fuzzy model shows consistent improvements in RMSE and MAE for all six collaborative filtering models.

    “TripAdvisor dataset: The neuro-fuzzy model shows consistent improvements in RMSE and MAE for all models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On the TripAdvisor dataset, User-based CF and Item-based CF exhibit the largest gains, with RMSE improvements of up to 0.62 and MAE reductions over 0.40.

    “Notably, User-based CF and Item-based CF exhibit the largest gains, with RMSE improvements of up to 0.62 and MAE reductions over 0.40.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The study ensured that every user and item appearing in the test set appears at least once in the training set to avoid cold-start leakage.

    “We also ensure that every user and item appearing in the test set appears at least once in the training set to avoid cold-start leakage.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Training settings were kept the same across datasets unless noted.

    “Training settings are kept the same across datasets unless noted.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The study repeated the 80/20 train-test split five times with different random splits to reduce the effect of randomness.

    “To reduce the effect of randomness, we repeat the 80/20 split five times with different random splits.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    The proposed neuro-fuzzy aggregation method is model-agnostic and can be inserted into a variety of collaborative filtering models, including UserCF, ItemCF, SVD, SVD++, PMF, and DeepMF.

    “In contrast with current methods, the technique is model-agnostic and can be inserted into a variety of collaborative filtering models, including User-based CF, Item-based CF, SVD, SVD++, PMF, and DeepMF.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The neuro-fuzzy solution has strong robustness and generality.

    “validating the strong robustness and generality of the solution.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Early stopping on validation RMSE with patience of 10 epochs was used.

    “We use Adam as the optimizer and early stopping on validation RMSE with patience of 10 epochs, with the main hyperparameters listed in Table 3.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The neuro-fuzzy aggregation approach enhances the accuracy of overall-rating prediction while offering an interpretable and adaptable alternative to traditional aggregation techniques.

    “The findings demonstrate that the neuro-fuzzy aggregation approach enhances the accuracy of overall-rating prediction while offering an interpretable and adaptable alternative to traditional aggregation techniques.”
    Quote in the paperSubstantial claim
10 of 10

  • The paper relies on this without support.

    Fixed-weight averaging assumes linearity and fixed criterion importance, which may not reflect real user behavior.

    “While simple and efficient, this assumes linearity and fixed criterion importance, which may not reflect real user behavior.”

  • The paper relies on this without support.

    The key problem in MCRS is how to merge several criterion ratings into one overall score for ranking; the choice of aggregation strongly influences accuracy and transparency.

    “The key problem in MCRS is how to merge several criterion ratings into one overall score for ranking; the choice of aggregation strongly influences accuracy and transparency.”

  • The paper relies on this without support.

    The majority of methods adopt static aggregation and use hand-crafted or fixed weights to merge criterion-level ratings.

    “Firstly, the majority of methods adopt static aggregation and use hand-crafted or fixed weights to merge criterion-level ratings.”

  • The paper relies on this without support.

    A majority of models using deep learning or evolutionary optimization is inadequate in interpretability.

    “Secondly, a majority of models using deep learning or evolutionary optimization is inadequate in interpretability.”

  • The paper relies on this without support.

    Many aggregation techniques are strongly linked to specific model architectures, limiting generalizability and flexibility when aggregating with larger families of recommendation algorithms.

    “Lastly, many aggregation techniques are strongly linked to specific model architectures, limiting generalizability and flexibility when aggregating with larger families of recommendation algorithms.”

  • The paper relies on this without support.

    Multi-criteria recommender systems provide more effective recommendations than single-criterion models.

    “Multi-criteria recommender systems have gained popularity recently because they provide more effective recommendations than single-criterion models.”
6 of 6

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Claim of uniform improvement conflicts with Yahoo!Movies results

    Major

    The conclusion says neuro-fuzzy aggregation improves both metrics for every model, but Yahoo!Movies DeepMF has a higher RMSE and MAE with neuro-fuzzy aggregation than with fixed weighting.

  • Item-based CF does not have a Yahoo!Movies MAE reduction

    Minor

    The Yahoo!Movies table shows Item-based CF MAE increasing from 2.1642 with fixed weighting to 2.1810 with neuro-fuzzy aggregation.

  • Low fuzzy set violates the stated triangle definition

    Minor

    The definition requires a < b < c, but the specified Low triangle has a = b = 0.0.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Zhai, 2020, A survey on fuzzy recommender systems, Information Sciences, 509, 255

Related

Authors
Singh, R., Pinheiro, G., Ajagalla, M., & Patel, T.
Affiliations
Bennett University (IN)
Corresponding author
  • Singh, Reetu
  • Pinheiro, Greetta
  • Ajagalla, Mahadev
  • Patel, Tanya
ORCID
Ajagalla, Mahadev
Venue
Procedia Computer Science
Publisher
Elsevier BV
Published
2026
Volume
283
Pages
3925-3933
ISSN
1877-0509 (ISSN-L)
DOI
10.1016/j.procs.2026.06.453
OpenAlex
W7167732030
Cited by
0
References
10
FWCI
0.00
Type
conference paper
Language
en
Open access
Licence
Locations
Procedia Computer Science (publishedVersion, cc-by-nc-nd)
Topics
  • Physical Sciences › Computer Science › Information Systems › Recommender Systems and Techniques
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Multi-Criteria Decision Making
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
Keywords
multi-criteria recommender systems, collaborative filtering, implicit feedback, probabilistic matrix factorization, deep matrix factorization, SVD, user-based collaborative filtering, item-based collaborative filtering, personalized recommendation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-29.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-29.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review