FCNHSMRA_HRS: Improve the performance of the movie hybrid recommender system using resource allocation approach

Mostafa Khalaji; Nilufar Mohammadnejad

REVIEW 4 major objections 5 minor 17 references

FCNHSMRA_HRS: Improve the performance of the movie hybrid recommender system using resource allocation approach

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Multiplying NHSM neighbor similarity by a resource-allocation reliability weight improves the fuzzy-clustered hybrid recommender's accuracy, precision, and recall on MovieLens, reducing average MAE from 0.7742 to 0.7666.

desk verdict A plausible incremental idea—NHSM similarity weighted by a resource-allocation term—but the reported gain is a small point estimate with no variance, so the central claim is not yet supported. read the letter →

arxiv 1908.05608 v1 pith:ICQW5MJD submitted 2019-08-13 cs.IR cs.SI

Mostafa Khalaji , Nilufar Mohammadnejad This is my paper · ORCID

classification cs.IRcs.SI

keywords RecommendersystemsCollaborativefilteringFuzzyclusteringHybridsystemResourceallocationSimilaritymeasuresMovieLensNHSM

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that a hybrid movie recommender can be made more accurate by treating neighbor similarity as a signal that needs a reliability adjustment. The proposed system, FCNHSMRA_HRS, combines fuzzy clustering with the NHSM similarity measure and multiplies each neighbor's similarity weight by a resource-allocation term that down-weights agreements on popular movies and up-weights agreements on rare ones. On the MovieLens 100K data, the system reports better mean absolute error, accuracy, precision, and recall than the FNHSM_HRS baseline and than fuzzy collaborative filtering with correlation, cosine, and the two neighbor-weighting baselines used in the comparison. The practical interest is that a simple popularity penalty might improve recommendation quality without changing the underlying prediction model.

What carries the argument

The load-bearing object is the resource-allocation reliability weight $R_{RA}(u,v)$, a sum over movies that two users both rated of $1/k_z$, where $k_z$ is the number of users who rated that movie. It is a popularity penalty: a shared rating of an obscure movie contributes more than a shared rating of a popular one. The prediction equation multiplies the NHSM similarity by $R_{RA}$ before summing neighbor deviations from their mean rating, so the weight reshapes which neighbors dominate the prediction. The fuzzy-clustering phase (three clusters, center-of-gravity defuzzification) is inherited from FNHSM_HRS and supplies the scalability claim, but the new mechanism is the RA multiplier.

What would settle it

Run the same five-fold evaluation with 3 clusters and 50 neighbors while replacing $R_{RA}$ with a constant 1 and with randomly permuted values across neighbor pairs; if the permuted version reproduces the reported improvement or the constant-1 version falls inside the confidence interval of FCNHSMRA_HRS, the resource-allocation term is not carrying the claimed signal.

Watch

Extended reading notes

Core claim

The paper claims that adding a resource-allocation reliability weight to the established FNHSM_HRS hybrid recommender materially improves its predictions. Neighbor selection still uses the NHSM heuristic similarity measure inside fuzzy clusters, but the prediction equation weights each neighbor's contribution by the product of NHSM similarity and the resource-allocation coefficient $R_{RA}(u,v)=\sum_{z \in \Gamma(u) \cap \Gamma(v)}1/k_z$, where $k_z$ is the number of users who rated the common movie $z$. Thus agreements on obscure movies are trusted more than agreements on blockbusters. Across Top-5 through Top-30 lists, the system reports average MAE 0.7666, accuracy 59.3578, precision 91.8406, and recall 52.575, beating FNHSM_HRS and the alternative similarity measures on every metric in the reported tables. The paper concludes that these results indicate improved performance and increased accuracy, while fuzzy clustering addresses scalability.

Load-bearing premise

The claimed improvement rests on the assumption that the resource-allocation weight supplies reliability information that the NHSM similarity measure does not already encode; if the two are redundant, the reported MAE gain of about 0.008 could easily be chance variation.

Editorial extensions

If this is right

Across Top-5 to Top-30 lists, FCNHSMRA_HRS reports lower MAE and higher accuracy, precision, and recall than FNHSM_HRS and than fuzzy CF with correlation, cosine, and the two neighbor-weighting baselines.
The gains are largest at Top-5, where accuracy rises from 60.203 to 64.146 and MAE falls from 0.756 to 0.741, so the reliability weight helps most when the recommendation list is short.
Because the RA term depends only on the degrees of co-rated movies in the rating matrix, it adds a per-pair sum to the existing NHSM computation rather than a new learned model.
Restricting neighbor search to the active user's fuzzy cluster preserves the scalability benefit of clustering while the RA weight operates inside that cluster.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct ablation, replacing $R_{RA}$ with a constant or with randomly permuted values, would settle whether the reported gains are real; the paper reports no significance testing, so the 0.008 MAE gap should be treated as provisional.
The same reliability weighting could plausibly be attached to correlation or cosine similarity; if the mechanism is general, it should improve those baselines by a similar margin, offering a cheap test outside NHSM.
An implicit consequence is that niche movies carry more weight in neighbor selection, which may improve recommendations for eclectic users but could reduce trust weights for users whose histories are dominated by mainstream hits.
Because $R_{RA}$ is exactly a link-prediction heuristic, the method translates directly to friend-suggestion and social-network recommendation tasks where item degree is defined by interaction counts.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The punchline: this is a modest incremental system—multiply the NHSM similarity by a resource-allocation weight inside a fuzzy-clustered collaborative filter—and the reported wins over the FNHSM_HRS baseline are small point estimates without any measure of uncertainty. The central claim that the RA term improves performance is not currently established.

What is genuinely new is the specific combination. I don't know of another paper that puts the Liben-Nowell/Kleinberg RA index into the NHSM neighbor-weighting step. The equations are laid out, and the evaluation uses MovieLens 100K with MAE, precision, recall, and accuracy over several Top-N settings. That is a workable starting point.

The soft spots are real and they are load-bearing. First, Tables 2 and 3 report single numbers from 5-fold CV with no standard deviations, confidence intervals, or paired tests. The average MAE difference is 0.0076 and accuracy difference is about 2 points; on folds of a 100K dataset, that can easily be fold noise. Second, the evaluation protocol has holes: "correct recommendation" in Eq. (13) is never defined, and the confusion-matrix labels in Table 1 are confusing, so precision/recall are hard to interpret. Third, the two free parameters, number of clusters and number of neighbors, are taken from earlier work without sensitivity analysis. Fourth, the main baselines, FNHSM_HRS and the neural-RA system from [5], are not public, so an independent run is impossible.

The stress-test note about RA being entangled with overlap is fair. Equation (10)'s sum over co-rated movies is closely related to the Jaccard-style overlap already in Eq. (2)/(3), and the popularity normalization inside the Singularity term arguably already does part of what RA does. The paper never disentangles whether the improvement (if real) comes from resource allocation or simply from giving more weight to high-overlap neighbors. That attack hits the same step as the missing statistics: the attribution of the gain to the introduced term.

The citation pattern is not a problem per se—the RA index is from the standard Liben-Nowell/Kleinberg paper, and self-citation to [3,5] is natural when the baselines are the authors' own—but the lack of public code and baselines makes verification hard.

Who gets value? Someone working on similarity measures for memory-based CF or on transferring link-prediction weights into recommenders. They would read it as a suggestion, not as a settled result. I would not cite it in the next twelve months; the reported numbers aren't trustworthy enough. For peer review, I would actually send it to referees rather than desk reject—there is a coherent idea and standard data—but I would expect a major revision that adds error bars or significance tests, fixes the evaluation protocol, and at least analyzes the redundancy between RA and the existing overlap terms.

Referee Report

4 major / 5 minor

Summary. The paper proposes FCNHSMRA_HRS, a hybrid movie recommender that extends the authors' earlier FNHSM_HRS system by multiplying the NHSM user-similarity weight with a resource-allocation (RA) reliability term, R_RA, inside the rating-prediction rule. The system first clusters users with fuzzy clustering, then computes NHSM similarity and RA reliability to select neighbors and predict ratings. The authors report experiments on MovieLens 100K using 5-fold cross-validation and compare against FNHSM_HRS and several collaborative-filtering baselines on MAE, Accuracy, Precision, and Recall. They conclude that the proposed system improves performance, with average MAE of 0.7666 versus 0.7742 for FNHSM_HRS and higher Accuracy (59.3578 versus 57.2976).

Significance. If the claimed improvement were statistically supported and the RA term shown to contribute independently of the existing similarity components, the contribution would be a modest incremental advance in memory-based collaborative filtering: adding a link-prediction-inspired reliability weight to a heuristic similarity measure is a plausible idea. The paper does use a standard public dataset (MovieLens 100K), reports several metrics across five Top-N settings, and describes the prediction rule explicitly. However, the evaluation is purely point-estimate based, the protocol for converting predicted ratings into 'correct recommendations' is not specified, and the comparison relies on non-public self-cited baselines. As it stands, the central claim that the RA term improves performance is not established beyond fold-to-fold noise.

major comments (4)

[Section 4, Tables 2 and 3] All results are reported as point estimates, with no standard deviations, confidence intervals, or significance tests, despite the statement that each fold was repeated 5 times. The key comparisons are small: MAE 0.7666 versus 0.7742 and Accuracy 59.3578 versus 57.2976 against FNHSM_HRS. With only 5 folds, these differences can easily lie within cross-validation variability. Since the central claim is precisely that the RA term improves performance, the authors should report per-fold results, variance, and a paired significance test (e.g., paired t-test or Wilcoxon signed-rank) across the folds, or otherwise demonstrate that the difference is not noise.
[Section 4, Equations (12)-(15)] The definitions of Accuracy, Precision, and Recall require a threshold for what counts as a 'correct recommendation' and what counts as a 'recommended item,' but the manuscript never states this threshold. The text says only that top-N movies are offered; it does not specify how a predicted rating is binarized into a positive/negative recommendation or what rating in the test set counts as 'useful.' Without this protocol, the reported Accuracy, Precision, and Recall values are not reproducible, and the comparison across methods is not well-defined. This is a load-bearing issue for the empirical claim.
[Section 3, Equations (10) and (11)] The resource-allocation term R_RA in Eq. (10) is an unnormalized sum over co-rated items of the inverse item degree (1/k_z). In Eq. (11) it is multiplied by NHSM_Sim, which already contains a Jaccard' overlap-count term (Eq. (3)) and a Singularity term (Eq. (7)) that depends on each item's mean rating. Both R_RA and these NHSM components are increasing functions of shared rated items and item popularity. The paper provides no analysis, ablation, or experiment isolating the contribution of R_RA from the existing terms. Therefore, even if the point-estimate improvement were significant, the manuscript does not establish that the improvement is caused by the resource-allocation mechanism rather than by re-weighting high-overlap neighbors.
[Section 4, baselines and reproducibility] The main baselines, FNHSM_HRS and the F_CF variants, are described only by references to the authors' own prior work [3] and [5], neither of which appears to be publicly available. The manuscript does not specify the exact parameter settings, preprocessing steps, or implementation details needed to reproduce these baselines. Since the empirical claim is a comparison against these systems, the lack of a public implementation or detailed algorithmic description prevents independent verification of the reported improvements.

minor comments (5)

[Section 3, Equation (6)] The equation for Significance is missing the equals sign before the expression 1/(1+exp(...)); as printed, it is not a valid equation.
[Section 3, Equations (7)-(9)] Notation is inconsistent: Eq. (8) uses mu_u and sigma_u, while Eq. (9) writes r_u-bar for the mean; the relation between these symbols should be clarified.
[Section 4, Table 1] The confusion-matrix table is confusingly labeled ('Positive Negative Actual / Predicted') and the placement of A, B, C, D is not aligned with the formulas in Eqs. (13)-(15); this should be redrawn for clarity.
[Section 5, Conclusion] The conclusion contains grammatical errors, e.g., 'The proposed systems is' and 'indicate improved system performance and increased accuracy is'; these should be corrected.
[Section 3, offline phase] The number of fuzzy clusters (3) and number of neighbors (50) are fixed based on prior work with no sensitivity analysis in this paper; a brief robustness check would strengthen the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RA-augmented NHSM prediction is empirically compared against stated baselines, and no equation reduces the conclusion to its inputs.

full rationale

The paper's prediction rule, Eq. (11), is a weighted average in which the NHSM similarity (Eq. 1) is multiplied by the resource-allocation reliability R_RA (Eq. 10). R_RA is taken from the external link-prediction literature [12], not derived from the target recommendation metrics, and the MAE/Accuracy/Precision/Recall values are computed on held-out MovieLens test data. The claimed improvement over FNHSM_HRS [3] is an empirical comparison, not a logical consequence of the definitions: nothing in Eqs. (1)-(11) defines the evaluation metrics in terms of R_RA or defines R_RA in terms of the evaluation outcomes. The use of the authors' own earlier FNHSM_HRS as a baseline and the citation of the authors' thesis [5] for the reliability idea are self-citations, but they are not load-bearing in a circular sense: the RA formula itself is attributed to Liben-Nowell and Kleinberg, and the baseline results are presented as experimental numbers rather than as assumptions that force the conclusion. The lack of significance testing and the small MAE gap (0.7666 vs 0.7742) are correctness/reproducibility concerns, not evidence that the derivation assumes its own conclusion. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported to forbid alternatives, and no known result is presented under new coordinates. The central claim therefore has independent empirical content, even though its statistical support is weak.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its central result rests on three existing techniques plus two hand-set parameters (cluster count and neighbor count) that are chosen from prior work with no sensitivity analysis.

free parameters (2)

Number of fuzzy clusters = 3
Selected based on reference [11]; no sensitivity analysis is provided. Introduced in Section 3.1 (Offline Phase).
Number of neighboring users m = 50
Stated in Section 4 with no justification; used in equation (11).

assumptions (4)

domain assumption Fuzzy c-means clustering accurately groups users with similar tastes.
Relies on Koohi and Kiani [11]; the paper does not validate cluster quality.
domain assumption NHSM similarity from Liu et al. [4] is a valid measure of user similarity.
Adopted wholesale from [4]; equations (1)-(9) are taken from that work.
domain assumption The resource allocation index of Liben-Nowell and Kleinberg [12] is a valid reliability signal for user pairs.
Equation (10) applies the RA link-prediction index to user-movie co-ratings without derivation of its relevance to rating prediction.
ad hoc to paper The weighted deviation prediction rule in equation (11) is an appropriate model.
The form of equation (11) is asserted without derivation; the weighting by the product NHSM * R_RA is the key modeling choice of the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FCNHSMRA_HRS: Improve the performance of the movie hybrid recommender system using resource allocation approach." pith.science (2026). https://pith.science/paper/ICQW5MJD

@misc{pith2026190805608,
  author       = {Pith},
  title        = {Pith review of: FCNHSMRA_HRS: Improve the performance of the movie hybrid recommender system using resource allocation approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICQW5MJD}},
  note         = {Machine review of arXiv:1908.05608}
}

read the original abstract

Recommender systems are systems that are capable of offering the most suitable services and products to users. Through specific methods and techniques, the recommender systems try to identify the most appropriate items, such as types of information and goods and propose the closest to the user's tastes. Collaborative filtering offering active user suggestions based on the rating of a set of users is one of the simplest and most comprehensible and successful models for finding people in the same tastes in the recommender systems. In this model, with increasing number of users and movie, the system is subject to scalability. On the other hand, it is important to improve the performance of the system when there is little information available on the ratings. In this paper, a movie hybrid recommender system based on FNHSM_HRS structure using resource allocation approach called FCNHSMRA_HRS is presented. The FNHSM_HRS structure was based on the heuristic similarity measure (NHSM), along with fuzzy clustering. Using the fuzzy clustering method in the proposed system improves the scalability problem and increases the accuracy of system suggestions. The proposed systems is based on collaborative filtering and, by using the heuristic similarity measure and applying the resource allocation approach, improves the performance, accuracy and precision of the system. The experimental results using MAE, Accuracy, Precision and Recall metrics based on MovieLens dataset show that the performance of the system is improved and the accuracy of recommendations in comparison of FNHSM_HRS and collaborative filtering methods that use other similarity measures for finding similarity, is increased

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[5]

Improvement of performance of hybrid recommender system using a neural network and resource allocation, Master's Thesis, K.N.Toosi University of Technology, Iran, Tehran, 2017

Khalaji, M. Improvement of performance of hybrid recommender system using a neural network and resource allocation, Master's Thesis, K.N.Toosi University of Technology, Iran, Tehran, 2017

work page 2017
[3]

FNHSM_HRS: Hybrid recommender system based on fuzzy clustering heuristic similarity measure, 7th Iranian Fuzzy and Intelligent Systems Congress, 2019

Khalaji, M, Dadkhah, C. FNHSM_HRS: Hybrid recommender system based on fuzzy clustering heuristic similarity measure, 7th Iranian Fuzzy and Intelligent Systems Congress, 2019

work page 2019
[1]

Recommender systems

Aggarwal C. Recommender systems. Springer International Publishing, 2016. International Conference on Researches in Science & Engineering th4 & International Congress on Civil, Architecture and Urbanism in Asia 19 March. 2019, Kasem Bundit University, Bangkok, Thailand

work page 2016
[2]

Cacheda F, Carneiro V, Fernández D, Formoso V. Comparison of collaborative filtering algorithms: Limitations of current techniques and proposals for scalable, high -performance recommender systems, ACM Transactions on the Web (TWEB), 5.1, 2011

work page 2011
[4]

A new user similarity model to improve the accuracy of collaborative filtering, Knowledge-Based Systems, 56, 156-166, 2014

Liu H, Hu Z, Mian A, Tian H, Z hu X. A new user similarity model to improve the accuracy of collaborative filtering, Knowledge-Based Systems, 56, 156-166, 2014

work page 2014
[6]

Using collaborative filtering to weave an information tapestry, Communications of the ACM, 35.12, 61-70, 1992

Goldberg D, Nichols D, Oki B, Terry D. Using collaborative filtering to weave an information tapestry, Communications of the ACM, 35.12, 61-70, 1992

work page 1992
[7]

A new similarity measure for collaborative filtering to alleviate the new user cold -starting Problem, Information Sciences, 178.1, 37-51, 2008

Ahn H. A new similarity measure for collaborative filtering to alleviate the new user cold -starting Problem, Information Sciences, 178.1, 37-51, 2008

work page 2008
[8]

Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge & Data Engineering, 6, 734- 749, 2005

Adomavicius G, Tuzhilin A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge & Data Engineering, 6, 734- 749, 2005

work page 2005

Show all 17 references

[9]

Bellogín A, Castells P, Cantador I. Improving memory -based collaborative filtering by neighbour selection based on user preference overlap, Proceedings of the 10th Conference on Open Research Areas in Information Retrieval, 2013

2013
[10]

A collaborative filtering a roach to mitigate the new user cold start problem, Knowledge-Based Systems, 26, 225-238, 2012

Bobadilla J, Ortega F, Hernando A, Bernal J. A collaborative filtering a roach to mitigate the new user cold start problem, Knowledge-Based Systems, 26, 225-238, 2012

2012
[11]

User based Collaborative Filtering using fuzzy C-means, Measurement, 91,134- 139, 2016

Koohi H, Kiani K. User based Collaborative Filtering using fuzzy C-means, Measurement, 91,134- 139, 2016

2016
[12]

The link ‐ prediction problem for social networks

Liben‐ Nowell D, Kleinberg J. The link ‐ prediction problem for social networks. journal of the Association for Information Science and Technology 58.7, 1019-1031, 2007

2007
[13]

Link prediction in complex networks: A survey

Lü L, Zhou T . Link prediction in complex networks: A survey. Physica A: statistical mechanics and its applications, 390(6), 1150-1170, 2011

2011
[14]

Suggesting friends using the implicit social graph

Roth M, Ben-David A, Deutscher D, Flysher G, Horn I, Leichtberg A, Leiser N, Matias Y, Merom, R. Suggesting friends using the implicit social graph. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 233-242, ACM, 2010

2010
[15]

Grocery shopping recommendations based on basket-sensitive random walk

Li M, Dias B.M, Jarman I, El -Deredy W, Lisboa P.J. Grocery shopping recommendations based on basket-sensitive random walk. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 1215-1224). ACM, 2009

2009
[16]

GroupLens: a lying collaborative filtering to Usenet news, Communications of the ACM, 40.3, 77-87, 1997

Konstan J, Miller B, Maltz D, Herlocker J, Gordon L, Riedl J. GroupLens: a lying collaborative filtering to Usenet news, Communications of the ACM, 40.3, 77-87, 1997

1997
[17]

Hahsler M. Developing and testing top -n recommendation algorithms for 0 -1 data usin g recommenderlab, NSF Industry University Cooperative Research Center for Net-Centric Software and System, 2011

2011

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [5]

Improvement of performance of hybrid recommender system using a neural network and resource allocation, Master's Thesis, K.N.Toosi University of Technology, Iran, Tehran, 2017

Khalaji, M. Improvement of performance of hybrid recommender system using a neural network and resource allocation, Master's Thesis, K.N.Toosi University of Technology, Iran, Tehran, 2017

work page 2017

[2] [3]

FNHSM_HRS: Hybrid recommender system based on fuzzy clustering heuristic similarity measure, 7th Iranian Fuzzy and Intelligent Systems Congress, 2019

Khalaji, M, Dadkhah, C. FNHSM_HRS: Hybrid recommender system based on fuzzy clustering heuristic similarity measure, 7th Iranian Fuzzy and Intelligent Systems Congress, 2019

work page 2019

[3] [1]

Recommender systems

Aggarwal C. Recommender systems. Springer International Publishing, 2016. International Conference on Researches in Science & Engineering th4 & International Congress on Civil, Architecture and Urbanism in Asia 19 March. 2019, Kasem Bundit University, Bangkok, Thailand

work page 2016

[4] [2]

Cacheda F, Carneiro V, Fernández D, Formoso V. Comparison of collaborative filtering algorithms: Limitations of current techniques and proposals for scalable, high -performance recommender systems, ACM Transactions on the Web (TWEB), 5.1, 2011

work page 2011

[5] [4]

A new user similarity model to improve the accuracy of collaborative filtering, Knowledge-Based Systems, 56, 156-166, 2014

Liu H, Hu Z, Mian A, Tian H, Z hu X. A new user similarity model to improve the accuracy of collaborative filtering, Knowledge-Based Systems, 56, 156-166, 2014

work page 2014

[6] [6]

Using collaborative filtering to weave an information tapestry, Communications of the ACM, 35.12, 61-70, 1992

Goldberg D, Nichols D, Oki B, Terry D. Using collaborative filtering to weave an information tapestry, Communications of the ACM, 35.12, 61-70, 1992

work page 1992

[7] [7]

A new similarity measure for collaborative filtering to alleviate the new user cold -starting Problem, Information Sciences, 178.1, 37-51, 2008

Ahn H. A new similarity measure for collaborative filtering to alleviate the new user cold -starting Problem, Information Sciences, 178.1, 37-51, 2008

work page 2008

[8] [8]

Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge & Data Engineering, 6, 734- 749, 2005

Adomavicius G, Tuzhilin A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge & Data Engineering, 6, 734- 749, 2005

work page 2005