pith. sign in

arxiv: 1907.10416 · v1 · pith:GOHBIHRLnew · submitted 2019-07-24 · 💻 cs.SI · cs.LG· physics.soc-ph

Semi-Supervised Tensor Factorization for Node Classification in Complex Social Networks

Georgios Katsimpras , Georgios Paliouras This is my paper

Pith reviewed 2026-05-24 16:43 UTC · model grok-4.3

classification 💻 cs.SI cs.LGphysics.soc-ph
keywords tensor factorizationsemi-supervised learningnode classificationsocial networksmulti-relational networksRESCALspammer detection
0
0 comments X

The pith

Extending RESCAL with a joint classification error term produces more accurate node classifications in multi-relational networks.

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

The paper introduces a semi-supervised version of the RESCAL tensor factorization method for multi-relational social networks. It adds a classification error term to the standard factorization objective and optimizes both together. This allows the model to use available class labels to guide the factorization process. Evaluation on real-world data shows improved accuracy for tasks like identifying spammers compared to unsupervised factorization.

Core claim

The paper proposes extending RESCAL to a semi-supervised factorization method that combines a classification error term with the standard factor optimization process. This coupled optimization models the tensorial data from all relations while accounting for classification performance, leading to more accurate models when supervision is available.

What carries the argument

The semi-supervised RESCAL extension, which jointly optimizes tensor factorization and a classification error term to incorporate label information into the factorization of multi-relational data.

If this is right

  • Models become more accurate for node classification when class labels are incorporated during factorization.
  • The method can identify nodes with special roles such as spammers in social networks.
  • Collective use of all relations in the tensor improves predictions over separate models per relation.
  • Joint optimization assimilates observed information from relations and classification performance simultaneously.

Where Pith is reading between the lines

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

  • Similar joint optimization approaches might improve other factorization methods in network analysis.
  • Testing on larger or different types of networks could reveal scalability limits of the coupled optimization.
  • The approach might generalize to other prediction tasks beyond node classification in social networks.

Load-bearing premise

That adding and jointly optimizing the classification error term does not destabilize the factorization or require dataset-specific tuning that invalidates the accuracy gains.

What would settle it

Running the semi-supervised RESCAL and standard RESCAL on the same real-world social network datasets and measuring if the classification accuracy is consistently higher with the added error term.

Figures

Figures reproduced from arXiv: 1907.10416 by Georgios Katsimpras, Georgios Paliouras.

Figure 1
Figure 1. Figure 1: A multi-relational network can be easily mapped to a ten [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The process of k-nn classification. The top-k similar en [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The folding-in process of CLASS-RESCAL. Note that R remains unchanged [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Behavior of CLASS-RESCAL while using different num [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Running times of all methods for different tensor sizes. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

This paper proposes a method to guide tensor factorization, using class labels. Furthermore, it shows the advantages of using the proposed method in identifying nodes that play a special role in multi-relational networks, e.g. spammers. Most complex systems involve multiple types of relationships and interactions among entities. Combining information from different relationships may be crucial for various prediction tasks. Instead of creating distinct prediction models for each type of relationship, in this paper we present a tensor factorization approach based on RESCAL, which collectively exploits all existing relations. We extend RESCAL to produce a semi-supervised factorization method that combines a classification error term with the standard factor optimization process. The coupled optimization approach, models the tensorial data assimilating observed information from all the relations, while also taking into account classification performance. Our evaluation on real-world social network data shows that incorporating supervision, when available, leads to models that are more accurate.

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

3 major / 1 minor

Summary. The paper proposes extending the RESCAL tensor factorization method for multi-relational networks to a semi-supervised variant by adding a classification error term to the factorization objective. It claims that this coupled optimization produces more accurate node classification on real-world social network data and is useful for identifying special nodes such as spammers.

Significance. If the accuracy gains can be shown to arise from the semi-supervised coupling rather than from per-dataset hyperparameter search, the method would provide a practical way to incorporate partial labels into collective tensor factorization for social network tasks.

major comments (3)
  1. [Abstract] The abstract states that the method 'extend[s] RESCAL to produce a semi-supervised factorization method that combines a classification error term with the standard factor optimization process' and that 'the coupled optimization approach... leads to models that are more accurate,' yet provides neither the explicit joint objective function nor any description of the balancing hyper-parameter between the RESCAL reconstruction loss and the classification term. Without this, the central claim cannot be evaluated.
  2. [Evaluation / Experiments] The evaluation asserts accuracy improvements from incorporating supervision, but the description gives no indication that the trade-off weight is fixed across experiments or derived without dataset-specific cross-validation. If the weight must be tuned per network, the reported gains cannot be attributed to the semi-supervised formulation itself rather than to the tuning procedure.
  3. [Abstract] No equations, error bars, or baseline comparisons (e.g., unsupervised RESCAL, other semi-supervised methods) are referenced in support of the accuracy claim, leaving the soundness of the central empirical result unverifiable from the given text.
minor comments (1)
  1. [Abstract] The phrasing 'models the tensorial data assimilating observed information' is awkward and should be clarified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive feedback. We address each major comment below, clarifying the presentation in the abstract and strengthening the evaluation to better demonstrate the benefits of the semi-supervised coupling.

read point-by-point responses

  1. Referee: [Abstract] The abstract states that the method 'extend[s] RESCAL to produce a semi-supervised factorization method that combines a classification error term with the standard factor optimization process' and that 'the coupled optimization approach... leads to models that are more accurate,' yet provides neither the explicit joint objective function nor any description of the balancing hyper-parameter between the RESCAL reconstruction loss and the classification term. Without this, the central claim cannot be evaluated.

    Authors: We agree the abstract is high-level and omits these details. Section 3.2 of the manuscript defines the joint objective as the RESCAL reconstruction loss plus a weighted classification error term, with hyper-parameter lambda controlling the trade-off. We will revise the abstract to include a concise reference to the joint objective and the balancing parameter. revision: yes

  2. Referee: [Evaluation / Experiments] The evaluation asserts accuracy improvements from incorporating supervision, but the description gives no indication that the trade-off weight is fixed across experiments or derived without dataset-specific cross-validation. If the weight must be tuned per network, the reported gains cannot be attributed to the semi-supervised formulation itself rather than to the tuning procedure.

    Authors: Lambda is selected via cross-validation per dataset, which is standard. To directly address the concern about attribution, we will add a sensitivity analysis in the revised manuscript showing results for a fixed lambda value across all datasets, confirming that the coupled optimization yields gains independent of per-dataset tuning. revision: yes

  3. Referee: [Abstract] No equations, error bars, or baseline comparisons (e.g., unsupervised RESCAL, other semi-supervised methods) are referenced in support of the accuracy claim, leaving the soundness of the central empirical result unverifiable from the given text.

    Authors: Abstract length constraints preclude equations and detailed results; these appear in Sections 4 and 5, which include comparisons to unsupervised RESCAL and other semi-supervised baselines along with error bars from repeated runs. We will revise the abstract to note the empirical improvements over baselines where space permits. revision: partial

Circularity Check

0 steps flagged

No circularity: new semi-supervised objective is independently defined and evaluated

full rationale

The paper defines an explicit extension to the RESCAL objective by adding a classification error term and jointly optimizing the combined loss. This is a constructive modeling choice rather than a re-expression of fitted quantities as predictions. The central claim rests on empirical accuracy improvements on real networks, not on any self-definitional reduction, fitted-input renaming, or load-bearing self-citation chain. No equations or steps in the abstract or method description reduce the reported gains to the inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that the added classification term integrates cleanly with the existing RESCAL loss.

pith-pipeline@v0.9.0 · 5689 in / 1027 out tokens · 20853 ms · 2026-05-24T16:43:46.157577+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

24 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    Un- derstanding data fusion within the framework of coupled matrix and tensor factorizations

    [Acar et al., 2013] Evrim Acar, Morten Arendt Rasmussen, Francesco Savorani, Tormod Næs, and Rasmus Bro. Un- derstanding data fusion within the framework of coupled matrix and tensor factorizations. Chemometrics and Intel- ligent Laboratory Systems, 129:53–63,

    work page 2013

  2. [2]

    Temporal analysis of semantic graphs using asalsan

    [Bader et al., 2007] Brett W Bader, Richard A Harshman, and Tamara G Kolda. Temporal analysis of semantic graphs using asalsan. In Data Mining,

    work page 2007

  3. [3]

    Discovering facts with boolean tensor tucker decom- position

    [Erdos and Miettinen, 2013] D´ora Erdos and Pauli Mietti- nen. Discovering facts with boolean tensor tucker decom- position. In Proceedings of the 22nd ACM international conference on Conference on information & knowledge management, pages 1569–1572. ACM,

    work page 2013

  4. [4]

    Collective spammer detection in evolving multi-relational social net- works

    [Fakhraei et al., 2015] Shobeir Fakhraei, James Foulds, Madhusudana Shashanka, and Lise Getoor. Collective spammer detection in evolving multi-relational social net- works. In Proceedings of the 21th ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining, pages 1769–1778. ACM,

    work page 2015

  5. [5]

    Sociology: A global perspec- tive

    [Ferrante, 2012] Joan Ferrante. Sociology: A global perspec- tive. Nelson Education,

    work page 2012

  6. [6]

    Triplerank: Ranking semantic web data by tensor decomposition

    [Franz et al., 2009] Thomas Franz, Antje Schultz, Sergej Sizov, and Steffen Staab. Triplerank: Ranking semantic web data by tensor decomposition. In International se- mantic web conference, pages 213–228. Springer,

    work page 2009

  7. [7]

    Multilinear pagerank

    [Gleich et al., 2015] David F Gleich, Lek-Heng Lim, and Yongyang Yu. Multilinear pagerank. SIAM Journal on Matrix Analysis and Applications , 36(4):1507–1541,

    work page 2015

  8. [8]

    Parafac: Parallel factor analysis

    [Harshman and Lundy, 1994] Richard A Harshman and Margaret E Lundy. Parafac: Parallel factor analysis. Com- putational Statistics & Data Analysis , 18(1):39–72,

    work page 1994

  9. [9]

    Mining social influence in microblogging via tensor factorization approach

    [Jingjing et al., 2013] Wei Jingjing, Tang Changhong, Liao Xiangwen, and Chen Guolong. Mining social influence in microblogging via tensor factorization approach. In Cloud Computing and Big Data (CloudCom-Asia), 2013 Interna- tional Conference on, pages 583–591. IEEE,

    work page 2013

  10. [10]

    Kolda and Brett W

    [Kolda and Bader, 2009] Tamara G. Kolda and Brett W. Bader. Tensor decompositions and applications. SIAM Rev., 51(3):455–500, August

    work page 2009

  11. [11]

    Non- negative matrix and tensor factorization based classifica- tion of clinical microarray gene expression data

    [Li and Ngom, 2010] Yifeng Li and Alioune Ngom. Non- negative matrix and tensor factorization based classifica- tion of clinical microarray gene expression data. 2010 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 438–443,

    work page 2010

  12. [12]

    Har: hub, authority and relevance scores in multi- relational data for query search

    [Li et al., 2012] Xutao Li, Michael K Ng, and Yunming Ye. Har: hub, authority and relevance scores in multi- relational data for query search. In Proceedings of the 2012 SIAM International Conference on Data Mining , pages 141–152. SIAM,

    work page 2012

  13. [13]

    Mining topic-level influence in het- erogeneous networks

    [Liu et al., 2010] Lu Liu, Jie Tang, Jiawei Han, Meng Jiang, and Shiqiang Yang. Mining topic-level influence in het- erogeneous networks. In Proceedings of the 19th ACM International Conference on Information and Knowledge Management, CIKM ’10, pages 199–208, New York, NY , USA,

    work page 2010

  14. [14]

    [McPherson et al., 2001] Miller McPherson, Lynn Smith- Lovin, and James M Cook

    ACM. [McPherson et al., 2001] Miller McPherson, Lynn Smith- Lovin, and James M Cook. Birds of a feather: Ho- mophily in social networks. Annual Review of Sociology , 27(1):415–444,

    work page 2001

  15. [15]

    Tensor factorization us- ing auxiliary information

    [Narita et al., 2011] Atsuhiro Narita, Kohei Hayashi, Ryota Tomioka, and Hisashi Kashima. Tensor factorization us- ing auxiliary information. In Joint European Confer- ence on Machine Learning and Knowledge Discovery in Databases, pages 501–516. Springer,

    work page 2011

  16. [16]

    Multirank: co-ranking for objects and relations in multi-relational data

    [Ng et al., 2011] Michaek Kwok-Po Ng, Xutao Li, and Yun- ming Ye. Multirank: co-ranking for objects and relations in multi-relational data. In Proceedings of the 17th ACM SIGKDD international conference on Knowledge discov- ery and data mining , pages 1217–1225. ACM,

    work page 2011

  17. [17]

    A three-way model for collective learning on multi-relational data

    [Nickel et al., 2011] Maximilian Nickel, V olker Tresp, and Hans-Peter Kriegel. A three-way model for collective learning on multi-relational data. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 809–816,

    work page 2011

  18. [18]

    Factorizing yago: Scalable machine learning for linked data

    [Nickel et al., 2012] Maximilian Nickel, V olker Tresp, and Hans-Peter Kriegel. Factorizing yago: Scalable machine learning for linked data. In Proceedings of the 21st In- ternational Conference on World Wide Web , WWW ’12, pages 271–280, New York, NY , USA,

    work page 2012

  19. [19]

    Relation Schema Induction using Tensor Factorization with Side Information

    ACM. [Nimishakavi et al., 2016] Madhav Nimishakavi, Uday Singh Saini, and Partha Talukdar. Relation schema induction using tensor factorization with side information. arXiv preprint arXiv:1605.04227,

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  20. [20]

    Learning optimal ranking with tensor factorization for tag recommendation

    [Rendle et al., 2009] Steffen Rendle, Leandro Balby Mar- inho, Alexandros Nanopoulos, and Lars Schmidt-Thieme. Learning optimal ranking with tensor factorization for tag recommendation. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discov- ery and data mining , pages 727–736. ACM,

    work page 2009

  21. [21]

    Structural analysis in multi-relational social networks

    [Tian Dai et al., 2012] Bing Tian Dai, Freddy Chong Tat Chua, and Ee-Peng Lim. Structural analysis in multi-relational social networks. In Proceedings of the 2012 SIAM International Conference on Data Mining , pages 451–462. SIAM,

    work page 2012

  22. [22]

    Some mathematical notes on three-mode factor analysis

    [Tucker, 1966] Ledyard R Tucker. Some mathematical notes on three-mode factor analysis. Psychometrika, 31(3):279– 311,

    work page 1966

  23. [23]

    Supervised nonnega- tive tensor factorization with maximum-margin constraint

    [Wu et al., 2013] Fei Wu, Xu Tan, Yi Yang, Dacheng Tao, Siliang Tang, and Yueting Zhuang. Supervised nonnega- tive tensor factorization with maximum-margin constraint. In AAAI,

    work page 2013

  24. [24]

    Mining topical influencers based on the multi- relational network in micro-blogging sites.China Commu- nications, 10(1):93–104, 2013

    [Zhaoyun et al., 2013] Ding Zhaoyun, Jia Yan, Zhou Bin, and Han Yi. Mining topical influencers based on the multi- relational network in micro-blogging sites.China Commu- nications, 10(1):93–104, 2013

    work page 2013