Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Learning to Match Jobs with Resumes from Sparse Interaction Data using Multi-View Co-Teaching Network

T0 review · 4 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Two-view co-teaching lifts job-resume matching past the best baseline

desk verdict Solid applied paper on job-resume matching with plausible 2–3% AUC gains, but the re-weighting equation is backwards as written and the evaluation lacks error bars. read the letter →

arxiv 2009.13299 v1 pith:WW5FTYMD submitted 2020-09-25 cs.CL cs.IRcs.LG

Shuqing Bian , Xu Chen , Wayne Xin Zhao , Kun Zhou , Yupeng Hou , Yang Song , Tao Zhang , Ji-Rong Wen This is my paper · ORCID
classification cs.CLcs.IRcs.LG
keywords Job-resumematchingMulti-viewco-teachingnetworkNoisylabelsSparseinteractiondataGraphneuralTextBERT
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

The paper tries to establish that job-resume matching can be made robust to sparse and noisy interaction data by combining two different views of the same problem: a text-based semantic matcher and a relation-based graph matcher, connected by a co-teaching mechanism. The two views share learned representations and then act as peer teachers, each one selecting or re-weighting the training instances the other will learn from. If this works, recruitment platforms can train accurate matchers from the noisy implicit feedback they already collect, without relying on large clean datasets. The authors report that the method outperforms all baselines on three real-world datasets, with average gains of 3.1%, 2.9%, and 3.5% over the best baseline on the Technology, Sales, and Design domains.

What carries the argument

The load-bearing object is the multi-view co-teaching network, a matched pair of models with complementary views. The text view is a hierarchical self-attention encoder (BERT sentence encoder plus Transformer document encoder); the relation view is a relational graph convolutional network (RGCN) over a job-resume graph whose edges are shared category labels and PageRank-selected keywords. The two views exchange representations and then serve as peer examiners: for each batch, one model computes confidence scores for instances, and the other trains on re-weighted or filtered versions of those instances. This co-teaching loop is what the paper credits for reducing the influence of noisy negative samples and for transferring evidence across similar jobs and resumes.

What would settle it

Take the randomly sampled negative instances used in training, have experts label their true match status, then measure whether the two views give confidently consistent predictions on the false negatives; the co-teaching mechanism's premise would fail if a substantial fraction of noisy negatives receive similar high-confidence scores from both views. A direct version of this test is already sketched in the paper's qualitative case study with four examples.

Watch

Extended reading notes

Core claim

The central claim is that a multi-view co-teaching network (MV-CoN) achieves the best reported job-resume matching performance on three real-world datasets by learning from sparse, noisy interaction data. The text-based component encodes job and resume sentences with BERT and a hierarchical Transformer, while the relation-based component runs a relational graph convolutional network over a graph where jobs and resumes are linked by shared category labels and keywords. The two components are fused twice: their representations are shared, and a co-teaching loop lets each component filter or re-weight the other's training batch based on prediction confidence. The paper reports that this design outperforms the strongest baselines, including a noisy-label method, on every metric and dataset tested, and that the improvement is largest when training data is most sparse.

Load-bearing premise

The method rests on the assumption that true job-resume pairs receive similar predictions from both the text and the relation view, while noisy pairs receive inconsistent or low-confidence predictions; if noisy pairs can fool both views, the co-teaching loop will discard useful training signal and the reported gains disappear.

Editorial extensions

If this is right

  • If the central claim holds, MV-CoN is the best known method for job-resume matching on these datasets and should transfer to similarly sparse recruitment platforms.
  • The co-teaching loop lets a platform use noisy implicit signals such as clicking and browsing as training negatives without degrading performance, because false negatives are down-weighted.
  • Representation sharing means the relation graph can bootstrap cold-start jobs and resumes that have few or no historical matches.
  • The soft re-weighting variant should be preferred over hard instance filtering, since it performs better in most reported settings.

Reading between the lines

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

  • The cross-view peer-selection trick is not specific to recruitment: any matching problem with a semantic view and a structural view (e.g., product recommendation from purchase edges, citation matching from text plus graph) could use the same co-teaching loop to denoise implicit feedback.
  • Because the paper finds that simple averaging of the two views (TR) is worse than either view alone, the gain is tied to the co-teaching interaction itself; a testable implication is that replacing co-teaching with attention-based fusion would not reproduce the reported improvement.
  • The relation graph is built from category labels and keyword co-occurrence; an obvious extension is to enrich those edges with learned skill embeddings and check whether the relation view alone improves correspondingly.
  • The released code and data make the headline numbers directly checkable: the reported average gains should be reproducible, and the advantage should shrink or vanish on a version of the data with clean, human-verified labels.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes MV-CoN, a multi-view co-teaching network for job-resume matching from sparse and noisy interaction data. The model combines a BERT-based hierarchical text encoder with an RGCN-based relation encoder, integrating them through representation sharing (concatenation or initialization) and a co-teaching data-enhancement procedure implemented either by instance re-weighting (Eq. 12) or instance filtering (Eq. 14). The authors evaluate on three real-world datasets from BOSS Zhipin and report consistent gains over eight baselines on AUC, Accuracy, Precision, Recall, and F1, with ablations, hyperparameter sensitivity, and qualitative case studies.

Significance. If the results hold, the paper addresses a practical and under-studied problem: job-resume matching with extremely sparse interaction data and noisy negative samples generated by random sampling from non-acceptance signals. The combination of a textual view and a relation-graph view, with a co-teaching loop to select reliable instances, is a sensible and original direction. The authors provide code and data, use three diverse real-world datasets, compare against eight baselines including a noise-robust method (UBD), and include ablations and qualitative analyses. The central practical claim is that MV-CoN, especially the re-weighting variant MV-CoN(R), outperforms the best baseline by 2.9-3.5% on average, which is meaningful for this task.

major comments (4)
  1. [Section 4.3.2, Eq. (12)] The re-weighting formula w^A_i = 1 - s_A(j_i, r_i, y_i) inverts the denoising objective stated in the surrounding text. The text says the peer model should 'increase the weight of high-quality samples and decrease the weight of unreliable samples', and §5.5 says that inconsistent predictions are penalized via Eq. (12). With s_A defined as the confidence score for the given label y_i, a sample on which the peer model is confident receives a low weight, and a sample on which it disagrees receives a high weight. Since the suspected noisy instances are the sampled negatives, this literal implementation would downweight reliable negative samples and upweight false negatives, which is the opposite of the intended effect. Because MV-CoN(R) is the best configuration and the one used in the ablation study, this apparent sign inversion is load-bearing for the reported 3.1%, 2.9%, and 3.5% improvements. Please clarify the definition of s_A or correct the formula to w^A_i = s_A(j_i, r_i, y_i) (or an equivalent monotone function).
  2. [Section 5.1.3 and Table 2] All reported results appear to come from a single run, with no standard deviations, confidence intervals, or significance tests. The reported margins over the best baseline are 2.9-3.5% on fixed test sets, which is within the range of run-to-run variation for deep models, especially with stochastic negative sampling and co-teaching. Please provide multiple independent runs (or seeded runs) with mean and variance, and where feasible a paired significance test across methods, to support the claim of consistent superiority.
  3. [Sections 4.2 and 5.1.3] The paper does not state whether the job-resume relation graph includes validation and test nodes, nor whether the 8:1:1 split is by interaction pair or by node (job/resume). If the graph is constructed over the full dataset and the RGCN node representations are computed on the full graph during training, the relation-based component is evaluated in a transductive setting that uses test-node features (and potentially test-node connectivity) during training, while the text-only baselines are inductive. This is an asymmetric evaluation setup. Please specify the graph construction time, the splitting procedure (pair-level versus node-level), and discuss the impact of transductive information on the comparison with baselines. If the split is pair-level, there is likely leakage in the form of the same job or resume appearing in both training and test pairs.
  4. [Sections 4.3.1 and 4.3.2] The co-teaching assumption is that true samples receive similar predictions under 'different model views' while noisy ones are inconsistent. However, the two views are made highly dependent through representation enhancement: the relation-based model initializes its node states with the text encoder's output (Eqs. 10-11), and the text-based model uses concatenated relation representations (Eqs. 8-9). This coupling weakens the independence of the two peer signals and may cause correlated mistakes, undermining the premise that a peer model can reliably identify the other view's noisy samples. Please provide an empirical analysis of view agreement, e.g., the correlation or disagreement rate between the two components on clean versus noisy training instances, to validate this core assumption.
minor comments (6)
  1. [Algorithm 1] The loop structure is confusing: variable b is used for the epoch loop and e for the batch loop, but in the text 'for each batch update' the batch is the inner loop; please rewrite the algorithm so that the batch loop is innermost and the variable names are consistent.
  2. [Algorithm 1, lines 5-6] The notation 'Learning( eDB, LB) from model A and DB' is ambiguous; please specify what function is being called, what it returns, and how the filtered/re-weighted sets are used in the loss updates.
  3. [Section 5.1.1] The sentence 'By equally sampling from these two resources, the final ratio between the numbers of positive and negative samples is set to 1:1' should clarify whether this ratio applies only to the training set, since the validation and test sets use explicit accept/reject status and remain imbalanced.
  4. [Table 2] The marker definition says 'we use bold fonts and * to mark the best performance and the best baseline performance, respectively', but the * marker appears only on the best baseline in each column; the text should state the convention clearly and distinguish column-wise best from overall best.
  5. [Eq. (7)] The clause 'where is the predicted n(L')_j and n(L')_r' is grammatically incomplete; please revise to a full sentence.
  6. [References and figures] References [23] and [24] are duplicates of the same ESWC paper; Figures 3 and 4 contain typos ('Traing ratio' should be 'Training ratio'), and Figure 5 is crowded and difficult to read at the current resolution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's co-teaching loop is self-referential by design, but no prediction reduces to its inputs by construction.

full rationale

This is an empirical machine-learning paper whose central claims are benchmark comparisons, not mathematical derivations from first principles. The text-based and relation-based views are independently specified (BERT/Transformer versus RGCN), and the reported gains are measured on held-out job-resume pairs with explicit accept/reject labels. The co-teaching mechanism uses each model's own predictions to select or weight training instances, which is self-referential as a training procedure, but it does not make a test-time prediction equivalent to an input by construction. The only self-citations ([2] for DGMN and [30] for JRMPM) appear as baselines or related work and are not load-bearing for the proposed method; no uniqueness theorem or imported ansatz carries the argument. One internal issue worth noting is that Eq. 12 defines wA_i = 1 - sA(ji, ri, yi), which conflicts with the surrounding text's claim that the peer model should 'increase the weight of high-quality samples'; a high-confidence correct label receives a low weight. This is a potential implementation or formulation inconsistency that could affect the reported results, but it is a correctness concern rather than a circularity: the re-weighting formula is not equivalent to the paper's output by definition, and the core comparison with external baselines remains independently evaluable. Therefore no circular step is established, and the appropriate score is 0.

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

The method is empirical; no derivation is claimed. The main external inputs are pretrained BERT, the co-teaching assumption, the relation-graph construction heuristic, and several hyperparameters. The paper reports tuning for delta and L but does not report K or RGCN depth, leaving part of the model under-specified.

free parameters (4)
  • selection ratio delta = 0.8
    Held-out rate for instance filtering; tuned over {0.2, 0.4, 0.6, 0.8, 1.0} and best at 0.8 in Figure 3(a).
  • Transformer layers L = 2
    Tuned over {1, 2, 3, 4}; best at 2 in Figure 3(b).
  • keyword count K
    Number of top PageRank keywords used to create relation-graph links; value is not reported in the paper, though the model depends on it.
  • positive-to-negative sampling ratio = 1:1
    Set by equally sampling clicking and browsing instances as negative training examples; this determines the label noise level in the training set.
assumptions (4)
  • domain assumption True samples receive similar predictions under different model views, while noisy ones are not easy to cheat all the models.
    Stated in Section 4.3.2; the co-teaching filtering and re-weighting mechanisms fail if this premise is false.
  • domain assumption Randomly sampled clicking and browsing pairs are useful noisy negatives whose label noise can be learned through.
    Used in Section 5.1.1 to build training data; the noise level and structure are assumed manageable.
  • domain assumption Category labels and top-PageRank keywords create links that capture useful job-resume correlations.
    The entire relation graph in Section 4.2.1 is built on this heuristic.
  • domain assumption Pretrained BERT-Base-Chinese transfers to Chinese recruitment documents.
    Sentence encoder is initialized with BERT in Section 5.1.3; no domain adaptation is demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Match Jobs with Resumes from Sparse Interaction Data using Multi-View Co-Teaching Network." pith.science (2026). https://pith.science/paper/WW5FTYMD

@misc{pith2026200913299,
  author       = {Pith},
  title        = {Pith review of: Learning to Match Jobs with Resumes from Sparse Interaction Data using Multi-View Co-Teaching Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WW5FTYMD}},
  note         = {Machine review of arXiv:2009.13299}
}
read the original abstract

With the ever-increasing growth of online recruitment data, job-resume matching has become an important task to automatically match jobs with suitable resumes. This task is typically casted as a supervised text matching problem. Supervised learning is powerful when the labeled data is sufficient. However, on online recruitment platforms, job-resume interaction data is sparse and noisy, which affects the performance of job-resume match algorithms. To alleviate these problems, in this paper, we propose a novel multi-view co-teaching network from sparse interaction data for job-resume matching. Our network consists of two major components, namely text-based matching model and relation-based matching model. The two parts capture semantic compatibility in two different views, and complement each other. In order to address the challenges from sparse and noisy data, we design two specific strategies to combine the two components. First, two components share the learned parameters or representations, so that the original representations of each component can be enhanced. More importantly, we adopt a co-teaching mechanism to reduce the influence of noise in training data. The core idea is to let the two components help each other by selecting more reliable training instances. The two strategies focus on representation enhancement and data enhancement, respectively. Compared with pure text-based matching models, the proposed approach is able to learn better data representations from limited or even sparse interaction data, which is more resistible to noise in training data. Experiment results have demonstrated that our model is able to outperform state-of-the-art methods for job-resume matching.

Figures

Figures reproduced from arXiv: 2009.13299 by the authors.

Figure 1
Figure 1. A motivating example for relation-based match. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of our proposed approach. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance tuning with the selection ratio [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance comparison by varying the amount [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: An illustrative example for showing the effect of different views in our co-teaching network. Here, we use circles and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In ICML 2009. 41–48

    work page 2009

  2. [2]

    Shuqing Bian, Wayne Xin Zhao, Yang Song, Tao Zhang, and Ji-Rong Wen. 2019. Domain Adaptation for Person-Job Fit with Transferable Deep Global Match Network. In EMNLP-IJCNLP 2019. 4809–4819

    work page 2019

  3. [3]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL-HLT 2019. 4171–4186

    work page 2019

  4. [4]

    Mamadou Diaby, Emmanuel Viennet, and Tristan Launay. 2013. Toward the next generation of recruitment tools: an online social network-based job recommender system. In ASONAM 2013. IEEE, 821–828

    work page 2013

  5. [5]

    Yang Fan, Fei Tian, Tao Qin, Xiang-Yang Li, and Tie-Yan Liu. 2018. Learning to Teach. In ICLR 2018. 350–357

    work page 2018

  6. [6]

    Jiazhan Feng, Chongyang Tao, Wei Wu, Yansong Feng, Dongyan Zhao, and Rui Yan. 2019. Learning a Matching Model with Co-teaching for Multi-turn Response Selection in Retrieval-based Dialogue Systems. In ACL 2019. 3805–3815

    work page 2019

  7. [7]

    Tsang, and Masashi Sugiyama

    Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS 2018. 8536–8546

    work page 2018

  8. [8]

    Haibo He and Edwardo A Garcia. 2009. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering 21, 9 (2009), 1263–1284

    2009

Show all 33 references
  1. [9]

    Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse Predictive Analytics. In SIGIR 2017. 355–364

    2017

  2. [10]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In CIKM 2013. 2333–2338

    2013

  3. [11]

    Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. 2018. MentorNet: Learning Data-Driven Curriculum for Very Deep Neural Networks on Corrupted Labels. In ICML 2018. 2309–2318

    2018

  4. [12]

    Miao Jiang, Yi Fang, Huangming Xie, Jike Chong, and Meng Meng. 2019. User click prediction for personalized job recommendation. WWW 2019 (2019), 325– 345

    2019

  5. [13]

    Krishnaram Kenthapadi, Benjamin Le, and Ganesh Venkataraman. 2017. Personalized Job Recommendation System at LinkedIn: Practical Challenges and Lessons Learned. In RecSys 2017. 346–347

    2017

  6. [14]

    Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification. In EMNLP 2014. 1746–1751

    2014

  7. [15]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR 2017. 436–444

    2017

  8. [16]

    Pawan Kumar, Benjamin Packer, and Daphne Koller

    M. Pawan Kumar, Benjamin Packer, and Daphne Koller. 2010. Self-Paced Learning for Latent Variable Models. In NeurIPS 2010. 1189–1197

    2010

  9. [17]

    Yao Lu, Sandy El Helou, and Denis Gillet. 2013. A recommender system for job seeking and recruiting website. In WWW 2013. 963–966

    2013

  10. [18]

    Yong Luo, Huaizheng Zhang, Yonggang Wen, and Xinwen Zhang. 2019. ResumeGAN: An Optimized Deep Representation Learning Framework for Talent- Job Fit via Adversarial Learning. In CIKM 2019. 1101–1110

    2019

  11. [19]

    when to update

    Eran Malach and Shai Shalev-Shwartz. 2017. Decoupling "when to update" from "how to update". In NeurIPS 2017. 960–970

    2017

  12. [20]

    Chuan Qin, Hengshu Zhu, Tong Xu, Chen Zhu, Liang Jiang, Enhong Chen, and Hui Xiong. 2018. Enhancing Person-Job Fit for Talent Recruitment: An Ability- aware Neural Network Approach. In SIGIR 2018

    2018

  13. [21]

    Rohan Ramanath, Hakan Inan, Gungor Polatkan, Bo Hu, Qi Guo, Cagri Ozcaglar, Xianren Wu, Krishnaram Kenthapadi, and Sahin Cem Geyik. 2018. Towards Deep and Representation Learning for Talent Search at LinkedIn. In CIKM 2018. 2253–2261

    2018

  14. [22]

    Konstan, and John Riedl

    Badrul Munir Sarwar, George Karypis, Joseph A. Konstan, and John Riedl. 2010. Item-based collaborative filtering recommendation algorithms. In WWW 2010

    2010

  15. [24]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling Relational Data with Graph Convolutional Networks. In ESWC 2018. 593–607

    2018

  16. [25]

    Walid Shalaby, BahaaEddin AlAila, Mohammed Korayem, Layla Pournajaf, Khalifeh AlJadda, Shannon Quinn, and Wlodek Zadrozny. 2017. Help me find a job: A graph-based approach for job recommendation at scale. In IEEE BigData

    2017

  17. [26]

    Dazhong Shen, Hengshu Zhu, Chen Zhu, Tong Xu, Chao Ma, and Hui Xiong

  18. [27]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In NeurIPS 2017. 5998–6008

    2017

  19. [28]

    Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Jian-Huang Lai, and Tie-Yan Liu. 2018. Learning to Teach with Dynamic Loss Functions. In NeurIPS 2018. 6467–6478

    2018

  20. [29]

    Kun Xu, Yansong Feng, Songfang Huang, and Dongyan Zhao. 2015. Semantic Relation Classification via Convolutional Neural Networks with Simple Negative Sampling. In EMNLP 2015. 536–540

    2015

  21. [30]

    Rui Yan, Ran Le, Yang Song, Tao Zhang, Xiangliang Zhang, and Dongyan Zhao

  22. [31]

    Fengxiang Yang, Ke Li, Zhun Zhong, Zhiming Luo, Xing Sun, Hao Cheng, Xiaowei Guo, Feiyue Huang, Rongrong Ji, and Shaozi Li. 2020. Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification. In AAAI 2020. 12597– 12604

    2020

  23. [32]

    Peng Zhou, Zhenyu Qi, Suncong Zheng, Jiaming Xu, Hongyun Bao, and Bo Xu. 2016. Text Classification Improved by Integrating Bidirectional LSTM with Two-dimensional Max Pooling. In COLING 2016. 3485–3495

    2016

  24. [2018]

    In IJCAI 2018

    A Joint Learning Approach to Intelligent Job Interview Assessment. In IJCAI 2018. 3542–3548

    2018

  25. [2019]

    In KDD 2019

    Interview Choice Reveals Your Preference on the Market: To Improve Job-Resume Matching through Profiling Memories. In KDD 2019. 914–922

    2019

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.