Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Hierarchical Job Classification with Similarity Graph Integration

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By embedding job postings, SOC categories, Carotene categories, and a Carotene similarity graph into one shared space, the paper reports job-classification accuracy of 0.948 at the SOC level and 0.893 at the Carotene level, above all…

desk verdict Novel similarity-graph integration for hierarchical job classification, but the printed triplet losses are sign-inverted and the TRA metric is measured on training triplets, so the results as written are not reproducible. read the letter →

arxiv 2507.09949 v1 pith:6US2HMKH submitted 2025-07-14 cs.LG

Md Ahsanul Kabir , Kareem Abdelfatah , Mohammed Korayem , Mohammad Al Hasan This is my paper
classification cs.LG
keywords hierarchicaltextclassificationjobjointembeddingtripletlosssimilaritygraphhardnegativeminingSOCCarotene
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 addresses a practical problem in online recruitment: assigning each job posting to a coarse category (SOC) and a fine category (Carotene) at the same time. It proposes to learn one shared embedding space for jobs, SOCs, and Carotenes, with training losses that enforce both the parent-child hierarchy and a similarity graph among Carotenes. The authors claim this joint approach reaches 0.948 SOC accuracy and 0.893 Carotene accuracy on a 456K-postings proprietary dataset, beating text-classification and hierarchical-classification baselines. The broader point is that label structure, both the tree and the similarity edges, can be encoded directly into the metric space used for classification rather than treated as a post-processing step.

What carries the argument

The central object is a shared $q$-dimensional embedding space in which each job is embedded by a pretrained transformer (BERT or Sentence-Transformer), and each SOC and Carotene node is a trainable embedding row, $E_s \in \mathbb{R}^{m\times q}$ and $E_c \in \mathbb{R}^{n\times q}$. Two categorical cross-entropy losses classify a job embedding into its SOC and Carotene; four margin-based triplet losses act on the normalized dot products of (SOC, child Carotene, non-child Carotene), (Carotene, graph-neighbor Carotene, non-neighbor), (job, true SOC, false SOC), and (job, true Carotene, false Carotene). The intended effect is to make positive pairs more similar than negatives by margin $\alpha$, so the hierarchy and the Carotene similarity graph are both encoded in the geometry of the latent space. As printed, the sign convention in the equations has the opposite effect, so the reported results depend on a correction not stated in the text.

What would settle it

Implement the four triplet losses exactly as printed with margin $\alpha>0$ and train on a two-level taxonomy dataset with known positive and negative triplets: the triplet ranking accuracy on held-out triplets should be near zero, not the reported 0.998, since the loss rewards the wrong ordering. A corrected-sign implementation on the same data is the direct test of whether the claimed accuracy gains are caused by the hierarchy and similarity losses.

Watch

Extended reading notes

Core claim

The paper's central claim is that a joint embedding of jobs, SOCs, and Carotenes, trained with a loss that combines two cross-entropy classifiers with four margin-based triplet losses, can classify a job into both levels of a two-level taxonomy at once while preserving the taxonomy's hierarchical structure and a separately given similarity graph over the leaf Carotene nodes. On a proprietary dataset of 456K job postings with 23 SOCs and 4,778 Carotenes, the best variant, Hierarchical-Sim-Classification (Hard), reaches SOC accuracy 0.948 and Carotene accuracy 0.893, surpassing BERT, Sentence-Transformer, HQC, HiAGM, and HPT baselines on accuracy and on the paper's triplet ranking accuracy measure. The ablation study attributes the gain to each loss component: removing the SOC or Carotene cross-entropy term collapses the corresponding accuracy, and removing the soc-carotene or carotene-carotene triplet term drops the corresponding triplet ranking accuracy to about 0.69.

Load-bearing premise

The load-bearing premise is that the four margin-based triplet losses really pull positive pairs closer than negatives by margin $\alpha$; the equations as printed are minimized when the negative is more similar than the positive, so the reported accuracies rest on an unstated corrected implementation.

Editorial extensions

If this is right

  • A job can be labeled at both taxonomy levels from one embedding, so no separate classifiers or cascades are needed for coarse and fine labels.
  • Hard-negative triplet mining, which ranks non-child Carotenes by similarity to signature text, yields higher accuracy than random negatives, so negative selection matters as much as the loss itself.
  • The method transfers to any domain with a class taxonomy plus a class-similarity graph, making it a candidate recipe for other large-scale hierarchical labeling problems.
  • Ablation results imply that each of the six loss terms is load-bearing: removing the classification loss for one level destroys that level's accuracy, and removing a triplet term degrades the corresponding ordering metric.
  • TRA values near 0.999 indicate the learned geometry preserves hierarchy and similarity ordering on held-out triplets, which is what the joint loss is designed to do.

Reading between the lines

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

  • A public-domain replication on a non-proprietary two-level hierarchy, for example product listings or court documents with a category tree, would settle how much of the gain comes from the joint embedding method versus from the in-house Carotene taxonomy and its similarity graph.
  • The same recipe could be applied to hierarchical problems where labels have known pairwise similarity but no text, by replacing the transformer job encoder with any entity embedding; the triplet losses do not depend on text.
  • Because the printed margin losses are sign-inverted, anyone building on the paper from the text alone would likely train a model that rewards the wrong ordering; the reported numbers imply an unstated corrected implementation, and a code release would resolve this.
  • Varying the maximum degree of the Carotene similarity graph, capped at five in the paper, is a natural testable extension to see how accuracy and triplet ranking depend on the density and quality of the similarity edges.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Hierarchical-Sim-Classification, a model that embeds job postings and two levels of a job taxonomy (SOC and Carotene) into a shared latent space, using two cross-entropy losses for classification and four margin-based triplet losses to encode SOC-Carotene hierarchy, Carotene-Carotene similarity, and job-label relations. The method is evaluated on a proprietary dataset of 456K job postings with 23 SOCs and 4,778 Carotenes. The authors report that the proposed 'Hard' variant achieves SOC accuracy 0.948 and Carotene accuracy 0.893, outperforming BERT, Sentence-Transformer, HQC, HiAGM, and HPT, and they include an ablation study and a case study.

Significance. The industrial problem is relevant and the dataset is unusually large, which makes the paper potentially valuable. The idea of augmenting hierarchical classification with a similarity graph among leaf nodes and encoding both through triplet losses is reasonable and goes beyond existing hierarchical text classification methods. The paper also provides a useful ablation showing the contribution of each loss term. However, the manuscript's central claims are weakened by a sign error in all four margin loss equations, by a TRA metric evaluated on training triplets, by the absence of error bars or multiple splits, and by an implausibly low HiAGM baseline. These issues must be resolved before the claimed gains can be accepted.

major comments (4)
  1. [4.3 (Eqs. Lsoc-car, Lcar-car, Ljob-soc, Ljob-car)] All four margin losses are printed in the form [alpha + sim(positive) - sim(negative)]_+. Since the hinge is minimized when sim(negative) >= sim(positive) + alpha, a literal implementation would push negative pairs closer than positive pairs, the opposite of the stated objective in Section 3.2 and Section 4.3. The same inversion appears in the contrastive loss Lcontrastive earlier in Section 4.3, where the positive pair is in the denominator and the negative pair is in the numerator. Because Ljob-soc and Ljob-car are the only mechanisms that pass hierarchy and similarity information into job embeddings, the reported high TRA scores and accuracy gains are inconsistent with the equations as written; the experiments must have used a different sign convention. Please correct the equations, state the actual implemented losses, and provide code or pseudo-code so the method is reproducible.
  2. [5.4 and 5.1] The TRA metric is computed on 'all the generated triplets stored in the file,' and Section 5.1 shows these files are the training triplets used to fit Lsoc-car, Lcar-car, Ljob-soc, and Ljob-car. Near-perfect TRA values (0.998-0.999) on the training triplets therefore only show that the model fit the training triples; they do not measure whether the learned embeddings generalize. Please report TRA on held-out triplets (e.g., triplets created from test jobs or from edges withheld during training) and compare it with a random-embedding baseline.
  3. [5.2 and Table 1] All results come from a single random 6:2:2 split with no error bars, no repeated runs, and no statistical tests. Moreover, Section 6 reports that hyperparameters (lambdas, margin, n_neg) were tuned to maximize accuracy, but the paper does not state that a separate validation split was used for this tuning; if the test set was used for tuning, the reported numbers are optimistically biased. Please report mean and standard deviation over multiple seeds or splits, and use a validation set for all hyperparameter selection.
  4. [Table 1, HiAGM row] HiAGM's SOC accuracy of 0.403 and Carotene accuracy of 0.332 are far below every other method, including non-hierarchical BERT baselines (0.867-0.886). The text attributes this to HiAGM being a 'non-scalable bi-LSTM based model,' but such a large gap suggests a possible misconfiguration or a mismatch between HiAGM and the 4,778-label Carotene taxonomy. Please verify the HiAGM implementation, report its hyperparameters and training details, and, if possible, include a sanity check on a public hierarchical text classification benchmark; otherwise the baseline comparison is not trustworthy.
minor comments (5)
  1. [Title page] The author affiliation on the first page identifies CareerBuilder LLC, while the text replaces the company name with 'X Company' to preserve anonymity; please make this consistent or remove the anonymization claim.
  2. [Section 8] The sentence introducing Table 3 ends with 'which achieved 95', which is incomplete.
  3. [Table 3 caption] The table caption refers to 'Hierarchical Text Classification (Hard)' but the model in Section 5.4 is called 'Hierarchical-Sim-Classification (Hard)'; use one consistent name throughout.
  4. [Figures] Figure 1 and Figure 2 are referenced in Sections 2 and 5.4, but no figures appear in the manuscript text; please include them.
  5. [Notation] The notation is inconsistent: Section 4.2 defines Es as R^{m x q} with columns e^T_sj, but the loss equations use e_sj and e_ji without transposes, and the dot product is written as <a,b> in some places and as a centered dot in others; please standardize the notation.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported Triplet Ranking Accuracy (TRA) is computed on the same triplets used to fit the triplet losses, so the near-perfect TRA values are forced by construction rather than evidence of generalization.

  1. fitted input called prediction [Section 5.1 (Triplet Generation Technique) and Section 5.4 (Results), Table 1]
    "All these triplets are then saved to a file... The exact same n_sample hyperparameter is used to sample from the list of triplet files... To compute this metric, we considered all the generated triplets stored in the file."

    The four margin losses in Section 4.3 (Lsoc-car, Lcar-car, Ljob-soc, Ljob-car) are optimized on triplets sampled from files generated in Section 5.1. TRA ('Triplet Ranking Accuracy') is then measured on the same stored triplets: 'we considered all the generated triplets stored in the file.' Thus TRA reports how well the model ranks the very triplets it was trained on, not how well the learned embeddings generalize to unseen hierarchy/similarity relations. The near-perfect TRA values (0.998–0.999) are therefore expected by construction from fitting the triplet losses to those same triplets, and cannot independently validate the claim that the embeddings preserve hierarchy and similarity. This is a fitted input being reported as a predictive/validation result.

full rationale

The paper's main classification claim (SOC accuracy 0.948, Carotene accuracy 0.893) is evaluated on a held-out test split (6:2:2 random split, 91K test instances) and is therefore not circular: it measures generalization to unseen job postings. The TRA metric, however, is circular because it is computed on 'all the generated triplets stored in the file,' which are exactly the triplets used to fit the triplet margin losses; high TRA is forced by construction and does not demonstrate that the embeddings preserve hierarchy or similarity in a generalizable way. The self-citations in the paper (CareerBuilder patents, prior job-skill embedding work) appear only as related work and are not load-bearing justifications of the central claim, so no self-citation-circularity step is scored. The inverted sign in the printed triplet losses (Section 4.3) is a serious correctness concern but is not a circularity issue; it makes the reported results internally inconsistent rather than equivalent to the inputs by definition. Overall, the circularity is partial: the TRA-based validation reduces by construction, while the main held-out accuracy comparison retains independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are postulated. The Carotene taxonomy and similarity graph are proprietary data objects that predate the paper, not invented explanatory constructs.

free parameters (5)
  • lambda1-lambda6 (loss weights) = lambda2 = 0.5 in Section 6 vs 0.7 in Table 2; others 0.3
    Hand-tuned loss weights controlling the balance of classification and triplet losses.
  • n_neg (number of negatives per triplet edge) = 15 or 20 for job-soc, 200 for other triplet types
    Number of negative Carotenes sampled per positive edge; tuned per relation type.
  • n_sample (number of triplets sampled per batch) = unspecified
    Said to be fixed for all variants, but no value is given in the paper.
  • margin (alpha in hinge losses) = 0.4
    Tuned over [0.1, 0.6]; the paper states 0.4 gives best accuracy.
  • similarity graph max degree = 5
    Each Carotene is connected to at most five others via an internal threshold and ranking criteria; this design choice shapes the graph signal.
assumptions (5)
  • domain assumption The Carotene and SOC taxonomies are correct and complete labels for the dataset.
    The entire classification target depends on the quality of the proprietary Carotene taxonomy, which is not publicly documented.
  • domain assumption The Carotene similarity graph, built from proprietary transition data, is a meaningful signal and does not leak future test information.
    The graph is constructed using internal similarity scores and ranking criteria that are not specified; if it is built using test-period transitions, the reported gains would not generalize.
  • domain assumption The pretrained language model embeddings provide adequate job semantics.
    The job encoder is a frozen pretrained model (BERT or Sentence-Transformer); the paper does not fine-tune it, so its quality is assumed.
  • domain assumption The 6:2:2 random split is i.i.d. with no leakage between train, validation, and test.
    The split is described as random, but no precautions against duplicate or near-duplicate postings are mentioned.
  • ad hoc to paper The margin-loss equations in Section 4.3 are assumed to be implemented with the correct sign, since the printed sign is inverted.
    The equations as written minimize similarity between positives and maximize similarity between negatives, contrary to the prose; the paper's results depend on an unstated sign correction in the implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Job Classification with Similarity Graph Integration." pith.science (2026). https://pith.science/paper/6US2HMKH

@misc{pith2026250709949,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Job Classification with Similarity Graph Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6US2HMKH}},
  note         = {Machine review of arXiv:2507.09949}
}
read the original abstract

In the dynamic realm of online recruitment, accurate job classification is paramount for optimizing job recommendation systems, search rankings, and labor market analyses. As job markets evolve, the increasing complexity of job titles and descriptions necessitates sophisticated models that can effectively leverage intricate relationships within job data. Traditional text classification methods often fall short, particularly due to their inability to fully utilize the hierarchical nature of industry categories. To address these limitations, we propose a novel representation learning and classification model that embeds jobs and hierarchical industry categories into a latent embedding space. Our model integrates the Standard Occupational Classification (SOC) system and an in-house hierarchical taxonomy, Carotene, to capture both graph and hierarchical relationships, thereby improving classification accuracy. By embedding hierarchical industry categories into a shared latent space, we tackle cold start issues and enhance the dynamic matching of candidates to job opportunities. Extensive experimentation on a large-scale dataset of job postings demonstrates the model's superior ability to leverage hierarchical structures and rich semantic features, significantly outperforming existing methods. This research provides a robust framework for improving job classification accuracy, supporting more informed decision-making in the recruitment industry.

Figures

Figures reproduced from arXiv: 2507.09949 by the authors.

Figure 1
Figure 1. Framework of the Hierarchical Job Classification [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE Plot of SOCs and Their Corresponding [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 24 canonical work pages

  1. [1]

    Khalifeh Al Jadda, Mohammed Korayem, Brandon Tripp, Amber Soley, and Steve Proell. 2024. Interactive Job Recommendation and Application Submission Tools of Employment Website Entities. US Patent App. 18/478,442

    work page 2024

  2. [2]

    Siddhartha Banerjee, Cem Akkaya, Francisco Perez-Sorrosal, and Kostas Tsiout- siouliklis. 2019. Hierarchical Transfer Learning for Multi-label Text Classifica- tion. In Proceedings of the 57th Annual Meeting of the Association for Compu- tational Linguistics, Anna Korhonen, David Traum, and Lluís Màrquez (Eds.). The 48’th ACM SIGIR conference, July 13th t...

    doi:10.18653/v1/p1 2019

  3. [3]

    Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information.Transactions of the association for computational linguistics 5 (2017), 135–146

    work page 2017

  4. [4]

    Huanhuan Cao, Derek Hao Hu, Dou Shen, Daxin Jiang, Jian-Tao Sun, Enhong Chen, and Qiang Yang. 2009. Context-aware query classification. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (Boston, MA, USA) (SIGIR ’09). Association for Computing Machinery, New York, NY, USA, 3–10. https://doi.o...

    arXiv 2009

  5. [5]

    William Cavnar and John Trenkle. 2001. N-Gram-Based Text Categorization. Proceedings of the Third Annual Symposium on Document Analysis and Information Retrieval (05 2001)

    work page 2001

  6. [6]

    Xiquan Cui, Estelle Afshar, Khalifeh Al-Jadda, Srijan Kumar, Julian McAuley, Tao Ye, Kamelia Aryafar, Vachik Dave, and Mohammad Korayem. 2021. Workshop on Online and Adaptative Recommender Systems (OARS). In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 4116–4117

    work page 2021

  7. [7]

    Vachik S Dave, Baichuan Zhang, Mohammad Al Hasan, Khalifeh AlJadda, and Mohammed Korayem. 2018. A combined representation learning approach for better job and skill recommendation. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management . 1997–2005

    work page 2018

  8. [8]

    Dave, Baichuan Zhang, Mohammad Al Hasan, Khalifeh AlJadda, and Mohammed Korayem

    Vachik S. Dave, Baichuan Zhang, Mohammad Al Hasan, Khalifeh AlJadda, and Mohammed Korayem. 2018. A Combined Representation Learning Approach for Better Job and Skill Recommendation. InProceedings of the 27th ACM International Conference on Information and Knowledge Management (Torino, Italy) (CIKM ’18). Association for Computing Machinery, New York, NY, U...

    arXiv 2018

Show all 36 references
  1. [9]

    Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

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

    2019

  2. [10]

    Brian Gaspar, Khalifeh Aljadda, Mohammed Korayem, Alan Chern, Phuong Hoang, Robert Malony, Kristen Ruel, Humair Ghauri, Vachik S Dave, Mohammad Al Hasan, et al. 2020. Automated Systems and Methods for Determining Jobs, Skills, and Training Recommendations. US Patent App. 16/528,467

    2020

  3. [11]

    Brian Gaspar, Mohammed Korayem, WANG Jingya, Kareem Abdelfatah, Janani Balaji, Robert Malony, Eric Presely, and Humair Ghauri. 2020. Classification of job titles via machine learning. US Patent App. 16/383,019

    2020

  4. [12]

    Shiqiang Guo, Folami Alamudun, and Tracy Hammond. 2016. RésuMatcher: A personalized résumé-job matching system. Expert Systems with Applications 60 (2016), 169–182

    2016

  5. [13]

    Bing He, Sreyashi Nag, Limeng Cui, Suhang Wang, Zheng Li, Rahul Goutam, Zhen Li, and Haiyang Zhang. 2024. Hierarchical Query Classification in E- commerce Search. Companion Proceedings of the ACM on Web Conference 2024 (2024). https://api.semanticscholar.org/CorpusID:268358421

    2024

  6. [14]

    Yunzhong He, Cong Zhang, Ruoyan Kong, Chaitanya Kulkarni, Qing Liu, Ashish Gandhe, Amit Nithianandan, and Arul Prakash. 2023. HierCat: Hierarchical Query Categorization from Weakly Supervised Data at Facebook Marketplace. In Companion Proceedings of the ACM Web Conference 2023...

    2023

  7. [15]

    Xu Jin, Desheng Kong, Maoqiang Xie, Yalou Huang, Mingming Liu, Weiwei Yang, Hao Shi, and Yue Liu. 2024. Self-supervised reconstructed graph learning for link prediction in bipartite graphs. Neurocomputing 602 (2024), 128250. https: //doi.org/10.1016/j.neucom.2024.128250

    2024

  8. [16]

    Mohammed Korayem, Camilo Ortiz, Khalifeh AlJadda, and Trey Grainger. 2015. Query sense disambiguation leveraging large scale user behavioral data. In 2015 IEEE International Conference on Big Data (Big Data) . IEEE, 1230–1237

    2015

  9. [17]

    Ajay Kumar, Shashank Sheshar Singh, Kuldeep Singh, and Bhaskar Biswas. 2020. Link prediction techniques, applications, and performance: A survey. Physica A: Statistical Mechanics and its Applications 553 (2020), 124289

    2020

  10. [18]

    Jia Li, Dhruv Arya, Viet Ha-Thuc, and Shakti Sinha. 2016. How to get them a dream job? Entity-aware features for personalized job search ranking. InProceed- ings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 501–510

    2016

  11. [19]

    Ryan N Lichtenwalter, Jake T Lussier, and Nitesh V Chawla. 2010. New perspec- tives and methods in link prediction. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining . 243–252

    2010

  12. [20]

    Xianjing Liu, Hejia Zhang, Mingkuan Liu, and Alan Lu. 2019. System Design of Extreme Multi-label Query Classification using a Hybrid Model. In eCOM@SIGIR. https://api.semanticscholar.org/CorpusID:198120401

    2019

  13. [21]

    Dale T Mortensen. 1986. Job search and labor market analysis. Handbook of labor economics 2 (1986), 849–919

    1986

  14. [22]

    Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global Vectors for Word Representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Alessandro Moschitti, Bo Pang, and Walter Daelemans (Eds.). Asso...

    2014 doi

  15. [23]

    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 Proceedings of the 27th ACM international conference on informat...

    2018

  16. [24]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing . Association for Computational Linguistics. http://arxiv.org/abs/1908.10084

    2019 arXiv

  17. [25]

    Kazuya Shimura, Jiyi Li, and Fumiyo Fukumoto. 2018. HFT-CNN: Learning Hierarchical Category Structure for Multi-label Short Text Categorization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Ellen Riloff, David Chiang, Julia Hockenm...

    2018 doi

  18. [26]

    Nicole Strah, Deborah E Rupp, and Scott B Morris. 2022. Job analysis and job classification for addressing pay inequality in organizations: Adjusting our meth- ods within a shifting legal landscape. Industrial and Organizational Psychology 15, 1 (2022), 1–45

    2022

  19. [27]

    Ying Sun, Fuzhen Zhuang, Hengshu Zhu, Qing He, and Hui Xiong. 2021. Cost- Effective and Interpretable Job Skill Recommendation with Deep Reinforcement Learning. In Proceedings of the Web Conference 2021 (Ljubljana, Slovenia) (WWW ’21). Association for Computing Machinery, New ...

    2021

  20. [28]

    Jingya Wang, Kareem Abdelfatah, Mohammed Korayem, and Janani Balaji. 2019. Deepcarotene-job title classification with multi-stream convolutional neural network. In 2019 IEEE International Conference on Big Data (Big Data) . IEEE, 1953–1961

    2019

  21. [29]

    Zihan Wang, Peiyi Wang, Lianzhe Huang, Xin Sun, and Houfeng Wang. 2022. Incorporating Hierarchy into Text Encoder: a Contrastive Learning Approach for Hierarchical Text Classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (...

    2022 doi

  22. [30]

    Zihan Wang, Peiyi Wang, Tianyu Liu, Binghuai Lin, Yunbo Cao, Zhifang Sui, and Houfeng Wang. 2022. HPT: Hierarchy-aware Prompt Tuning for Hierarchical Text Classification. In Proceedings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing. Association f...

    2022 doi

  23. [31]

    Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. 2018. Hierarchical multi- label classification networks. In International conference on machine learning . PMLR, 5075–5084

    2018

  24. [32]

    Shuo Yang, Mohammed Korayem, Khalifeh AlJadda, Trey Grainger, and Sriraam Natarajan. 2017. Combining content-based and collaborative filtering for job rec- ommendation system: A cost-sensitive Statistical Relational Learning approach. Knowledge-Based Systems 136 (2017), 37–45

    2017

  25. [33]

    Muhan Zhang and Yixin Chen. 2018. Link prediction based on graph neural networks. Advances in neural information processing systems 31 (2018)

    2018

  26. [34]

    Yingya Zhang, Cheng Yang, and Zhixiang Niu. 2014. A research of job recom- mendation system based on collaborative filtering. In 2014 seventh international symposium on computational intelligence and design , Vol. 1. IEEE, 533–538

    2014

  27. [35]

    Jing Zhao, Jingya Wang, Madhav Sigdel, Bopeng Zhang, Phuong Hoang, Mengshu Liu, and Mohammed Korayem. 2021. Embedding-based recommender system for job to candidate matching on scale. arXiv preprint arXiv:2107.00221 (2021)

    2021 arXiv

  28. [36]

    Jie Zhou, Chunping Ma, Dingkun Long, Guangwei Xu, Ning Ding, Haoyu Zhang, Pengjun Xie, and Gongshen Liu. 2020. Hierarchy-Aware Global Model for Hier- archical Text Classification. In Proceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics , Da...

    2020 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.