Pith. sign in

REVIEW 4 major objections 5 minor 55 references

A Scalable and Efficient Signal Integration System for Job Matching

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 Feeding fine-tuned LLM embeddings into a graph neural network raises job-matching accuracy across three products.

desk verdict A credible industrial case study of LLM-in-GNN integration with strong online A/B evidence, but the offline AUC gain needs a clearer temporal split to rule out label leakage. read the letter →

arxiv 2507.09797 v1 pith:2IDLBOHY submitted 2025-07-13 cs.LG

Ping Liu , Rajat Arora , Xiao Shi , Benjamin Le , Qianqi Shen , Jianqiang Shen , Chengming Jiang , Nikita Zhiltsov
show 8 more authors
Priya Bannur Yidan Zhu Liming Dong Haichao Wei Qi Guo Luke Simon Liangjie Hong Wenjing Zhang
This is my paper
classification cs.LG
keywords LLMembeddingsgraphneuralnetworksjobmatchingrecommendersystemslinkpredictionfeatureintegrationindustrialdeploymentA/Btesting
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 tries to establish that large language model embeddings and graph neural networks are complementary at industrial scale: the LLM supplies semantic understanding of long job postings and member profiles, while the GNN supplies relational structure from interaction histories. The authors built the STAR system to feed fine-tuned LLM embeddings into a large GNN as node features rather than co-training the two models, which they argue is cheaper and more scalable. Offline, adding LLM embeddings raised validation AUC (area under the ROC curve) from 0.8447 to 0.8489 for job recommendation and search. Online A/B tests across three products showed reported metric gains—more job applications, fewer dismissals, and more recruiter responses—all stated as statistically significant. The broader claim is that this decoupled LLM-plus-GNN architecture is a reusable template for other relevance systems.

What carries the argument

The load-bearing mechanism is the decoupling of the LLM and the GNN. The LLM is trained separately as a bi-encoder: one tower encodes job postings, the other encodes member profiles and resumes, with a binary cross-entropy plus semi-hard triplet contrastive loss on apply/no-apply pairs, using LoRA fine-tuning and gradient checkpointing to fit long sequences. Its frozen 4096-dimensional output embeddings are then concatenated into the GNN node features alongside ID and categorical features, and the GNN is trained with multi-task link-prediction losses over a graph of interaction edges (apply, save, InMail reply) and attribute edges (title, skill, company, geo). Two engineering mechanisms carry the scalability claims: adaptive neighbor sampling, which grows the sample count only when validation stops improving and cut a 48-hour training job to 36 hours, and backward-compatible embedding versioning, which learns a linear map from a new embedding version to the old one so downstream retraining can be postponed.

What would settle it

A decisive check would be to retrain the LLM embeddings on a temporally earlier slice of activity data, then train and evaluate the GNN only on interactions that occurred after that slice, so no apply event ever appears in both the embedding training set and the GNN test labels. If the validation AUC gain from adding LLM embeddings disappears or reverses under this strict time split, the paper's central claim is not supported. A cheaper observational version: count how many of the GNN's positive test edges correspond to member-job pairs that were used as positive pairs when fine-tuning the LLM.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that pre-computed embeddings from a fine-tuned 7-billion-parameter text model, when concatenated into the node features of a heterogeneous graph neural network trained for link prediction, improve job-matching quality over a GNN that uses categorical features and legacy two-tower embeddings. The authors call this the GNN+LLM Embedding strategy and report that it outperforms the alternatives they compared: freezing a small text encoder inside the GNN, co-training the encoder and GNN, or encoding graph structure into an LLM prompt. The gains appear across three products—job recommendation and search, recruiter InMail messages, and top applicant jobs—with offline AUC rising from 0.8447 to 0.8489 and online metrics moving in the expected direction for all three. They also claim that LLM embeddings can substitute for a large set of hand-maintained categorical features with comparable accuracy and lower maintenance burden.

Load-bearing premise

The claim rests on the assumption that LLM embeddings trained on apply/no-apply activity contribute fresh signal to the GNN rather than leaking the same labels that define the GNN's training edges.

Editorial extensions

If this is right

  • If the reported gains are real, job recommendation systems can adopt this decoupled architecture without re-engineering the ranking model to host a generative LLM.
  • LLM embeddings can replace many taxonomy-based categorical features, cutting the maintenance burden of title, skill, and industry dictionaries while keeping accuracy.
  • Co-training or freezing a text encoder inside a GNN is not the right operating point at scale; the authors' measurements put it at roughly ten times the training time or worse with no accuracy gain.
  • Backward-compatible embedding versioning lets embedding upgrades ship without forcing immediate retraining of all downstream models.
  • The online tests reinforce that small but statistically significant lifts in matching metrics are achievable together with lower operational cost.

Reading between the lines

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

  • A direct extension would test whether the same recipe transfers to other two-sided marketplaces with long textual profiles, such as housing or freelance markets; the paper's graph construction is general, but its evidence comes from one platform.
  • Because the LLM and the GNN are trained on the same apply/no-apply labels, the offline AUC gain might partly reflect label sharing rather than genuine semantic generalization; a temporally disjoint retraining of the LLM would settle this.
  • The online metric shifts (+1.5% applies, -3.2% dismiss-to-apply ratio, +3.0% InMail acceptance) are small relative to the roughly 0.4-point offline AUC change, suggesting that modest embedding quality gains can still move business metrics at platform scale.
  • One implicit consequence of the backward-compatibility scheme is that embeddings become a versioned data product with a compatibility contract, making cross-team feature engineering more like infrastructure management than model research.
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 presents STAR, LinkedIn's system that combines fine-tuned LLM embeddings with a large-scale heterogeneous GNN for job matching and related talent products. The LLM component is a bi-encoder built on Mistral-7B, fine-tuned with binary cross-entropy plus a contrastive loss on member activity data; the GNN component uses a multi-task link-prediction framework over a graph with hundreds of millions of nodes and billions of edges, consuming the LLM embeddings as node features. The authors report offline validation AUC gains for the GNN (from 0.8447 to 0.8489 when LLM embeddings are added), along with online A/B test improvements across three products, including +1.5% total applies, -3.2% dismiss-to-apply ratio, +3.0% InMail acceptance, and +2.4% positive hearing-back. The paper also describes engineering contributions such as adaptive sampling, embedding version management with backward compatibility, and deployment infrastructure details.

Significance. If the results hold, the paper is a useful industrial case study: it demonstrates a practical recipe for training and serving long-context LLM embeddings with modest GPU resources, a scalable GNN-LLM integration, and concrete operational lessons (adaptive sampling, versioned embedding lifecycles, and latency/throughput trade-offs). The explicit resource numbers and infrastructure descriptions are valuable for practitioners. However, the empirical evidence is currently not fully convincing: the offline comparison lacks uncertainty quantification, the online tests are only summarized with p-value thresholds, and a potential temporal overlap between LLM training data and GNN test labels is not addressed. The backward-compatibility contribution is claimed but not evaluated in the paper. These issues need to be resolved before the central claims can be fully credited.

major comments (4)
  1. [Section 3.3 and Section 5.1.2, Table 4] The paper does not rule out label leakage between the LLM training data and the GNN test edges. Section 3.3 states that the LLM is fine-tuned on member activities with positive labels for job applications and negatives from skips/clicks, using a 2-day training window and 4.5M samples, but it does not state whether these training pairs are temporally or entity-wise disjoint from the 15-day GNN test period described in Section 5.1.2. Since the GNN graph contains APPLY edges constructed from the same kind of activity data, and the LLM embeddings are frozen node features fed into the GNN, any member-job pair appearing in both the LLM training set and the GNN test set could make the embedding encode the exact label the GNN is asked to predict. The reported +0.0042 AUC improvement (Table 4) could then reflect memorization rather than generalizable semantic understanding. The sentence in Section 3.3 about solving 'time-travel issues' with text snapshots addresses text-version staleness, not label overlap. Please clarify the temporal split of the LLM training data relative to the GNN test window, or provide an ablation in which LLM embeddings are trained only on data preceding the GNN test period.
  2. [Table 4 and Section 5.1.2] The offline AUC results are reported as single point estimates with no error bars, confidence intervals, or significance tests. The headline difference between the baseline (0.8447) and the model with LLM embeddings (0.8489) is 0.0042 absolute, which is small enough that without replication or uncertainty quantification it is unclear whether the difference is meaningful or within run-to-run noise. The paper states that all hyper-parameters were held constant, but this does not account for random seed variation, initialization, or sampling stochasticity. Please report multiple runs with standard deviations or bootstrap confidence intervals, and state whether the observed differences are statistically significant.
  3. [Section 5.2, Tables 5-7] The online A/B test descriptions are too thin to assess the claimed business impact. The paper only reports that a two-tailed t-test gave p < 0.05, with no effect-size confidence intervals, sample sizes, variance estimates, or information about how many metrics were tested and whether multiple-testing corrections were applied. For example, Table 5 reports '+1.5% Total Applies' as a site-wide metric in a 50/50 split, but it does not state the unit of analysis, the length of the pre-period used for normalization, or whether the treatment and control groups were well balanced. Given that these online results are a central part of the paper's evidence, please provide a more detailed experimental protocol, including confidence intervals for the reported relative ratios and a clear statement of the primary versus secondary metrics.
  4. [Section 4.6] The backward-compatible embedding mechanism is presented as a key contribution, but the paper provides no evaluation for it. Section 4.6 states that 'Both offline and online A/B tests demonstrate that the backward-compatible embedding achieves parity results without any degradation in metrics,' yet no tables, numbers, or test descriptions are given anywhere in the paper, including the appendix. This is an unsupported load-bearing claim for the version-management contribution. Please add the actual parity results, including the offline metric and the online A/B setup, or clearly mark this as future work.
minor comments (5)
  1. [Section 2 and Section 4.3] The citation for GraphSAGE appears to be incorrect: Section 4.3 says 'Our framework supports GraphSAGE [20]', but reference [20] is Kipf and Welling's GCN paper, while the GraphSAGE paper is reference [12]. Please fix the citation.
  2. [Equations (1)-(4)] The notation in the contrastive loss is imprecise: the summation index i in Equation (3) is not used inside the summand, and z_d, z_r, and z_r^- are not indexed by i, so it is unclear whether they are per-sample embeddings or batch-level aggregates. Please clarify the indexing and the role of the batch B in Equation (4).
  3. [Table 2] The reported percentage improvements in Table 2 are ambiguous: 'BCE +4.5%' appears to be 0.0445 absolute AUC points (0.6886 to 0.7331), which is a relative improvement of about 6.5%, not 4.5%. Please state explicitly whether all changes are absolute or relative.
  4. [Section 5.1.2] The text says 'We used the full training and testing datasets described above for training and validation in each epoch', which is confusing because it suggests the test set is used for validation. Please distinguish the training, validation, and test splits, and report results on the held-out test set rather than only validation AUC.
  5. [Section 5.3] Section 1 promises a discussion of limitations in the final part of Section 5, but Section 5.3 does not contain an explicit limitations paragraph. Please add one, or revise the introduction's roadmap to match the actual structure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STAR is an empirical system paper whose performance claims rest on measured offline and online evaluations, not on a derivation that reduces to its inputs.

full rationale

The paper's central claim is that adding fine-tuned LLM embeddings as node features to a large GNN improves job-matching metrics. This is presented as an experimental result, not as a derived prediction from first principles. No equation in the paper defines the GNN's link-prediction target as a function of the LLM embedding by construction; the LLM is trained separately with a BCE-plus-contrastive objective (Eq. 1), and the GNN is trained with a cross-entropy link-prediction objective (Eq. 6). The offline gains in Table 4 are measured AUC values over a validation set, not quantities forced by the model definitions. The backward-compatible embedding transform in Eq. 8 is a learned linear mapping evaluated for parity, which is a model component rather than a circular prediction. The paper does cite prior work by overlapping authors (e.g., [3], [24], [31]), but none of these citations carries the load of the paper's main empirical claim, and no uniqueness theorem or forced-choice argument is imported from those works. The most plausible concern is label leakage: the LLM is trained on member activities that include apply/no-apply labels, and the GNN graph is built from the same kind of activity, so the offline AUC improvement could in principle reflect overlap between LLM training pairs and GNN test edges. However, that is a data-contamination validity risk, not a circular-reasoning reduction. The paper does not state a temporal or entity-wise split for the LLM training data, and it reports no significance intervals for Table 4; these are correctness risks that should be checked, but they do not make the argument circular by construction. The online A/B tests over future traffic provide an independent, external check on the system's effectiveness. Accordingly, no circular step can be exhibited from the paper's text, and the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's claims are empirical; there is no mathematical derivation. The main unstated assumptions are the transferability of LLM embeddings across tasks and the reliability of the graph and online metrics. Hyperparameters are tuned but reported without full details, and no new theoretical entities are introduced.

free parameters (6)
  • temperature tau = not reported
    Temperature parameter in the contrastive loss (Eq. 3); chosen by hand but value not given.
  • lambda (loss weight) = not reported
    Weight balancing BCE and contrastive losses in Eq. 1; tuned but not specified.
  • GNN encoder dimension = 200
    Selected based on experiments showing positive correlation in range 50 to 200.
  • LLM effective batch size = 3172
    Chosen after observing consistent improvements up to this value; further increases did not help.
  • max context length = 1800 (jobs/resumes), 1024 (profiles)
    Set based on memory constraints on A100 GPUs.
  • LoRA rank = 8
    Used for fine-tuning; no ablation is provided.
assumptions (4)
  • domain assumption The bi-encoder with average pooling over all layer CLS tokens produces embeddings sufficient for downstream GNN tasks.
    Section 3.1: the embedding is defined as the average of <CLS> token across all layers with L2 normalization; this is a design choice, not proven optimal.
  • domain assumption LLM embeddings fine-tuned on job application data transfer to the three downstream products (Job Recommendation, Recruiter InMail, Top Applicant).
    Section 3.3: training data is member job applications; embeddings are then used as node features for all three tasks, assuming transferability.
  • domain assumption The heterogeneous graph from six months of LinkedIn activity captures sufficient relational information for link prediction.
    Section 4.1 and Appendix: graph with 763M nodes and 12.3B edges is used; no analysis of graph completeness or sampling bias.
  • domain assumption Multi-task learning with a shared GNN encoder and weighted loss (Eq. 7) does not cause task interference that harms individual task performance.
    Section 4.5: multi-task losses are combined with tuned weights, but no analysis of gradient conflicts or task interference is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Scalable and Efficient Signal Integration System for Job Matching." pith.science (2026). https://pith.science/paper/2IDLBOHY

@misc{pith2026250709797,
  author       = {Pith},
  title        = {Pith review of: A Scalable and Efficient Signal Integration System for Job Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IDLBOHY}},
  note         = {Machine review of arXiv:2507.09797}
}
read the original abstract

LinkedIn, one of the world's largest platforms for professional networking and job seeking, encounters various modeling challenges in building recommendation systems for its job matching product, including cold-start, filter bubbles, and biases affecting candidate-job matching. To address these, we developed the STAR (Signal Integration for Talent And Recruiters) system, leveraging the combined strengths of Large Language Models (LLMs) and Graph Neural Networks (GNNs). LLMs excel at understanding textual data, such as member profiles and job postings, while GNNs capture intricate relationships and mitigate cold-start issues through network effects. STAR integrates diverse signals by uniting LLM and GNN capabilities with industrial-scale paradigms including adaptive sampling and version management. It provides an end-to-end solution for developing and deploying embeddings in large-scale recommender systems. Our key contributions include a robust methodology for building embeddings in industrial applications, a scalable GNN-LLM integration for high-performing recommendations, and practical insights for real-world model deployment.

Figures

Figures reproduced from arXiv: 2507.09797 by the authors.

Figure 1
Figure 1. The illustration of the STAR workflow. The figure’s left side depicts the LLM workflow, which generates embeddings for member profiles, resumes, and job descriptions. The GNN then processes these embeddings during the signal integration stage to support LinkedIn’s downstream products. Job Seeker Job Posting Preimium JobPosting Recruiter Recruiter Learning Course Learning Course Appl y Cl i c k w ATCH Save Repl y Ema… view at source ↗
Figure 2
Figure 2. An example of LinkedIn’s heterogeneous graph. [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

55 extracted references · 31 canonical work pages

  1. [1]

    Ethem Alpaydin. 2020. Introduction to machine learning . MIT press

    work page 2020

  2. [2]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems . 1007–1014

    work page 2023

  3. [3]

    Fedor Borisyuk, Shihai He, Yunbo Ouyang, Morteza Ramezani, Peng Du, Xiaochen Hou, Chengming Jiang, Nitin Pasumarthy, Priya Bannur, et al. 2024. Lignn: Graph neural networks at linkedin. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4793–4803

    work page 2024

  4. [4]

    Martin Juan José Bucher and Marco Martini. 2024. Fine-Tuned ’Small’ LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models in Text Classification. arXiv:2406.08660 [cs.CL]

    arXiv 2024

  5. [5]

    Gal Chechik, Varun Sharma, Uri Shalit, and Samy Bengio. 2010. Large scale online learning of image similarity through ranking. Journal of Machine Learning Research 11, 3 (2010). KDD ’25, August 3–7, 2025, Toronto, ON, Canada Ping Liu et al. Table 9: Performance and resource comparison of GNN+LLM training configurations Text Encoder ModelCo-train Max Seque...

    work page 2010

  6. [6]

    Nurendra Choudhary, Edward W Huang, Karthik Subbian, and Chandan K Reddy

    work page

  7. [7]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research 25, 70 (2024), 1–53

    2024

  8. [8]

    Andreas Damianou, Francesco Fabbri, Paul Gigioli, Marco De Nadai, Alice Wang, Enrico Palumbo, and Mounia Lalmas. 2024. Towards Graph Foundation Models for Personalization. In Companion Proceedings of the ACM on Web Conference

    work page 2024

Show all 55 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human...

    2019

  2. [10]

    Ahmed El-Kishky, Thomas Markovich, Serim Park, Chetan Verma, Baekjin Kim, Ramy Eskander, Yury Malkov, Frank Portman, Sofía Samaniego, Ying Xiao, et al

  3. [11]

    Bahare Fatemi, Jonathan Halcrow, and Bryan Perozzi. 2023. Talk like a Graph: Encoding Graphs for Large Language Models. In NeurIPS 2023 Workshop: New Frontiers in Graph Learning

    2023

  4. [12]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  5. [13]

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards universal sequence representation learning for recommender systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Dis- covery and Data Mining . 585–593

    2022

  6. [14]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. In European Conference on Information Retrieval. Springer, 364–381

    2024

  7. [15]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  8. [16]

    Weihua Hu, Rajas Bansal, Kaidi Cao, Nikhil Rao, Karthik Subbian, and Jure Leskovec. 2022. Learning backward compatible embeddings. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3018–3028

    2022

  9. [17]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 1857–1867

    2020

  10. [18]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thoma...

    2023 arXiv

  11. [19]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

  12. [20]

    Thomas N Kipf and Max Welling. 2016. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  13. [21]

    Chaozhuo Li, Bochen Pang, Yuming Liu, Hao Sun, Zheng Liu, Xing Xie, Tianqi Yang, Yanling Cui, Liangjie Zhang, and Qi Zhang. 2021. Adsgnn: Behavior-graph augmented relevance modeling in sponsored search. In Proceedings of the 44th international ACM SIGIR conference on research ...

    2021

  14. [22]

    Ping Liu, Karthik Shivaram, Aron Culotta, Matthew Shapiro, and Mustafa Bilgic

  15. [23]

    Ping Liu, Karthik Shivaram, Aron Culotta, Matthew A Shapiro, and Mustafa Bilgic. 2021. The interaction between political typology and filter bubbles in news recommendation algorithms. In Proceedings of the Web Conference 2021 . 3791–3801

    2021

  16. [24]

    Ping Liu, Haichao Wei, Xiaochen Hou, Jianqiang Shen, Shihai He, Kay Qianqi Shen, Zhujun Chen, Fedor Borisyuk, Daniel Hewlett, Liang Wu, et al . 2024. LinkSAGE: Optimizing Job Matching Using Graph Neural Networks. arXiv preprint arXiv:2402.13430 (2024)

    2024 arXiv

  17. [25]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. In Pro- ceedings of the ACM Web Conference 2023 . 417–428

    2023

  18. [26]

    In Proceedings of the International AAAI Conference on Web and Social Media, Vol

    How Does Empowering Users with Greater System Control Affect News Filter Bubbles?. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 18. 943–957

  19. [27]

    Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. 2022. PEFT: State-of-the-art Parameter-Efficient Fine-Tuning methods. https://github.com/huggingface/peft

    2022

  20. [28]

    Phu Pham, Loan TT Nguyen, Witold Pedrycz, and Bay Vo. 2023. Deep learning, graph-based text representation and classification: a survey, perspectives and challenges. Artificial Intelligence Review 56, 6 (2023), 4893–4927

    2023

  21. [29]

    N Reimers. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT- Networks. arXiv preprint arXiv:1908.10084 (2019)

    2019 arXiv

  22. [30]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  23. [31]

    Jianqiang Shen, Yuchin Juan, Ping Liu, Wen Pu, Shaobo Zhang, Qianqi Shen, Liangjie Hong, and Wenjing Zhang. 2024. Learning Links for Adaptable and Explainable Retrieval. In Proceedings of the 33rd CIKM . 4046–4050

    2024

  24. [32]

    Kyuyong Shin, Hanock Kwak, Kyung-Min Kim, Minkyu Kim, Young-Jin Park, Jisu Jeong, and Seungjae Jung. 2021. One4all user representation for recommender systems in e-commerce. arXiv preprint arXiv:2106.00573 (2021)

    2021 arXiv

  25. [33]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568 (2024), 127063. LinkedIn STAR System KDD ’25, August 3–7, 2025, Toronto, ON, Canada

    2024

  26. [34]

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. FaceNet: A Unified Embedding for Face Recognition and Clustering. CoRR abs/1503.03832 (2015). arXiv:1503.03832 http://arxiv.org/abs/1503.03832

    2015 arXiv

  27. [35]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  28. [36]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In International Con- ference on Learning Representations

    2018

  29. [37]

    Binbin Wang, Mingming Li, Zhixiong Zeng, Jingwei Zhuo, Songlin Wang, Sulong Xu, Bo Long, and Weipeng Yan. 2023. Learning Multi-Stage Multi-Grained Semantic Embeddings for E-Commerce Search. In Companion Proceedings of the ACM Web Conference 2023. 411–415

    2023

  30. [38]

    Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2021. Augmented SBERT: Data Augmentation Method for Improving Bi-Encoders for Pairwise Sentence Scoring Tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Comp...

    2021

  31. [39]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672 (2024)

    2024 arXiv

  32. [40]

    Songhao Wu, Quan Tu, Hong Liu, Jia Xu, Zhongyi Liu, Guannan Zhang, Ran Wang, Xiuying Chen, and Rui Yan. 2024. Unify Graph Learning with Text: Unleashing LLM Potentials for Session Search. In Proceedings of the ACM on Web Conference 2024. 1509–1518

    2024

  33. [41]

    Han Xie, Da Zheng, Jun Ma, Houyu Zhang, Vassilis N Ioannidis, Xiang Song, Qing Ping, Sheng Wang, Carl Yang, Yi Xu, et al. 2023. Graph-aware language model pre-training on a large graph corpus can help multiple graph applications. In Proceedings of the 29th ACM SIGKDD Conferenc...

    2023

  34. [42]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving Text Embeddings with Large Language Models.arXiv preprint arXiv:2401.00368 (2023)

    2023 arXiv

  35. [43]

    Yuhao Yang, Lianghao Xia, Da Luo, Kangyi Lin, and Chao Huang. 2024. GraphPro: Graph Pre-training and Prompt Learning for Recommendation. In Proceedings of the ACM on Web Conference 2024 . 3690–3699

    2024

  36. [44]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 974–983

    2018

  37. [45]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. MultiGPrompt for multi-task pre-training and prompting on graphs. In Proceedings of the ACM on Web Conference 2024. 515–526

    2024

  38. [46]

    Junhan Yang, Zheng Liu, Shitao Xiao, Chaozhuo Li, Defu Lian, Sanjay Agrawal, Amit Singh, Guangzhong Sun, and Xing Xie. 2021. Graphformers: Gnn-nested transformers for representation learning on textual graph. Advances in Neural Information Processing Systems 34 (2021), 28798–28810

    2021

  39. [47]

    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

  40. [48]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI open 1 (2020), 57–81

    2020

  41. [49]

    Jason Zhu, Yanling Cui, Yuming Liu, Hao Sun, Xue Li, Markus Pelger, Tianqi Yang, Liangjie Zhang, Ruofei Zhang, and Huasha Zhao. 2021. TextGNN: Improving Text Encoder via Graph Neural Network in Sponsored Search. In Proceedings of the Web Conference. Association for Computing M...

    2021

  42. [50]

    Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. 2021. Decoupling the depth and scope of graph neural networks. Advances in Neural Information Processing Systems 34 (2021), 19665–19679

    2021

  43. [51]

    Yaochen Zhu, Liang Wu, Binchi Zhang, Song Wang, Qi Guo, Liangjie Hong, Luke Simon, and Jundong Li. 2024. Understanding and Modeling Job Marketplace with Pretrained Language Models. arXiv preprint arXiv:2408.04381 (2024)

    2024 arXiv

  44. [54]

    Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collab- orative large language model for recommender systems. In Proceedings of the ACM on Web Conference 2024. 3162–3172

    2024

  45. [2020]

    CoRR abs/2001.08361 (2020)

    Scaling Laws for Neural Language Models. CoRR abs/2001.08361 (2020). arXiv:2001.08361

    2020 arXiv

  46. [2022]

    In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining

    Twhin: Embedding the twitter heterogeneous information network for personalized recommendation. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining . 2842–2850

  47. [2024]

    In Companion Proceedings of the ACM on Web Conference 2024

    An interpretable ensemble of graph and language models for improving search relevance in e-commerce. In Companion Proceedings of the ACM on Web Conference 2024. 206–215

    2024

Pith tools

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