arxiv: 2604.17411 · v1 · submitted 2026-04-19 · 💻 cs.CL · cs.AI

Recognition: unknown

DuConTE: Dual-Granularity Text Encoder with Topology-Constrained Attention for Text-attributed Graphs

Lexuan Liang , Tao Zou , Xuxiang Ta , Zekun Qiu

Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:45 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords text-attributed graphsdual-granularity encodingtopology-constrained attentionlanguage modelsgraph neural networksnode representationsattention maskssemantic correlations

0 comments

The pith

Dual-granularity text encoder uses topology-constrained attention to integrate graph structure into node text representations

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

The paper seeks to establish that language models can better encode the texts of nodes in text-attributed graphs when they incorporate information about the graph's topology during their attention computations. Current methods encode each node's text in isolation at the token level before passing features to a graph neural network, which overlooks how connected nodes' texts should influence each other semantically. DuConTE runs two language models in cascade, first processing at the word token level and then at the node level, while dynamically masking attention according to whether nodes share an edge. It further refines node embeddings by weighting tokens differently when viewed from the central node versus from its neighbors. A sympathetic reader would care because this tighter integration of structure and semantics at the encoding stage could yield more accurate results on tasks that combine text and graph data, such as classifying documents or extracting information from connected sources.

Core claim

DuConTE is a dual-granularity text encoder with topology-constrained attention that employs a cascaded architecture of two pretrained language models. It encodes semantics first at the word-token granularity and then at the node granularity. In each language model's self-attention, the attention mask matrix is dynamically adjusted based on node connectivity to guide the learning of semantic correlations informed by the graph structure. When composing node representations from word-token embeddings, the importance of tokens is evaluated separately under the center-node context and the neighborhood context.

What carries the argument

The topology-constrained attention mechanism, which dynamically adjusts the attention mask matrix according to node connectivity in the graph and separately evaluates token importance in center-node versus neighborhood contexts during representation composition.

If this is right

The model captures semantic correlations among node texts that are guided by the underlying graph structure.
Node representations incorporate both local and neighborhood textual context more effectively.
The approach achieves state-of-the-art performance on the majority of benchmark datasets for tasks on text-attributed graphs.
It provides a way to infuse structural information directly into the text encoding phase before applying graph neural networks.

Where Pith is reading between the lines

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

This method might allow simpler graph neural networks to suffice downstream since some structural info is already baked into the features.
Similar dual-context token weighting could be tested on other structured data like knowledge graphs or social networks with text.
If the dynamic masking proves robust, it could inspire attention modifications in other domains where data has both sequential and relational aspects.
An extension might involve applying this to temporal text-attributed graphs where connectivity changes over time.

Load-bearing premise

Dynamically adjusting the attention mask matrix based on node connectivity and separately evaluating token importance under center-node versus neighborhood contexts will reliably capture more contextually relevant semantic information without introducing noise or overfitting to the graph structure.

What would settle it

Running the model on the benchmark datasets but with the topology-based attention masks replaced by random or fully connected masks, and observing whether performance drops below the reported levels or matches standard language model baselines.

Figures

Figures reproduced from arXiv: 2604.17411 by Lexuan Liang, Tao Zou, Xuxiang Ta, Zekun Qiu.

**Figure 2.** Figure 2: Overview of DuConTE with the dual-granularity cascaded architecture (middle), the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Sensitive analysis of the fusion coefficient [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Sensitive analysis of the fusion coefficient [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

read the original abstract

Text-attributed graphs integrate semantic information of node texts with topological structure, offering significant value in various applications such as document classification and information extraction. Existing approaches typically encode textual content using language models (LMs), followed by graph neural networks (GNNs) to process structural information. However, during the LM-based text encoding phase, most methods not only perform semantic interaction solely at the word-token granularity, but also neglect the structural dependencies among texts from different nodes. In this work, we propose DuConTE, a dual-granularity text encoder with topology-constrained attention. The model employs a cascaded architecture of two pretrained LMs, encoding semantics first at the word-token granularity and then at the node granularity. During the self-attention computation in each LM, we dynamically adjust the attention mask matrix based on node connectivity, guiding the model to learn semantic correlations informed by the graph structure. Furthermore, when composing node representations from word-token embeddings, we separately evaluate the importance of tokens under the center-node context and the neighborhood context, enabling the capture of more contextually relevant semantic information. Extensive experiments on multiple benchmark datasets demonstrate that DuConTE achieves state-of-the-art performance on the majority of them.

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.

Desk Editor's Note private letter to a colleague

DuConTE offers a concrete cascaded LM design with graph-based attention masks and dual-context token scoring, but the abstract gives no numbers or ablations to back the SOTA claim.

read the letter

The main thing here is a two-stage text encoder that runs one pretrained LM on word tokens then another on nodes, while dynamically masking attention according to graph edges and scoring token importance once for the center node and once for its neighbors. This combination is new enough that it is not just a rehash of standard LM-plus-GNN pipelines. It does a clean job of naming the gap: most prior work lets the LM ignore topology entirely during semantic encoding and only adds structure afterward. That observation is fair and points to a practical direction for document networks and similar graphs. The soft spot is obvious and central: the abstract states that extensive experiments show SOTA results on most benchmarks, yet supplies zero quantitative details, baselines, ablations, or error bars. Without those, it is impossible to judge whether the mask adjustments and dual-context scoring actually help or whether they risk zeroing out useful semantic links between non-adjacent nodes or amplifying local artifacts. The stress-test concern about noise from the mask mechanism lands because the paper offers no internal math or derivation that guarantees the benefit. This is aimed at researchers who already work on text-attributed graphs and want to try tighter LM-graph fusion. A reader who needs reproducible evidence or clear performance deltas will get little value until the experimental section is examined. I would send it to peer review because the architectural idea is straightforward, the problem statement is honest, and the work is coherent on its own terms even if the empirical support is still missing.

Referee Report

3 major / 2 minor

Summary. The paper proposes DuConTE, a dual-granularity text encoder with topology-constrained attention for text-attributed graphs. It employs a cascaded architecture of two pretrained LMs to first encode semantics at the word-token granularity and then at the node granularity. During self-attention in each LM, the attention mask matrix is dynamically adjusted based on node connectivity to incorporate graph structure. When composing node representations, token importance is evaluated separately under center-node and neighborhood contexts. The authors report that extensive experiments on multiple benchmark datasets show state-of-the-art performance on the majority of them.

Significance. If the empirical results hold with proper controls, the approach could provide a more integrated method for combining semantic content from LMs with topological structure directly in the encoding stage, potentially improving over decoupled LM+GNN pipelines for tasks such as node classification on text-attributed graphs.

major comments (3)

Abstract: the claim of state-of-the-art performance is asserted without any quantitative results, specific baselines, ablation studies, error bars, or dataset details, so the central empirical claim cannot be evaluated from the provided description.
Topology-constrained attention mechanism: dynamically adjusting the attention mask based on node connectivity during self-attention risks zeroing out attention between semantically related but non-adjacent nodes; the manuscript provides no analysis, ablation, or mitigation strategy for this potential loss of relevant semantics.
Dual-context token importance evaluation: separately scoring tokens under center-node versus neighborhood contexts when forming node embeddings may amplify local graph artifacts or produce inconsistent representations; no details on combination method, normalization, or overfitting safeguards are supplied.

minor comments (2)

Clarify whether the two cascaded LMs share parameters, are fine-tuned jointly or sequentially, and how the node-granularity LM receives input from the token-level outputs.
Specify the exact form of the dynamic mask adjustment (e.g., additive bias, hard zeroing) and any hyper-parameters controlling its strength.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We sincerely thank the referee for the constructive and detailed feedback. We have carefully addressed each major comment and revised the manuscript to strengthen the presentation of results, add necessary analyses, and provide missing technical details. Point-by-point responses follow.

read point-by-point responses

Referee: Abstract: the claim of state-of-the-art performance is asserted without any quantitative results, specific baselines, ablation studies, error bars, or dataset details, so the central empirical claim cannot be evaluated from the provided description.

Authors: We agree that the abstract should better support the SOTA claim. In the revised version, we will update the abstract to include specific quantitative gains (e.g., average accuracy improvements), name the key baselines, and reference the benchmark datasets. Full experimental details with error bars and ablations remain in Section 5 and Tables 2-4; we will add an explicit cross-reference from the abstract to these sections. revision: yes
Referee: Topology-constrained attention mechanism: dynamically adjusting the attention mask based on node connectivity during self-attention risks zeroing out attention between semantically related but non-adjacent nodes; the manuscript provides no analysis, ablation, or mitigation strategy for this potential loss of relevant semantics.

Authors: This concern is valid. While our results show performance benefits from the topology-constrained mask, the original manuscript lacks explicit analysis of attention between non-adjacent nodes. In the revision, we will add an ablation comparing masked vs. unmasked attention, include attention visualization or statistics demonstrating preservation of relevant semantics via the node-granularity stage, and discuss a hybrid masking strategy as a potential mitigation. revision: yes
Referee: Dual-context token importance evaluation: separately scoring tokens under center-node versus neighborhood contexts when forming node embeddings may amplify local graph artifacts or produce inconsistent representations; no details on combination method, normalization, or overfitting safeguards are supplied.

Authors: We thank the referee for this observation. The manuscript describes the dual-context scoring but does not detail the combination method, normalization, or safeguards. In the revision, we will provide the exact formulation (weighted combination of the two scores), specify normalization (e.g., softmax), and add ablation experiments in Section 5.3 to demonstrate robustness, including variance across graph densities and regularization effects to address potential artifacts or overfitting. revision: yes

Circularity Check

0 steps flagged

No significant circularity: architectural proposal with external empirical validation

full rationale

The paper presents DuConTE as a new cascaded LM architecture with topology-constrained attention masks and dual-context token importance scoring. No derivation chain, equations, or first-principles predictions are described that reduce to fitted inputs or self-citations. The central claims rest on benchmark experiments rather than internal self-definition or renamed known results. This is the expected non-finding for an empirical architecture paper whose validity is externally testable.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper introduces an architectural design rather than new mathematics; no free parameters, axioms, or invented entities are defined or fitted in the provided abstract.

pith-pipeline@v0.9.0 · 5525 in / 1043 out tokens · 50570 ms · 2026-05-10T05:45:37.362097+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

141 extracted references · 14 canonical work pages · 4 internal anchors

[1]

Weilin Cong and Si Zhang and Jian Kang and Baichuan Yuan and Hao Wu and Xin Zhou and Hanghang Tong and Mehrdad Mahdavi , title =
[2]

Srijan Kumar and Xikun Zhang and Jure Leskovec , title =
[3]

Rakshit Trivedi and Mehrdad Farajtabar and Prasenjeet Biswal and Hongyuan Zha , title =
[4]

Inductive representation learning on temporal graphs , booktitle =

Da Xu and Chuanwei Ruan and Evren K. Inductive representation learning on temporal graphs , booktitle =
[5]

Bronstein , title =

Emanuele Rossi and Ben Chamberlain and Fabrizio Frasca and Davide Eynard and Federico Monti and Michael M. Bronstein , title =. CoRR , volume =
[6]

Inductive Representation Learning in Temporal Networks via Causal Anonymous Walks , booktitle =

Yanbang Wang and Yen. Inductive Representation Learning in Temporal Networks via Causal Anonymous Walks , booktitle =
[7]

NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , year =

Farimah Poursafaei and Shenyang Huang and Kellin Pelrine and Reihaneh Rabbany , title =. NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , year =

2022
[8]

CoRR , volume =

Lu Wang and Xiaofu Chang and Shuang Li and Yunfei Chu and Hui Li and Wei Zhang and Xiaofeng He and Le Song and Jingren Zhou and Hongxia Yang , title =. CoRR , volume =
[9]

AI magazine , volume=

Collective classification in network data , author=. AI magazine , volume=
[10]

and Cangea, C

A wikipedia-based benchmark for graph neural networks. arXiv 2020 , author=. arXiv preprint arXiv:2007.02901 , year=

work page arXiv 2020
[11]

Proceedings of the third ACM conference on Digital libraries , pages=

CiteSeer: An automatic citation indexing system , author=. Proceedings of the third ACM conference on Digital libraries , pages=
[12]

NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

Le Yu and Leilei Sun and Bowen Du and Weifeng Lv , title =. NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

2023
[13]

Aditya Grover and Jure Leskovec , title =
[14]

Le Yu and Zihang Liu and Leilei Sun and Bowen Du and Chuanren Liu and Weifeng Lv , title =
[15]

Tao Zou and Le Yu and Leilei Sun and Bowen Du and Deqing Wang and Fuzhen Zhuang , title =
[16]

Yongjie Huang and Meng Yang , title =
[17]

Guanhao Hou and Xingguang Chen and Sibo Wang and Zhewei Wei , title =. Proc
[18]

RecSys '21, The Netherlands, 27 September 2021 - 1 October 2021 , pages =

Alexander Dallmann and Daniel Zoller and Andreas Hotho , title =. RecSys '21, The Netherlands, 27 September 2021 - 1 October 2021 , pages =

2021
[19]

Performance-Adaptive Sampling Strategy Towards Fast and Accurate Graph Neural Networks , booktitle =

Minji Yoon and Th. Performance-Adaptive Sampling Strategy Towards Fast and Accurate Graph Neural Networks , booktitle =
[20]

Bronstein and Haggai Maron , title =

Fabrizio Frasca and Beatrice Bevilacqua and Michael M. Bronstein and Haggai Maron , title =. NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , year =

2022
[21]

MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing , booktitle =

Sami Abu. MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing , booktitle =
[22]

Networks beyond pairwise interactions: structure and dynamics , journal =

Federico Battiston and Giulia Cencetti and Iacopo Iacopini and Vito Latora and Maxime Lucas and Alice Patania and Jean. Networks beyond pairwise interactions: structure and dynamics , journal =
[23]

Harrington and Michael T

Christian Bick and Elizabeth Gross and Heather A. Harrington and Michael T. Schaub , title =
[24]

Motif-based Graph Self-Supervised Learning for Molecular Property Prediction , booktitle =

Zaixi Zhang and Qi Liu and Hao Wang and Chengqiang Lu and Chee. Motif-based Graph Self-Supervised Learning for Molecular Property Prediction , booktitle =
[25]

Maciej Besta and Raphael Grob and Cesare Miglioli and Nicola Bernold and Grzegorz Kwasniewski and Gabriel Gjini and Raghavendra Kanakagiri and Saleh Ashkboos and Lukas Gianinazzi and Nikoli Dryden and Torsten Hoefler , title =
[26]

Xingyi Zhang and Shuliang Xu and Wenqing Lin and Sibo Wang , title =
[27]

Yu , title =

Xixi Wu and Yun Xiong and Yao Zhang and Yizhu Jiao and Caihua Shan and Yiheng Sun and Yangyong Zhu and Philip S. Yu , title =
[28]

Seyed Mehran Kazemi and Rishab Goel and Kshitij Jain and Ivan Kobyzev and Akshay Sethi and Peter Forsyth and Pascal Poupart , title =. J. Mach. Learn. Res. , volume =
[29]

Claudio Daniel Tenorio de Barros and Matheus R. F. Mendon. A Survey on Embedding Dynamic Graphs , journal =
[30]

Dynamic Graph Representation Learning with Neural Networks:

Leshanshui Yang and S. Dynamic Graph Representation Learning with Neural Networks:. CoRR , volume =
[31]

Schardl and Charles E

Aldo Pareja and Giacomo Domeniconi and Jie Chen and Tengfei Ma and Toyotaro Suzumura and Hiroki Kanezashi and Tim Kaler and Tao B. Schardl and Charles E. Leiserson , title =
[32]

Jiaxuan You and Tianyu Du and Jure Leskovec , title =
[33]

Aravind Sankar and Yanhong Wu and Liang Gou and Wei Zhang and Hao Yang , title =
[34]

Learning on large-scale text- attributed graphs via variational inference,

Learning on large-scale text-attributed graphs via variational inference , author=. arXiv preprint arXiv:2210.14709 , year=

work page arXiv
[35]

Kaike Zhang and Qi Cao and Gaolin Fang and Bingbing Xu and Hongjian Zou and Huawei Shen and Xueqi Cheng , title =
[36]

Rossi and Nesreen K

Giang Hoang Nguyen and John Boaz Lee and Ryan A. Rossi and Nesreen K. Ahmed and Eunyee Koh and Sungchul Kim , title =
[37]

Aggarwal and Kai Zhang and Haifeng Chen and Wei Wang , title =

Wenchao Yu and Wei Cheng and Charu C. Aggarwal and Kai Zhang and Haifeng Chen and Wei Wang , title =
[38]

Neural Temporal Walks: Motif-Aware Representation Learning on Continuous-Time Dynamic Graphs , booktitle =

Ming Jin and Yuan. Neural Temporal Walks: Motif-Aware Representation Learning on Continuous-Time Dynamic Graphs , booktitle =
[39]

Yanhong Wu and Nan Cao and Daniel Archambault and Qiaomu Shen and Huamin Qu and Weiwei Cui , title =
[40]

Xin Liu and Mingyu Yan and Lei Deng and Guoqi Li and Xiaochun Ye and Dongrui Fan , title =
[41]

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Sentence-bert: Sentence embeddings using siamese bert-networks , author=. arXiv preprint arXiv:1908.10084 , year=

work page internal anchor Pith review arXiv 1908
[42]

Hamilton and Zhitao Ying and Jure Leskovec , title =

William L. Hamilton and Zhitao Ying and Jure Leskovec , title =. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA,

2017
[43]

Hamilton and Jure Leskovec , title =

Rex Ying and Ruining He and Kaifeng Chen and Pong Eksombatchai and William L. Hamilton and Jure Leskovec , title =
[44]

Jianfei Chen and Jun Zhu and Le Song , title =
[45]

Jie Chen and Tengfei Ma and Cao Xiao , title =
[46]

SubMix: Learning to Mix Graph Sampling Heuristics , booktitle =

Sami Abu. SubMix: Learning to Mix Graph Sampling Heuristics , booktitle =
[47]

Rong Chen and Jiaxin Shi and Yanzhe Chen and Binyu Zang and Haibing Guan and Haibo Chen , title =
[48]

Prasanna , title =

Hanqing Zeng and Hongkuan Zhou and Ajitesh Srivastava and Rajgopal Kannan and Viktor K. Prasanna , title =
[49]

Liangzhe Han and Xiaojian Ma and Leilei Sun and Bowen Du and Yanjie Fu and Weifeng Lv and Hui Xiong , title =
[50]

Libing Wu and Min Wang and Dan Wu and Jia Wu , title =
[51]

Xiaofu Chang and Xuqin Liu and Jianfeng Wen and Shuang Li and Yanming Fang and Le Song and Yuan Qi , title =
[52]

CoRR , volume =

Yifang Qin and Wei Ju and Hongjun Wu and Xiao Luo and Ming Zhang , title =. CoRR , volume =
[53]

Yanping Zheng and Hanzhi Wang and Zhewei Wei and Jiajun Liu and Sibo Wang , title =
[54]

Representation Learning in Continuous-Time Dynamic Signed Networks , booktitle =

Kartik Sharma and Mohit Raghavendra and Yeon. Representation Learning in Continuous-Time Dynamic Signed Networks , booktitle =
[55]

Mozhdeh Ariannezhad and Sami Jullien and Ming Li and Min Fang and Sebastian Schelter and Maarten de Rijke , title =
[56]

Predicting Music Relistening Behavior Using the

Markus Reiter. Predicting Music Relistening Behavior Using the. RecSys '21, 27 September 2021 - 1 October 2021 , pages =

2021
[57]

Proceedings of the Twenty-Ninth

Jun Chen and Chaokun Wang and Jianmin Wang , title =. Proceedings of the Twenty-Ninth
[58]

Senrong Xu and Liangyue Li and Yuan Yao and Zulong Chen and Han Wu and Quan Lu and Hanghang Tong , title =
[59]

CoRR , volume =

Andres Ferraro and Dmitry Bogdanov and Kyumin Choi and Xavier Serra , title =. CoRR , volume =
[60]

Dumais , title =

Eytan Adar and Jaime Teevan and Susan T. Dumais , title =
[61]

Personal Web Revisitation by Context and Content Keywords with Relevance Feedback , journal =

Li Jin and Ling Feng and Gang. Personal Web Revisitation by Context and Content Keywords with Relevance Feedback , journal =
[62]

Junqi Zhang and Yiqun Liu and Jiaxin Mao and Xiaohui Xie and Min Zhang and Shaoping Ma and Qi Tian , title =
[63]

Ashton Anderson and Ravi Kumar and Andrew Tomkins and Sergei Vassilvitskii , title =
[64]

Hui Fang and Danning Zhang and Yiheng Shu and Guibing Guo , title =
[65]

Modeling Personalized Item Frequency Information for Next-basket Recommendation , booktitle =

Haoji Hu and Xiangnan He and Jinyang Gao and Zhi. Modeling Personalized Item Frequency Information for Next-basket Recommendation , booktitle =
[66]

Ilya O. Tolstikhin and Neil Houlsby and Alexander Kolesnikov and Lucas Beyer and Xiaohua Zhai and Thomas Unterthiner and Jessica Yung and Andreas Steiner and Daniel Keysers and Jakob Uszkoreit and Mario Lucic and Alexey Dosovitskiy , title =. NeurIPS 2021, December 6-14, 2021 , pages =

2021
[67]

Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun , title =
[68]

Hinton , title =

Lei Jimmy Ba and Jamie Ryan Kiros and Geoffrey E. Hinton , title =. CoRR , volume =
[69]

CoRR , volume =

Dan Hendrycks and Kevin Gimpel , title =. CoRR , volume =
[70]

Bronstein and Guillaume Rabusseau and Reihaneh Rabbany , title =

Shenyang Huang and Farimah Poursafaei and Jacob Danovitch and Matthias Fey and Weihua Hu and Emanuele Rossi and Jure Leskovec and Michael M. Bronstein and Guillaume Rabusseau and Reihaneh Rabbany , title =. CoRR , volume =
[71]

CoRR , volume =

Le Yu , title =. CoRR , volume =
[72]

9th International Conference on Learning Representations,

Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby , title =. 9th International Conference on Learning Representations,
[73]

Andreas Steiner and Alexander Kolesnikov and Xiaohua Zhai and Ross Wightman and Jakob Uszkoreit and Lucas Beyer , title =. Trans. Mach. Learn. Res. , volume =
[74]

HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions , booktitle =

Yongming Rao and Wenliang Zhao and Yansong Tang and Jie Zhou and Ser. HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions , booktitle =
[75]

The Twelfth International Conference on Learning Representations,

Xue Wang and Tian Zhou and Qingsong Wen and Jinyang Gao and Bolin Ding and Rong Jin , title =. The Twelfth International Conference on Learning Representations,
[76]

Muyao Wang and Wenchao Chen and Bo Chen , title =
[77]

NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

Zelin Ni and Hang Yu and Shizhan Liu and Jianguo Li and Weiyao Lin , title =. NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

2023
[78]

Wipf and Junchi Yan , title =

Qitian Wu and Wentao Zhao and Zenan Li and David P. Wipf and Junchi Yan , title =. NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , year =

2022
[79]

Kim , title =

Seongjun Yun and Minbyul Jeong and Raehyun Kim and Jaewoo Kang and Hyunwoo J. Kim , title =. NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada , pages =

2019
[80]

Borgwardt , title =

Dexiong Chen and Leslie O'Bray and Karsten M. Borgwardt , title =

Showing first 80 references.