Gated Multimodal Graph Learning for Personalized Recommendation

Chengwei Feng; Hao Yang; Sibei Liu; Xiang Li; Yuanzhe Zhang; Yunbo Liu

REVIEW 3 major objections 6 minor 1 cited by

Gated Multimodal Graph Learning for Personalized Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A gated fusion of image and text embeddings, paired with a two-layer LightGCN, outperforms six recommendation baselines on the Amazon Clothing dataset.

desk verdict The gated-fusion-plus-LightGCN design is plausible, but the random pseudo-timestamp split invalidates the temporal claim and the reported gains are not consistent, so the central empirical contribution fails. read the letter →

arxiv 2506.00107 v1 pith:NATXQB3A submitted 2025-05-30 cs.IR cs.AI

Sibei Liu , Yuanzhe Zhang , Xiang Li , Yunbo Liu , Chengwei Feng , Hao Yang This is my paper

classification cs.IRcs.AI

keywords multimodalrecommendationgatedfusionLightGCNcollaborativefilteringtop-KevaluationAmazonClothingdatasetleave-one-outimplicitfeedback

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 proposes RLMultimodalRec, a lightweight recommender that combines a gated fusion of image and text item features with a two-layer LightGCN, a graph encoder that propagates embeddings over the user-item interaction graph without nonlinear transformations. Its central claim is that this modular separation, content-aware item encoding plus collaborative user encoding, outperforms matrix factorization, visual-aware, and multimodal GNN baselines on the Amazon Clothing, Shoes and Jewelry dataset. The gated fusion is the load-bearing novelty: it weights each modality per item and per embedding dimension rather than concatenating or averaging them. If the claim is right, multimodal recommendation can be effective without heavy modality-specific graph encoders, which matters for scalability and for deploying content-aware recommenders in practice.

What carries the argument

The central mechanism is the gated fusion module. For each item $i$, projected visual and textual embeddings $v_i^{\mathrm{img}}$ and $v_i^{\mathrm{txt}}$ are combined by a gate $g_i = \sigma(W_g[v_i^{\mathrm{img}}; v_i^{\mathrm{txt}}] + b_g)$, giving $z_i = g_i \odot v_i^{\mathrm{img}} + (1 - g_i) \odot v_i^{\mathrm{txt}}$, so each dimension of the fused item vector can lean toward image or text. A two-layer LightGCN propagates user and item ID embeddings over the bipartite interaction graph with symmetric degree normalization, and the final score is the dot product $s_{ui} = e_u^\top z_i$. Training uses binary cross-entropy with online negative sampling.

What would settle it

Recover real timestamps from the original Amazon Review data, rebuild the leave-one-out split by true chronological order, and check whether RLMultimodalRec's Recall@20 advantage over LightGCN persists; if the advantage vanishes or changes sign, the central performance claim depends on the arbitrary pseudo-timestamp split.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a dimension-wise gated combination of visual and textual item embeddings, scored against LightGCN-updated user embeddings, achieves Recall@20 of 0.0996 and NDCG@10 of 0.0285 on the Amazon Clothing dataset, outperforming MF-BPR, LightGCN, LayerGCN, VBPR, MMGCN, and DualGNN under the same leave-one-out protocol. The authors attribute the improvement to the gate's ability to emphasize the more informative modality for each item and to the clean separation between content encoding and collaborative message passing.

Load-bearing premise

The evaluation rests on treating random integers as pseudo-timestamps; if those random values do not reproduce the user's true chronological order, the held-out item is not a genuine next interaction and the reported gains are measured against an arbitrary split.

Editorial extensions

If this is right

A per-item, per-dimension gate can replace fixed concatenation or averaging when fusing multimodal item content.
Two LightGCN layers are sufficient to carry collaborative signals; deeper or dual-path graph encoders are not needed for the reported gains.
Keeping content encoding at the item level and graph propagation on ID embeddings makes the model modular and easier to interpret.
The full model can be trained with implicit feedback using binary cross-entropy and online negative sampling, with no auxiliary losses.

Reading between the lines

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

The pseudo-timestamps used for the leave-one-out split are random integers, so the held-out item is a genuine next interaction only by assumption; the reported gains should be interpreted as gains on an arbitrary split until confirmed with true temporal ordering.
A natural stress test the paper does not run is to mask one modality at inference and check that the gate shifts weight to the remaining modality.
The gate vectors themselves could be inspected per item category to see whether the model learns interpretable modality preferences, such as relying on images for clothing appearance and text for materials.

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.

Desk Editor's Note

read the letter

The short version: this is a simple gated-fusion plus two-layer LightGCN recommender, and the paper's central empirical claim—significant, consistent top-K gains on a next-interaction task—does not survive contact with the evaluation section.

What the paper does well: separating content encoding from graph propagation is a clean, sensible design choice, and the model itself is straightforward and easy to understand. The gating mechanism (Eqs. 3–4) is standard but reasonable, and the authors are right that lightweight fusion can beat more complex multimodal architectures. On the numbers they report, the model does beat most baselines on Recall@20 and NDCG@10.

The soft spots are not minor. Section 3.1 says the dataset has no timestamps, so the authors assign random integers as pseudo-timestamps and sort by them to build a "temporally consistent" leave-one-out split. That is not a temporal split. It is a random split, and the held-out item is not the user's next interaction. Every claim about next-interaction performance is therefore unsupported as presented. The stress-test note is correct, and it lands squarely on the paper's own text.

The results also undercut the "consistently outperforms" claim: Table 2 shows LayerGCN beats the proposed model on Recall@10 and NDCG@20. The paper says each experiment was repeated with three seeds, but no error bars or variance are reported. The policy network described in Section 3.7 is never actually used for scoring—Eqs. 7–8 discard it—so it is dead weight in the architecture. The discussion section is padded with unrelated finance, warehouse robotics, and UX references, and parts of Section 6 are duplicated almost verbatim. No code or data are provided.

I'd give credit for the core modeling idea: it is plausible that a per-dimension gate over fixed image and text features, combined with LightGCN, gives a modest boost over pure collaborative filtering on this kind of data. But the evaluation as written cannot support that conclusion. The temporal framing is broken, the head-to-head comparisons are selective, and the missing error bars make it impossible to judge whether the remaining differences are noise.

This is a desk reject for me. The right fix would be to either obtain real timestamps or drop the temporal claim entirely, add error bars, run on more than one dataset, and release code. If that were done, the model might become a modest but useful datapoint for the multimodal recommendation subfield. As it stands, the paper is not ready for referee time.

Referee Report

3 major / 6 minor

Summary. The paper proposes RLMultimodalRec, a multimodal recommendation model that combines a gated fusion of pre-extracted image and text features into item embeddings with a two-layer LightGCN that propagates user and item ID embeddings over the interaction graph. The model is trained with binary cross-entropy under dynamic negative sampling and evaluated on the Amazon Clothing, Shoes, and Jewelry dataset using a leave-one-out split that the authors describe as temporally consistent. The central claim is that this lightweight, modular design consistently outperforms collaborative filtering, visual-aware, and multimodal GNN baselines on Recall@10/20 and NDCG@10/20.

Significance. If the reported results were reliable, the paper would offer a simple and interpretable baseline for multimodal graph-based recommendation: a gated item-level fusion with two-layer LightGCN user modeling is a plausible design that could be useful for practitioners. The manuscript is clearly written in its architecture sections, and the authors honestly acknowledge limitations such as fixed pre-extracted features and candidate-set evaluation. However, the empirical contribution is not currently supported. The evaluation split is invalid for the claimed temporal task because the pseudo-timestamps are random, the reported numbers contradict the paper's own 'consistently outperforms' statement, no variance or significance testing is provided, and the policy network described in the architecture is not used in the training objective. These issues are load-bearing for the central claim of significant improvement, so the significance of the contribution cannot be assessed until they are addressed.

major comments (3)

[Section 3.1] The leave-one-out split is not temporally consistent. The paper states that the dataset has no timestamps and that the authors 'synthetically generate pseudo-timestamps by assigning random integers to each interaction' and then sort by these integers to hold out the 'most recent' interaction per user. Random integers carry no real chronological information, so the held-out item is effectively a uniformly random interaction per user. Consequently, the evaluation does not test the claimed 'next interaction' prediction task described in Section 3.1 and Section 5.1. This is a load-bearing flaw: the entire empirical comparison, including the abstract's claim of significant top-K improvements, rests on a test distribution that does not match the stated deployment scenario. The authors should either re-run the evaluation with a genuinely timestamped version of the dataset or explicitly reframe the paper as a non-temporal top-K recommendation study and remove all temporal claims.
[Section 5.4, Table 2] The statement that 'our model consistently outperforms all baselines on Recall@20 and NDCG@10' is contradicted by the paper's own reported numbers. In Table 2, LayerGCN achieves Recall@10 = 0.0529 and NDCG@20 = 0.0355, while RLMultimodalRec achieves Recall@10 = 0.0505 and NDCG@20 = 0.0341. Thus the proposed model is worse than LayerGCN on two of the four reported metrics. Additionally, the text claims 'significant improvements' without reporting standard deviations, confidence intervals, or significance tests, even though Section 5.1 says each experiment is repeated with three random seeds and the results are averaged. At minimum, the authors need to correct the overclaim, report per-seed variance, and provide a proper significance test for any claim of improvement.
[Sections 3.7 and 4.1] The policy network is not actually used in the reported training or inference procedure. Section 3.7 defines a two-layer policy network that outputs a score vector in Eq. (6), and then says the matching score is computed 'using either the policy net output or a dot product' (Eq. 7). However, Section 4.1 defines the training loss in Eq. (9) solely in terms of the dot product score in Eq. (8), with no loss term for the policy network. No experiment or ablation explains whether Eq. (6) is trained, used at inference, or simply unused. This ambiguity undermines the 'reinforcement-inspired' framing and makes the architecture description incomplete; the authors should clarify the role of the policy network or remove it from the model description.

minor comments (6)

[Section 3.1] The term 'pseudo-timestamps' is misleading; since the values are random, the sorted order is not a chronological order. The authors should replace all temporal language in Sections 3.1, 5.1, and 5.4 with language appropriate for a random split, or use real timestamps.
[Section 3.6 and 3.7] The LightGCN propagation in Eq. (5) updates both user and item ID embeddings, but the final scoring in Eq. (7) uses only the GCN-updated user embedding and the non-propagated fused item embedding z_i. The propagated item embeddings do not enter the score. If this is intentional, the item-side graph computation is wasted and should be removed for clarity; if not, the model description is inconsistent.
[Section 4.2 and Table 1] The maximum number of training epochs is inconsistent: Section 4.2 says 'a maximum of 150 epochs' while Table 1 lists 'Training epochs 100'. Please reconcile these values.
[Section 5.4] The sentence 'our method shows a substantial gain (+32' is truncated and incomplete. In addition, the Discussion section repeats an entire block of text verbatim, which should be removed.
[Section 5.1 and References] There are unresolved placeholders such as 'MF-BPR [?]' and 'Section ??', and the reference for MENTOR [31] has a malformed author list. Several cited papers are unrelated to the technical content of this manuscript (e.g., IMU sensor validation, warehouse robot scheduling, thyroid disease prediction); these citations should be either integrated into the narrative or removed.
[Section 5.1 and 5.4] The evaluation uses only 100 randomly sampled negative items per user. This is a standard protocol, but the authors should explicitly state that the reported Recall and NDCG are for the candidate set, not the full item catalog, and discuss the potential bias this introduces when comparing methods with different item representations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the random pseudo-timestamp split is an evaluation-validity concern, not a circular reduction.

full rationale

The paper's contribution is empirical. The derivation from Eq. (1) to Eq. (9) is a trainable architecture: modality projections, a per-dimension gate, two-layer LightGCN aggregation, and a dot-product scoring function. Every trainable quantity (projection weights, gate parameters, user/item embeddings) is learned from the interaction data and evaluated against external baselines (MF-BPR, LightGCN, LayerGCN, VBPR, MMGCN, DualGNN). No parameter is fitted to a subset of the data and then reported as a prediction of a closely related quantity; in particular, Eq. (7) is a model definition, not a fitted result renamed as a prediction. The LightGCN component is imported from external work [6] and the preprocessing protocol from MENTOR [31], so these citations carry independent support. The few self-citations (e.g., SETransformer [11] in Related Work) are motivational and do not justify the performance claim. The pseudo-timestamp procedure in Section 3.1 is a serious evaluation-validity flaw: random integers cannot make the leave-one-out split temporally consistent, so the 'next interaction' framing is unsupported. However, this is not circularity in the derivation chain—the model's scores do not reduce to the pseudo-timestamps. Section 6 also openly lists limitations (fixed features, binary labels, small candidate sets), which are acknowledged weaknesses rather than hidden circular steps. Overall, no derivation step reduces by construction to its own inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard learned model components rather than new physical entities. The so-called policy network of Section 3.7 is described but never used in the scoring equation, so it is a documentation artifact rather than an entity with independent evidence.

free parameters (5)

embedding_dimension = 64
This dimension is chosen by hand in Section 4.2 and Table 1, with no sensitivity analysis reported.
gcn_layers = 2
The number of GCN layers is set to two in Section 3.6 and Table 1, and this choice is central to the architecture.
negative_sampling_ratio = 1:1
The ratio of negative to positive samples is fixed at one to one in Table 1, which affects the BCE training objective.
learning_rate = 0.001
The learning rate is set to 0.001 for the Adam optimizer in Section 4.2, with no tuning curve shown.
batch_size = 256
The batch size is set to 256 in Section 4.2, and no ablation studies are provided.

assumptions (4)

domain assumption 5-core filtering yields a representative interaction distribution for evaluating recommenders.
Section 3.1 applies 5-core filtering without analyzing its effect on the reported metrics.
domain assumption Pre-extracted 4096-dimensional CNN image features and 384-dimensional Sentence Transformer text features are sufficient and fixed representations of item content.
Sections 3.1 and 3.4 rely on these frozen features, but the pretrained encoders are never identified.
ad hoc to paper Random synthetic timestamps can substitute for real chronological order in leave-one-out evaluation.
Section 3.1 assigns random integers to interactions and calls the split temporally consistent, which is an unjustified and likely false assumption.
domain assumption Binary cross-entropy loss with online uniform negative sampling and evaluation with 100 sampled negatives measures top-K ranking performance.
Sections 4.1 and 5.1 adopt this protocol with no analysis of ranking metric bias from candidate sampling.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Gated Multimodal Graph Learning for Personalized Recommendation." pith.science (2026). https://pith.science/paper/NATXQB3A

@misc{pith2026250600107,
  author       = {Pith},
  title        = {Pith review of: Gated Multimodal Graph Learning for Personalized Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NATXQB3A}},
  note         = {Machine review of arXiv:2506.00107}
}

read the original abstract

Multimodal recommendation has emerged as a promising solution to alleviate the cold-start and sparsity problems in collaborative filtering by incorporating rich content information, such as product images and textual descriptions. However, effectively integrating heterogeneous modalities into a unified recommendation framework remains a challenge. Existing approaches often rely on fixed fusion strategies or complex architectures , which may fail to adapt to modality quality variance or introduce unnecessary computational overhead. In this work, we propose RLMultimodalRec, a lightweight and modular recommendation framework that combines graph-based user modeling with adaptive multimodal item encoding. The model employs a gated fusion module to dynamically balance the contribution of visual and textual modalities, enabling fine-grained and content-aware item representations. Meanwhile, a two-layer LightGCN encoder captures high-order collaborative signals by propagating embeddings over the user-item interaction graph without relying on nonlinear transformations. We evaluate our model on a real-world dataset from the Amazon product domain. Experimental results demonstrate that RLMultimodalRec consistently outperforms several competitive baselines, including collaborative filtering, visual-aware, and multimodal GNN-based methods. The proposed approach achieves significant improvements in top-K recommendation metrics while maintaining scalability and interpretability, making it suitable for practical deployment.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Predicting Asphalt Pavement Friction Using Texture-Based Image Indicator
cs.CV 2025-07 conditional novelty 4.0 of 10

The paper defines aggregate protrusion area from 2D pavement images as a texture indicator and reports linear models with adjusted R2 above 0.90 for predicting DFT friction for three asphalt surface types.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages · cited by 1 Pith paper

[1]

Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation

Boris Baˇ ci´ c, Chengwei Feng, and Weihua Li. Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation. ISBS Proceedings Archive, 42(1):60, 2024

work page 2024
[2]

Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study

Boris Baˇ ci´ c, Claudiu V asile, Chengwei Feng, and Marian G Ciuc˘ a. Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study. arXiv preprint arXiv:2412.20733, 2024

arXiv 2024
[3]

A gan-based method to tune lstm hyperparameters for ﬁnancial forecasting

ADNANE EL OUARDI, BRAHIM ER-RAHA, MUSTAPHA RIAD, and KHALID TA TANE. A gan-based method to tune lstm hyperparameters for ﬁnancial forecasting. Journal of Theoretical and Applied Information T echnology, 103(9), 2025

work page 2025
[4]

Yi Fu, Yingzhou Lu, Yizhi Wang, Bai Zhang, Zhen Zhang, Guoqiang Y u, Chunyu Liu, Robert Clarke, David M Herrington, and Y ue Wang. Ddn3. 0: Determin- ing signiﬁcant rewiring of biological network structure with differential dependency networks. Bioinformatics, 40(6):btae376, 2024

work page 2024
[5]

Vbpr: visual bayesian personalized ranking from implicit feedback

Ruining He and Julian McAuley. Vbpr: visual bayesian personalized ranking from implicit feedback. In Pro- ceedings of the AAAI conference on artiﬁcial intelli- gence, volume 30, 2016

work page 2016
[6]

Lightgcn: Simplifying and powering graph convolution network for recommen- dation

Xiangnan He, Kuan Deng, Xiang Wang, Y an Li, Y ong- dong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommen- dation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Infor- mation Retrieval, pages 639–648, 2020

work page 2020
[7]

The impact of apple’s digital design on its success: An analysis of interaction and interface de- sign

Wanxin Li. The impact of apple’s digital design on its success: An analysis of interaction and interface de- sign. Academic Journal of Sociology and Management , 2(4):14–19, 2024

work page 2024
[8]

Transforming logistics with innovative inte r- action design and digital ux solutions

Wanxin Li. Transforming logistics with innovative inte r- action design and digital ux solutions. Journal of Com- puter T echnology and Applied Mathematics, 1(3):91–96, 2024

work page 2024

Show all 33 references

[9]

A con- trastive deep learning approach to cryptocurrency portfo- lio with us treasuries

Zichao Li, Bingyang Wang, and Ying Chen. A con- trastive deep learning approach to cryptocurrency portfo- lio with us treasuries. Journal of Computer T echnology and Applied Mathematics , 1(3):1–10, 2024

2024
[10]

Knowledge graph embedding and few-shot relational learning meth- ods for digital assets in usa

Zichao Li, Bingyang Wang, and Ying Chen. Knowledge graph embedding and few-shot relational learning meth- ods for digital assets in usa. Journal of Industrial Engi- neering and Applied Science , 2(5):10–18, 2024

2024
[11]

Setransformer: A hybrid attention-based ar- chitecture for robust human activity recognition

Y unbo Liu, Xukui Qin, Yifan Gao, Xiang Li, and Cheng- wei Feng. Setransformer: A hybrid attention-based ar- chitecture for robust human activity recognition. arXiv preprint arXiv:2505.19369, 2025

2025 arXiv
[12]

Cot: an efﬁcient and accurate method for detecting marker genes among many subtypes

Yingzhou Lu, Chiung-Ting Wu, Sarah J Parker, Zuolin Cheng, Georgia Saylor, Jennifer E V an Eyk, Guoqiang Y u, Robert Clarke, David M Herrington, and Y ue Wang. Cot: an efﬁcient and accurate method for detecting marker genes among many subtypes. Bioinformatics Ad- vances, 2(1):...

2022
[13]

Knowledge graph for query enrichment in retrieval augmented generation in domain speciﬁc ap- plication

Massimo Perna. Knowledge graph for query enrichment in retrieval augmented generation in domain speciﬁc ap- plication. Master’s thesis, University of Twente, 2025

2025
[14]

Convex optimization of markov decision pro- cesses based on z transform: A theoretical framework for two-space decomposition and linear programming recon- struction

Shiqing Qiu, Haoyu Wang, Y uxin Zhang, Zong Ke, and Zichao Li. Convex optimization of markov decision pro- cesses based on z transform: A theoretical framework for two-space decomposition and linear programming recon- struction. Mathematics, 13(11), 2025. 8 Journal of Emerging...

2025
[15]

Bpr: Bayesian person- alized ranking from implicit feedback

Steffen Rendle, Christoph Freudenthaler, Zeno Gant- ner, and Lars Schmidt-Thieme. Bpr: Bayesian person- alized ranking from implicit feedback. arXiv preprint arXiv:1205.2618, 2012

2012 arXiv
[16]

Evaluat- ing supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data, 2025

Chao Wang, Chuanhao Nie, and Y unbo Liu. Evaluat- ing supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data, 2025

2025
[17]

Mari: Mate- rial retrieval integration across domains

Jianhui Wang, Zhifei Y ang, Y angfan He, Huixiong Zhang, Y uxuan Chen, and Jingwei Huang. Mari: Mate- rial retrieval integration across domains. arXiv preprint arXiv:2503.08111, 2025

2025 arXiv
[18]

Dualgnn: Dual graph neural network for multimedia recommendation

Qifan Wang, Yinwei Wei, Jianhua Yin, Jianlong Wu, Xuemeng Song, and Liqiang Nie. Dualgnn: Dual graph neural network for multimedia recommendation. IEEE Transactions on Multimedia, 25:1074–1084, 2021

2021
[19]

A sys- tematic review of machine learning applications in in- fectious disease prediction, diagnosis, and outbreak fore - casting

Yiting Wang, Jiachen Zhong, and Rohan Kumar. A sys- tematic review of machine learning applications in in- fectious disease prediction, diagnosis, and outbreak fore - casting. 2025

2025
[20]

Study of artiﬁcial intelligence for visual defect inspection in industrial products

Y uxuan Wang et al. Study of artiﬁcial intelligence for visual defect inspection in industrial products. 2025

2025
[21]

Mmgcn: Multi- modal graph convolution network for personalized rec- ommendation of micro-video

Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. Mmgcn: Multi- modal graph convolution network for personalized rec- ommendation of micro-video. In Proceedings of the 27th ACM international conference on multimedia, pages 1437–1445, 2019

2019
[22]

Warehouse robot task scheduling based on reinforcement learning to maximize operational efﬁciency

Siye Wu, Lei Fu, Runmian Chang, Y uanzhou Wei, Y eyubei Zhang, Zehan Wang, Lipeng Liu, Haopeng Zhao, and Keqin Li. Warehouse robot task scheduling based on reinforcement learning to maximize operational efﬁciency. Authorea Preprints, 2025

2025
[23]

Advancing unsupervised graph anomaly detection: A multi-level contrastive learning framework to mitigate local consistency decep- tion

Wensen Wu and Yijun Gu. Advancing unsupervised graph anomaly detection: A multi-level contrastive learning framework to mitigate local consistency decep- tion. Neurocomputing, page 130507, 2025

2025
[24]

Multisage: Empowering gcn with contextualized multi- embeddings on web-scale multipartite networks

Carl Y ang, Aditya Pal, Andrew Zhai, Nikil Pancha, Jiawei Han, Charles Rosenberg, and Jure Leskovec. Multisage: Empowering gcn with contextualized multi- embeddings on web-scale multipartite networks. In Pro- ceedings of the 26th ACM SIGKDD international con- ference on knowle...

2020
[25]

Interpretable credit default prediction with ensem- ble learning and shap, 2025

Shiqi Y ang, Ziyi Huang, Wengran Xiao, and Xinyu Shen. Interpretable credit default prediction with ensem- ble learning and shap, 2025

2025
[26]

Machine learning optimizes the ef- ﬁciency of picking and packing in automated warehouse robot systems

Dezhi Y u, Lipeng Liu, Siye Wu, Keqin Li, Congyu Wang, Jing Xie, Runmian Chang, Yixu Wang, Zehan Wang, and Ryan Ji. Machine learning optimizes the ef- ﬁciency of picking and packing in automated warehouse robot systems. In 2025 IEEE International Conference on Electronics, Ene...

2025
[27]

Multi-scale video super-resolution transformer with polynomial approximation

Fan Zhang, Gongguan Chen, Hua Wang, Jinjiang Li, and Caiming Zhang. Multi-scale video super-resolution transformer with polynomial approximation. IEEE Transactions on Circuits and Systems for Video T echnol- ogy, 33(9):4496–4506, 2023

2023
[28]

Optimization and application of cloud-based deep learning architecture for multi-source data prediction, 2024

Y ang Zhang, Fa Wang, Xin Huang, Xintao Li, Sibei Liu, and Hansong Zhang. Optimization and application of cloud-based deep learning architecture for multi-source data prediction, 2024

2024
[29]

The role of machine learning in reducing healthcare costs: The impact of medication adherence and preventive care on hospital- ization expenses, 2025

Yixin Zhang and Yisong Chen. The role of machine learning in reducing healthcare costs: The impact of medication adherence and preventive care on hospital- ization expenses, 2025

2025
[30]

Contex- tual bandits for unbounded context distributions, 2025

Puning Zhao, Rongfei Fan, Shaowei Wang, Li Shen, Qixin Zhang, Zong Ke, and Tianhang Zheng. Contex- tual bandits for unbounded context distributions, 2025

2025
[31]

Mentor: Multi-level self-supervised learning for multimodal recommendation

Xiaotian Zhao, Xia Hu Jin, Fuli Feng Sun, and Tat-Seng Chua. Mentor: Multi-level self-supervised learning for multimodal recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2023

2023
[32]

Enhancing thyroid disease prediction using machine learning: A compar- ative study of ensemble models and class balancing tech- niques

Jiachen Zhong and Yiting Wang. Enhancing thyroid disease prediction using machine learning: A compar- ative study of ensemble models and class balancing tech- niques. 2025

2025
[33]

Layer-reﬁned graph convolutional networks for recom- mendation

Xin Zhou, Donghui Lin, Y ong Liu, and Chunyan Miao. Layer-reﬁned graph convolutional networks for recom- mendation. In 2023 IEEE 39th international conference on data engineering (ICDE) , pages 1247–1259. IEEE, 2023. 9

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation

Boris Baˇ ci´ c, Chengwei Feng, and Weihua Li. Jy61 imu sensor external validity: A framework for advanced pe- dometer algorithm personalisation. ISBS Proceedings Archive, 42(1):60, 2024

work page 2024

[2] [2]

Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study

Boris Baˇ ci´ c, Claudiu V asile, Chengwei Feng, and Marian G Ciuc˘ a. Towards nation-wide analytical healthcare infrastructures: A privacy-preserving aug- mented knee rehabilitation case study. arXiv preprint arXiv:2412.20733, 2024

arXiv 2024

[3] [3]

A gan-based method to tune lstm hyperparameters for ﬁnancial forecasting

ADNANE EL OUARDI, BRAHIM ER-RAHA, MUSTAPHA RIAD, and KHALID TA TANE. A gan-based method to tune lstm hyperparameters for ﬁnancial forecasting. Journal of Theoretical and Applied Information T echnology, 103(9), 2025

work page 2025

[4] [4]

Yi Fu, Yingzhou Lu, Yizhi Wang, Bai Zhang, Zhen Zhang, Guoqiang Y u, Chunyu Liu, Robert Clarke, David M Herrington, and Y ue Wang. Ddn3. 0: Determin- ing signiﬁcant rewiring of biological network structure with differential dependency networks. Bioinformatics, 40(6):btae376, 2024

work page 2024

[5] [5]

Vbpr: visual bayesian personalized ranking from implicit feedback

Ruining He and Julian McAuley. Vbpr: visual bayesian personalized ranking from implicit feedback. In Pro- ceedings of the AAAI conference on artiﬁcial intelli- gence, volume 30, 2016

work page 2016

[6] [6]

Lightgcn: Simplifying and powering graph convolution network for recommen- dation

Xiangnan He, Kuan Deng, Xiang Wang, Y an Li, Y ong- dong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommen- dation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Infor- mation Retrieval, pages 639–648, 2020

work page 2020

[7] [7]

The impact of apple’s digital design on its success: An analysis of interaction and interface de- sign

Wanxin Li. The impact of apple’s digital design on its success: An analysis of interaction and interface de- sign. Academic Journal of Sociology and Management , 2(4):14–19, 2024

work page 2024

[8] [8]

Transforming logistics with innovative inte r- action design and digital ux solutions

Wanxin Li. Transforming logistics with innovative inte r- action design and digital ux solutions. Journal of Com- puter T echnology and Applied Mathematics, 1(3):91–96, 2024

work page 2024