Pith. sign in

REVIEW 4 major objections 6 minor 17 references

LRA-GNN: Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual for Facial Age Estimation

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

Pith's one-line read The paper claims that a graph neural network which captures latent key-point relations and co-optimizes age estimation with reinforcement learning surpasses state-of-the-art baselines on four facial age benchmarks.

desk verdict A real but incremental GNN architecture for age estimation whose 'surpasses SOTA' claim is directly contradicted by its own Table 1. read the letter →

arxiv 2502.05423 v1 pith:JF7CEM2Q submitted 2025-02-08 cs.CV

classification cs.CV
keywords facialageestimationgraphneuralnetworklatentrelationrandomwalkresidualconvolutionreinforcementlearningkeypointsmulti-headattention
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 argues that facial age estimation can be improved by explicitly discovering latent relationships between facial key points instead of relying only on similarity-thresholded graph edges. It introduces LRA-GNN, a graph neural network that builds an initial face graph from key-point priors, enriches it with a random-walk global structure, and uses multi-head attention to produce fully connected graphs that capture overlooked relations such as a wrinkle connecting to a distant key point. A deep residual graph convolutional stack with adaptive initial and dynamic developmental residuals prevents over-smoothing, and a progressive reinforcement-learning stage co-optimizes age-group classification and regression. The authors report state-of-the-art or competitive mean absolute errors across MORPH II, FG-NET, ChaLearn LAP 2016, and UTK-Face with 13M parameters.

What carries the argument

The load-bearing object is the latent relation-aware fully connected graph. Starting from a key-point-initialized graph, a random walk updates the adjacency matrix with global structural evidence; multi-head self-attention then re-weights all pairwise connections, producing a set of fully connected graphs in which no potential relation is thresholded away. Deep feature extraction runs through residual GCN layers that fuse an adaptive initial residual term with a dynamic developmental residual term, both designed to counteract over-smoothing. Age estimation is framed as a Markov decision process on a grid whose rows are age groups and columns are within-group ages, solved with Double DQN; the reward function includes an imbalance ratio and label-distance terms.

What would settle it

Run LRA-GNN and the three baselines it trails on MORPH II Setting I under identical pretraining data, split, and preprocessing; if the MAE gap reverses or disappears, the paper's superiority claim fails. If public code reproduces the reported MAEs under the stated protocols, the claim stands.

Watch

Extended reading notes

Core claim

The central claim is that modeling faces as graphs with complete, attention-generated connections and residual-equipped deep convolution yields a compact age estimator that outperforms established CNN and Transformer baselines. Concretely, LRA-GNN reports MAE 1.79 on MORPH II Setting I and 1.94 on Setting II with external pre-training, 2.14 on FG-NET, 3.11 on ChaLearn LAP 2016, and 4.22 on UTK-Face, using about 13M parameters. The authors attribute the gains to three interacting mechanisms: random-walk guidance followed by multi-head attention that captures latent key-point relations, deep residual GCNs that preserve node individuality, and reinforcement learning that treats age estimation as a walk-to-the-end grid problem and jointly optimizes classification and regression with an imbalance-aware reward.

Load-bearing premise

The comparisons in Tables 1 to 4 assume the benchmark protocols are comparable, even though methods differ in external pretraining, backbone, and train/test splits; if those differences are not controlled, the reported state-of-the-art advantage could be an artifact of the evaluation setup rather than the architecture.

Editorial extensions

If this is right

  • If correct, a graph-based face representation can reach top-tier age-estimation accuracy at a fraction of the parameter count of VGG-16 or ViT-based systems.
  • Capturing latent relations beyond similarity-thresholded edges is a measurable source of accuracy: the ablation shows Latent Relation Capturing contributes the largest single gain, e.g., 0.28 MAE on ChaLearn LAP 2016.
  • Co-optimizing age-group classification and regression through reinforcement learning beats decoupled training; adding the imbalance ratio and label distance improves all four benchmarks, with the largest gain of 0.25 MAE under MORPH II Setting II.
  • Fusing adaptive initial and dynamic developmental residuals lets GCNs go deeper on fully connected graphs without over-smoothing; depth 12 is chosen as the balance between performance and parameter overhead.
  • The efficiency analysis indicates competitive runtime (141.96 ms on UTK-Face) with fewer parameters than the CNN and Transformer baselines, suggesting the architecture is practical to deploy.

Reading between the lines

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

  • Beyond the paper's chosen splits, the latent-relation mechanism should transfer to other facial attribute tasks, such as expression recognition or action-unit detection, where key-point relations are similarly informative and the graph construction is not age-specific.
  • A direct protocol-matched comparison against the three baselines it trails on MORPH II Setting I, under identical pretraining data, split, and preprocessing, would clarify whether the residual MAE differences reflect the architecture or the evaluation setup.
  • The random-walk-plus-attention pipeline can be seen as a cheap way to approximate global relational structure; a natural testable extension is replacing the hand-set thresholds (0.936, 0.824) with a learned edge predictor.
  • The reinforcement-learning grid formulation is a general template for any ordinal regression task with natural grouping, so applying it to other ordinal labels, such as health scores or quality ratings, is a plausible extension not explored in the paper.
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 / 6 minor

Summary. The manuscript proposes LRA-GNN, a graph neural network for facial age estimation. The method constructs an initial graph from facial key points, uses a random walk to capture global structure, applies multi-head attention to generate fully connected graphs that capture latent relations, and uses deep residual graph convolution with adaptive initial and dynamic developmental residuals to avoid over-smoothing. Age estimation is framed as a progressive reinforcement learning problem on a grid, optimized with a combined focal and MAE loss. The paper reports MAE values on MORPH II, FG-NET, ChaLearn LAP 2016, and UTK-Face, with ablations on the main components. The stated central claim is that the proposed model surpasses state-of-the-art baselines on several age estimation benchmarks.

Significance. If fully substantiated, the paper would show that a compact 13M-parameter GNN can match or exceed much larger CNN and Transformer models on facial age estimation. The architecture is plausible, the component-level ablations show internally consistent improvements, and the code is publicly available. However, the central claim is not supported as written: on MORPH II Setting I, the paper's own Table 1 shows that several existing methods achieve lower MAE than LRA-GNN, both with and without pre-training. In addition, the comparisons across datasets mix pre-trained and non-pre-trained baselines without protocol matching, and no uncertainty estimates are provided. The significance of the contribution therefore remains contingent on a corrected and protocol-matched empirical evaluation.

major comments (4)
  1. [Abstract, Section 4.4.1, Section 5, Table 1] The abstract and conclusion claim LRA-GNN 'surpasses the state-of-the-art baselines' on 'several age estimation benchmarks,' but Table 1 directly contradicts this for MORPH II Setting I. With pre-training, LRA-GNN achieves MAE 1.79, while HR reports 1.13*, GLAE reports 1.14*, and TAA-GCN reports 1.69; without pre-training, LRA-GNN achieves 2.02, which is worse than MetaAge's 1.81 and the same three methods. The text acknowledges underperforming HR, GLAE, and TAA-GCN in Section 4.4.1 but retains the global 'surpasses SOTA' phrasing in the abstract and conclusion. This is an internal inconsistency in a load-bearing claim and should be corrected by restricting the claim to the settings and datasets where LRA-GNN actually ranks first.
  2. [Section 4.1 and Tables 1-4] The comparison protocol is not matched across methods. The tables mix models with and without external pre-training, different backbones, and different preprocessing choices, with only an asterisk to indicate pre-training. For MORPH II Setting I, the train/test split is described as a random 80/20 partition with no fixed split or seed, so the reported differences of 0.02-0.1 MAE may reflect split variance rather than model quality. No confidence intervals, standard deviations, or significance tests are reported for any of the benchmark results. Consequently, even the results where LRA-GNN ranks first (e.g., Setting II, FG-NET, ChaLearn, UTK-Face) are not established as reliable state-of-the-art comparisons.
  3. [Section 4.3, Figure 6, Table 8] Several hyperparameters are selected using the same benchmarks on which the final results are reported. Section 4.3 fixes L=12, M=8, eta=0.5, initial-graph threshold 0.936, random-walk threshold 0.824, and Section 3.5.6 sets tau=1.3; Figure 6 tunes M on FG-NET and Table 8 tunes eta on MORPH II and FG-NET. The paper does not describe a validation split or nested tuning procedure. This selection on the test benchmarks makes the reported MAE values optimistic and weakens the generalization claim. The authors should either report results on a held-out validation split or quantify the sensitivity of the final numbers to these choices.
  4. [Section 3.5 and Algorithm 2] The progressive RL component is not specified precisely enough to assess its contribution. The MDP is defined by a five-tuple, but the transition function is not given: Algorithm 2 updates the state only as 'obtain next state s_{t+1}', and the episode termination condition is only described as 'reach the maximum cumulative reward.' It is also unclear how the DDQN loss interacts with the focal and MAE loss in Eq. (12). Since Table 7 attributes a substantial gain to the RL design, a precise formulation of the MDP transitions and an ablation separating the RL machinery from the combined loss would be needed to support that claim.
minor comments (6)
  1. [Section 3.2.2, Eq. (1)] The text says 'm, n are the adjustable parameters' but the equation uses p and q; the notation should be made consistent.
  2. [Section 4.4] The text says 'comprehensive experiments across three facial image datasets,' but the paper evaluates on four datasets (MORPH II, FG-NET, ChaLearn LAP 2016, and UTK-Face).
  3. [Table 3 caption] The caption reads 'ChaLearn LAP 2016 dataset' but the table title is 'CLAP2016'; use the consistent dataset name.
  4. [Section 4.5] The phrase 'without loading the eternal dataset pre-trained weights' should read 'external dataset pre-trained weights.'
  5. [Section 3.3, Eq. (3)] The symbol M is used both for the feature dimension in Eq. (3) and for the number of attention heads in Section 4.3; this should be disambiguated.
  6. [Section 5] The conclusion says the method outperforms SOTA on 'Morph II, FG-NET, and CLAP 2016' but omits UTK-Face, even though Table 4 reports results on that dataset; the list should be consistent with the experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported MAEs are empirical test-set results of a learned architecture, not derivations that reduce to their inputs; the paper's central 'surpasses SOTA' claim is undermined by its own Table 1, but that is a correctness/consistency failure, not circularity.

full rationale

The paper does not claim a derivation from first principles to its numerical results; LRA-GNN is an architecture plus training procedure, and Tables 1-4 report empirical MAE on held-out test splits. I checked for self-definitional reductions. The 'latent relations' in Eq. (3) are defined as the softmax-attention-weighted fully connected graphs produced from node features; this defines the module, and the ablation in Table 5 measures its empirical contribution rather than algebraically forcing the reported gains. Eq. (2) does use a cosine-similarity threshold for random-walk guidance, and thresholds (0.936, 0.824), layer count L=12, heads M=8, and loss weight eta are selected with reference to benchmark performance; this is tuning-induced optimism or selection bias, not circularity by construction, because those hyperparameters do not determine the test labels or the MAE formula. Self-citations to GroupFace (Zhang et al. 2024a) and MCGRL (Shou et al. 2025a) appear as baselines and related work, not as load-bearing justifications; no uniqueness theorem or cited prior result is invoked to force the architecture or the empirical outcomes. The strongest issue is that the abstract/conclusion claim 'surpasses the state-of-the-art baselines' is contradicted by the paper's own Table 1, where HR (1.13), GLAE (1.14), and TAA-GCN (1.69) achieve lower MAE than LRA-GNN's 1.79 on MORPH II Setting I with pre-training; the paper even concedes 'Under Setting I, we only underperform HR, GLAE and TAA-GCN.' This is an internal-consistency and benchmarking-protocol problem for the correctness of the headline, not a circular derivation. No equation-level or citation-level circular step can be exhibited, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central contribution rests on domain assumptions about facial keypoints and graph relations, plus roughly seven hand-tuned hyperparameters. No new physical or conceptual entities are introduced; 'latent relations' is a learned graph connectivity pattern, not a separate ontological addition.

free parameters (7)
  • initial graph similarity threshold = 0.936
    Set during experimental tuning in Section 4.3; controls initial adjacency edges.
  • random walk updating threshold = 0.824
    Set during tuning in Section 4.3; controls which cosine-similarity connections are added.
  • number of GCN layers L = 12
    Chosen in Section 4.5.3 as a trade-off between performance and parameter count.
  • number of multi-head attention heads M = 8
    Chosen from ablation in Section 4.5.2 (Figure 6).
  • loss weight eta = 0.5 (MORPH II), 0.4 (FG-NET)
    Tuned in Section 4.5.4 (Table 8); values differ per dataset.
  • focal loss focusing parameter tau = 1.3
    Fixed in Section 3.5.6 for better classification; no sensitivity analysis is reported.
  • random walk p and q parameters = not specified
    Eq. (1) defines p and q but the paper never reports their values; BFS+DFS is claimed to combine both regimes.
assumptions (6)
  • domain assumption Facial keypoints and local patches carry most age-discriminative information.
    Used in Section 3.2.1 to justify segmenting the face into keypoint-centered patches as graph nodes; if false, the graph representation loses needed signal.
  • domain assumption The landmark detector from Korban et al. (2023) provides sufficiently accurate keypoints for patch construction.
    Invoked in Section 3.2.1; inaccurate landmarks would misalign nodes and edges.
  • domain assumption Random walk with cosine-similarity thresholding reveals global graph structure useful for age estimation.
    Section 3.2.2; no theoretical guarantee, relies on the empirical benefit of node2vec-style walks.
  • domain assumption Multi-head attention scores approximate latent semantic relations between facial regions.
    Section 3.3; the core claim that missing relations are recovered depends on this learned-attention assumption.
  • domain assumption Deep residual graph convolutions with adaptive initial and dynamic residuals prevent over-smoothing.
    Section 3.4; the paper cites DRGCN (Zhang et al. 2023) but gives no proof that the proposed residual form controls over-smoothing.
  • domain assumption Benchmark splits, preprocessing, and pretraining are comparable across methods listed in Tables 1 to 4.
    Section 4.1 to 4.4; the SOTA comparison is only valid if protocols match.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LRA-GNN: Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual for Facial Age Estimation." pith.science (2026). https://pith.science/paper/JF7CEM2Q

@misc{pith2026250205423,
  author       = {Pith},
  title        = {Pith review of: LRA-GNN: Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual for Facial Age Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JF7CEM2Q}},
  note         = {Machine review of arXiv:2502.05423}
}
read the original abstract

Face information is mainly concentrated among facial key points, and frontier research has begun to use graph neural networks to segment faces into patches as nodes to model complex face representations. However, these methods construct node-to-node relations based on similarity thresholds, so there is a problem that some latent relations are missing. These latent relations are crucial for deep semantic representation of face aging. In this novel, we propose a new Latent Relation-Aware Graph Neural Network with Initial and Dynamic Residual (LRA-GNN) to achieve robust and comprehensive facial representation. Specifically, we first construct an initial graph utilizing facial key points as prior knowledge, and then a random walk strategy is employed to the initial graph for obtaining the global structure, both of which together guide the subsequent effective exploration and comprehensive representation. Then LRA-GNN leverages the multi-attention mechanism to capture the latent relations and generates a set of fully connected graphs containing rich facial information and complete structure based on the aforementioned guidance. To avoid over-smoothing issues for deep feature extraction on the fully connected graphs, the deep residual graph convolutional networks are carefully designed, which fuse adaptive initial residuals and dynamic developmental residuals to ensure the consistency and diversity of information. Finally, to improve the estimation accuracy and generalization ability, progressive reinforcement learning is proposed to optimize the ensemble classification regressor. Our proposed framework surpasses the state-of-the-art baselines on several age estimation benchmarks, demonstrating its strength and effectiveness.

Figures

Figures reproduced from arXiv: 2502.05423 by the authors.

Figure 1
Figure 1. The illustration of the latent relation (e.g., the dashed part). Existing GNN-based methods utilize the similarity threshold method and may ignore relations between facial key points and wrinkles such as 𝑟1 , which are not similar enough but important to attain more accurate estimation. Graph Representation Learning (MCGRL) to capture the rich structural information of the face, which outperforms the CNN and Transfo… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The illustration of Random Walk Updating. BFS tends to visit the immediate neighbors of the source node, and DFS tends to explore nodes that are further and further away from the source node. The combination of the two can effectively capture the local and global relations between nodes. Leskovec, 2016; Zhou et al., 2024) is employed in the initial graph to collect more information with fewer paths in the search spa… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The illustration of deep feature extraction with the adap￾tive initial residuals and dynamic developmental residuals. The adaptive initial residuals obtain the personalized characteristics from initial embedding 𝐻(0) and hidden embedding 𝐻(𝑙) . The dynamic developmenta…
Figure 5
Figure 5. Figure 5: The illustration of Progressive RL-based Age Estima￾tion. The age estimation through classification and then regression is defined as a walk-to-the-end problem on a grid. through classification and then regression on the samples as correctly as possible. We model the p…
Figure 6
Figure 6. Figure 6: The ablation study of Latent Relation Capturing on the FG-NET dataset. two strategies on latent relation capturing and to discuss the impact of the quantity of 𝑀 heads in the mechanism of multi-head attention, we conduct experiments on the FG￾NET dataset, which is eval…
Figure 7
Figure 7. Figure 7: The ablation study of Deep Feature Extraction on the FG-NET dataset. the upper limit of the performance without DDR is difficult to break through and improves slowly. Our Deep Feature Extraction fusing AIR and DDR can better alleviate the GCN over-smoothing problem and…
Figure 8
Figure 8. Figure 8: The examples of age estimation results of our LRA-GNN on three facial datasets. The factual label is the black number, the reliable estimation results are shown in the green number, and the poor estimation results are shown in the red number [PITH_FULL_IMAGE:figures/f…
Figure 9
Figure 9. Figure 9: The accuracy analysis of the baseline and our LRA-GNN on Morph II and FG-NET datasets. 4.6.1. Accuracy Analysis We discuss the estimation accuracy of LRA-GNN on Morph II (Setting I) and FG-NET datasets for different age groups. Baseline utilizes a common DeepGCN withou…
Figure 10
Figure 10. Figure 10: The visualization examples of LRA-GNN for facial feature extraction utilizing attention heat map. capturing latent relations. As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

  1. [8]

    arXiv preprint arXiv:1609.02907

    Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 . Korban, M., Youngs, P., Acton, S.T.,

  2. [12]

    18760– 18769

    Moving window regression: A novel approach to ordinal regression, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 18760– 18769. Shou, Y., Ai, W., Du, J., Meng, T., Liu, H., 2024a. Efficient long-distance latent relation-aware graph neural network for multi-modal emotion recognition in conversations. arXiv preprin...

  3. [13]

    Tan,Z.,Wan,J.,Lei,Z.,Zhi,R.,Guo,G.,Li,S.Z.,2017

    Deep conditional distribution learningforageestimation.IEEETransactionsonInformationForensics and Security 16, 4679–4690. Tan,Z.,Wan,J.,Lei,Z.,Zhi,R.,Guo,G.,Li,S.Z.,2017. Efficientgroup-n encoding and decoding for facial age estimation. IEEE transactions on pattern analysis and machine intelligence 40, 2610–2623. Van Hasselt, H., Guez, A., Silver, D.,

  4. [14]

    arXiv preprint arXiv:1710.10903

    Graph attention networks. arXiv preprint arXiv:1710.10903 . Wang, C., Li, Z., Mo, X., Tang, X., Liu, H.,

  5. [15]

    IEEE Transac- tions on Information Forensics and Security

    Exploiting unfairness with meta-set learning for chronological age estimation. IEEE Transac- tions on Information Forensics and Security . Wen,G.,Wu,K.,2021. Buildingdecisiontreeforimbalancedclassification via deep reinforcement learning, in: Asian Conference on Machine Learning, PMLR. pp. 1645–1659. Yang, J., El-Bouri, R., O’Donoghue, O., Lachapelle, A.S...

  6. [16]

    Prediction model of dow jones index based on lstm-adaboost, in: 2021 International Conference on Com- munications, Information System and Computer Engineering (CISCE), IEEE. pp. 808–812. Zhang, C., Liu, S., Xu, X., Zhu, C., 2019a. C3ae: Exploring the limits of compact model for age estimation, in: Proceedings of the IEEE/CVF conference on computer vision ...

  7. [17]

    IEEE signal processing letters 23, 1499–1503

    Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23, 1499–1503. Zhang,L.,Yan,X.,He,J.,Li,R.,Chu,W.,2023. Drgcn:Dynamicevolving initial residual for deep graph convolutional networks, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 11254–11261. Zhang, Y., Shou, Y., Ai,...

  8. [2014]

    arXiv preprint arXiv:1412.6980

    Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 . Kipf, T.N., Welling, M.,

Show all 17 references
  1. [2015]

    arXiv preprint arXiv:1506.05163

    Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 . Hiba, S., Keller, Y.,

  2. [2016]

    Chalearn looking at people and faces of the world: Face analysis workshop and challenge 2016, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 1–8. Gao, B.B., Zhou, H.Y., Wu, J., Geng, X.,

  3. [2017]

    2980–2988

    Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, pp. 2980–2988. Meng,T.,Shou,Y.,Ai,W.,Du,J.,Liu,H.,Li,K.,2024a. Amulti-message passing framework based on heterogeneous graphs in conversational emotion recognition. ...

  4. [2018]

    2304–2313

    Deep regression forests for age estimation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2304–2313. Shen,W.,Zhao,K.,Guo,Y.,Yuille,A.L.,2017. Labeldistributionlearning forests. Advances in neural information processing systems

  5. [2020]

    arXiv preprint arXiv:2010.11929

    An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 . Duan, M., Li, K., Li, K.,

  6. [2021]

    Artificial Intelligence Review 54, 179–213

    Deep learning approach for facial age classification: a survey of the state-of-the-art. Artificial Intelligence Review 54, 179–213. Ai, W., Deng, W., Chen, H., Du, J., Meng, T., Shou, Y., 2024a. Mcsff: Multi-modal consistency and specificity fusion framework for entity alignme...

  7. [2022]

    IEEE Transactions on Image Processing 31, 4761–4775

    Metaage: meta- learning personalized age estimators. IEEE Transactions on Image Processing 31, 4761–4775. Lin,E.,Chen,Q.,Qi,X.,2020.Deepreinforcementlearningforimbalanced classification. Applied Intelligence 50, 2488–2502. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.,

  8. [2023]

    long- tailedageestimation:Anapproachtokilltwobirdswithonestone

    General vs. long- tailedageestimation:Anapproachtokilltwobirdswithonestone. IEEE Transactions on Image Processing 32, 6155–6167. Bao,Z.,Tan,Z.,Wan,J.,Ma,X.,Guo,G.,Lei,Z.,2022. Divergence-driven consistency training for semi-supervised facial age estimation. IEEE Transactions o...

  9. [2024]

    Expert Systems with Applications 248, 123403

    Rakcr: Reviews sentiment-aware based knowledge graph convolutional networks for personalized recommendation. Expert Systems with Applications 248, 123403. Deng,Z.,Liu,H.,Wang,Y.,Wang,C.,Yu,Z.,Sun,X.,2021. Pml:Progres- sivemarginlossforlong-tailedageclassification,in:Proceeding...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.