acceptodds
Under review as a conference paper at ICLR 2027

Beyond Homophily: Dual-Frequency Fusion for Multi-Label Node Classification

Abstract

Multi-label node classification (MLNC) predicts multiple labels for each node in graph data, where connected nodes exhibit diverse label semantics, resulting in the coexistence of homogeneous and heterogeneous neighborhood information. Considering that homogeneous information preserves semantic consistency while heterogeneous one may harm consistency, existing methods mainly suppress heterogeneous neighbors while overlooking their potential discriminability. Integrating the merits of both information empowers representations but remains two coupled issues: (1) **Feature Entanglement.** Aggregating heterogeneous neighbors mixes different label semantics, resulting in ambiguous feature distributions and obscuring decision boundaries; (2) **Optimization Inconsistency.** These entangled features induce inconsistent prediction gradients among neighbors, making it difficult to reliably assign the contribution of neighborhood information. To mitigate these issues, we propose a Dual-Frequency Fusion Network (DFNet), which realizes effective fusion of homogeneous and heterogeneous neighborhood information from a frequency perspective. Specifically, DFNet models homogeneous and heterogeneous information as low- and high-frequency signals, respectively, and develops a dual-frequency fusion strategy to exploit their complementarity. Moreover, DFNet introduces a Joint Representation-Optimization Consistency (JROC) module containing Label-guided Feature Refinement (LFR) and Gradient Consistency Calibration (GCC) to align representation and optimization spaces in a mutually reinforcing manner. LFR aligns node representations with relevant labels to disentangle ambiguous features and clear decision boundaries, yielding reliable prediction gradients. Based on the similarities of these gradients, GCC realizes to quantify optimization consistency and calibrate the fusion process. Through this synergy, DFNet successfully exploits complementary homogeneous and heterogeneous information. Extensive experiments show that DFNet achieves superior performance over state-of-the-art MLNC methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
graph convolutional network, graph learning, multi-label learning, semi-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Ly3YOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on two coupled issues, namely Feature Entanglement and Optimization Inconsistency in MLNC, and proposes DF2NetDF^2Net, which tackles the two issues by modeling homogeneous and heterogeneous information as low- and high-frequency signals, respectively. A JROC module, including LFR and GCC, is introduced to align representation and optimization spaces, where LFR is introduced to disentangle features and clear decision boundaries and GCC quantifies optimization consistency and calibrate the fusion process. Extensive experiments on five datasets show DF2NetDF^2Net achieves superior performance over state-of-the-art MLNC baselines.

Critical strengths
  1. Unlike existing MLNC methods suppressing heterogeneous signals to avoid semantic confusion directly, this paper claims the benefits from the heterogeneous signals and appropriately uses them to mine complementary information with a technically sound framework. And the idea is interesting.

  2. The model effectiveness is evaluated on five datasets with existing MLNC methods, and the results show the advantages against the baseline models. Additionally, the authors also provide the statistics analysis to verify the superiority.

Critical weaknesses

None.

Questions for authors
  1. Though the paper has demonstrated the promising model performance, the model efficiency has not been evaluated. Could the authors provide a comparison of the training and inference time (e.g., seconds per epoch) and GPU memory usage between DF2NetDF^2Net and the baselines.

  2. The authors should provide the experimental results that the model uses either homogeneous information or heterogeneous information to further demonstrate the rationality of combining both types of information.

  3. The fusion weight α\alpha is generated from both the frequency signals and the label representation. Whether the performance improvement is attributable to the learnable label representations but not the frequency modeling? The authors could provide experiments to explain this question.

  4. The authors don’t provide the setting of data split, which makes it hard to reproduce the experiments.

3 comments
AuthorsOpenReview ↗

Responses to Questions 1 & 2

Response to Q1:

We evaluate the training time, inference time, and GPU memory usage for one epoch, and the experimental results are provided below. The results demonstrate that DF$^2$Net achieves promising computational efficiency. We have included these results in the refined version.

DatasetsMetricsMLGWML-GCNLANCSMLGLIPCorGCNDF$^2$Net
HumlocTraining (s)0.20250.52880.64890.86170.65450.63100.083
Inference (s)0.22760.30230.38700.24340.01500.50740.0078
GPU (MB)289.3058.1347.7065.8264.32167.8235.85
PCGTraining (s)0.26110.52080.64581.22520.49391.00330.022
Inference (s)0.28000.33070.35120.34800.01730.31260.0029
GPU (MB)289.7374.7949.6099.32187.55188.9242.90
DelveTraining (s)1.90374.08451.968416.340821.3543412.28432.436
Inference (s)3.99234.62924.02824.01730.5183104.79090.1381
GPU (MB)66.77466.50101.87199.932093.445840.7862.22
ProteinsTraining (s)12.82585.68092.998713.63991015.262571.53332.488
Inference (s)7.00705.52875.11176.80650.590329.40930.0286
GPU (MB)112.39823.21278.061372.09401.212327.35311.93

Response to Q2:

To demonstrate the necessity of integrating both types of information, we conducted experiments that isolate the high-frequency (α=0\alpha=0) and low-frequency (α=1\alpha=1) components. The experimental results show that the combination approach ("High-frequency+Low-frequency", namely the adoption of the learnable α\alpha) achieves the best performance. This confirms that combining homogeneous and heterogeneous information captures complementary information.

DatasetsFrequency CombinationHLRLMa-APMi-APMa-AUCMi-AUCLRAP
HumLocHigh-frequency8.23±0.1115.49±0.3219.71±0.6939.92±1.1365.37±1.1885.13±0.2459.30±0.67
Low-frequency7.42±0.1412.90±0.4528.17±0.8653.04±1.3274.20±1.2088.14±0.5166.93±0.74
High-frequency+Low-frequency7.29±0.0712.54±0.1928.39±0.7754.69±0.5574.24±0.7288.76±0.1068.46±0.45
PCGHigh-frequency13.13±0.0227.58±0.2816.02±0.6026.25±0.8254.12±1.0669.82±0.4746.94±0.57
Low-frequency12.98±0.1126.19±0.4725.17±0.4934.35±0.4365.18±0.4073.59±0.2251.20±0.65
High-frequency+Low-frequency12.68±0.0525.37±0.3025.31±0.3434.62±0.0865.29±0.2474.32±0.1951.38±0.38
AuthorsOpenReview ↗

Responses to Questions 3 & 4

Response to Q3:

To clarify the source of the performance improvements, we conduct experiments that decouple the frequency modeling and the label representations. We compare the full model ("W/ F&L") against variants lacking frequency modeling ("W/o F") and lacking label representations ("W/o L"), where "F" means the frequency signals and "L" means the label representations.

For constructing the variant without frequency modeling ("W/o F" variant), we employ a GCN to extract node representations and treat the output as the structural state representation HstateH_{state}, which then interacts with the label representation matrix EE. For DF$^2$Net without label embedding ("W/o L" variant), we remove the label representation matrix EE, and map the structural state representation HstateH_{state} to the class dimension.

The experimental results below indicate that removing either component degrades performance. The best performance is achieved when both components are included, confirming that the performance gain relies on both frequency modeling and label representations.

DatasetsComponentsHLRLMa-APMi-APMa-AUCMi-AUCLRAP
HumLocW/o F7.33±0.2114.20±0.6227.93±1.1652.78±1.6371.45±1.4986.85±0.6966.56±1.32
W/o L7.29±0.0613.06±0.5728.14±1.2154.13±0.6173.99±1.2588.36±0.3867.29±0.89
W/ F&L7.29±0.0712.54±0.1928.39±0.7754.69±0.5574.24±0.7288.76±0.1068.46±0.45
PCGW/o F12.79±0.0525.94±0.3022.80±0.5033.33±0.5662.85±0.6973.16±0.1650.86±0.25
W/o L12.76±0.0325.54±0.2723.39±0.4433.82±0.3363.60±0.2573.56±0.1751.30±0.47
W/ F&L12.68±0.0525.37±0.3025.31±0.3434.62±0.0865.29±0.2474.32±0.1951.38±0.38

Response to Q4:

In our experiments, all datasets are randomly partitioned into training, validation, and testing sets with a 6:2:2 data split. This data split has been implemented in previous works [1] and [2]. We have added the data split in the refined version.

[1]Correlation-aware graph convolutional networks for multi-label node classification. In SIGKDD 2025.

[2] Multilabel node classification with label influence propagation. In ICLR 2025.

Dear Reviewer Ly3Y,

Thank you very much for your thoughtful feedback and for taking the time to consider our responses. We sincerely appreciate your recognition of our work and your encouraging evaluation.

Thank you again for your time and consideration. We wish you all the best!

Authors

Reviewer_DUzDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multi-label node classification. While most existing studies focus on the negative effects of heterophilous neighbors, this paper argues that they are not necessarily noise. Rather than simply suppressing heterophilous information, it is important to identify which information is useful and how much it should contribute. Based on this idea, the paper proposes a dual-frequency fusion network that treats homophilous and heterophilous information as low- and high-frequency signals, respectively, and combines them to capture complementary information.

Critical strengths

The paper makes an interesting point: heterophilous neighbors are not always harmful and can provide useful discriminative information. The authors separate neighborhood information into low- and high-frequency signals and use adaptive fusion to combine them.

Critical weaknesses

The critical weaknesses are the limited baseline coverage and lack of dataset diversity.

Questions for authors

First, the baseline comparison could be more comprehensive. Including classic GNNs and other competitive methods for multi-label node classification would help determine whether the improvements come from the proposed modules. Second, dataset diversity is limited: four of the five datasets are from biological networks, while only one is a Wikipedia document network. This makes it difficult to assess how well the method generalizes to different graph structures and application domains.

5 comments
AuthorsOpenReview ↗

Comparative analysis with classic or competitive GNNs.

We incorporate GraphSAGE [1] and GAT [2] as representative classic GNN baselines, alongside DualFormer [3] as an additional competitive method. The experimental results in Table 1 demonstrate the superiority of our proposed method. These distinct performance gaps confirm that the improvements stem from the proposed modules.

Table 1. Comparative results with classic or competitive GNNs.

DatasetsMetricsGraphSAGEGATDualFormerOurs
HumlocRL14.50±0.5615.92±0.2113.21±0.7112.54±0.19
HL8.17±0.148.35±0.148.00±0.137.29±0.07
Ma-AUC65.10±1.8670.99±3.0974.35±1.5474.24±0.72
Mi-AUC84.76±0.7685.10±0.4187.43±0.4888.76±0.10
Ma-AP18.92±1.0420.38±1.5124.26±1.4428.39±0.77
Mi-AP38.40±1.8436.30±1.4743.83±1.4754.69±0.55
LRAP61.92±1.3457.47±0.8664.43±1.1368.46±0.45
PCGRL28.05±0.6226.10±0.2026.13±0.5225.37±0.30
HL12.86±0.5012.64±0.4013.05±0.4812.68±0.05
Ma-AUC54.68±0.4560.57±0.6262.04±0.9365.29±0.24
Mi-AUC69.26±0.4472.33±0.4172.67±0.8374.32±0.19
Ma-AP15.47±0.6320.13±1.0821.91±1.5925.31±0.34
Mi-AP24.92±0.4229.29±0.7929.95±1.5834.62±0.08
LRAP46.13±0.6948.83±0.6348.73±0.8551.38±0.38
PPIRL18.35±0.2120.39±0.1921.26±0.1514.11±0.23
HL22.51±0.2724.55±0.3026.32±0.4119.33±0.37
Ma-AUC73.19±0.3867.46±0.4267.78±0.4680.73±0.36
Mi-AUC80.44±0.2576.68±0.3375.78±0.3585.49±0.24
Ma-AP53.72±0.4945.95±0.8146.81±0.8465.55±0.63
Mi-AP67.69±0.4462.23±0.6060.37±0.7775.72±0.50
LRAP68.92±0.3766.40±0.3665.45±0.3576.62±0.46
ProteinsRL10.85±0.1312.60±0.0513.77±0.977.40±0.15
HL10.70±0.0511.21±0.0613.92±1.658.71±0.04
Ma-AUC78.61±0.3070.39±0.1572.51±0.7387.93±0.12
Mi-AUC81.95±0.2376.09±0.1073.89±2.0389.71±0.11
Ma-AP31.21±0.3021.60±0.2222.65±0.4951.42±0.43
Mi-AP41.69±0.3333.13±0.1231.11±1.8260.35±0.36
LRAP76.12±0.2274.31±0.1072.98±0.9381.76±0.20
DelveRL2.06±0.052.14±0.043.13±0.041.80±0.07
HL2.43±0.012.45±0.013.03±0.011.77±0.01
Ma-AUC96.30±0.1396.17±0.1695.49±0.1097.41±0.04
Mi-AUC98.03±0.0397.94±0.0496.87±0.0498.53±0.06
Ma-AP71.98±0.4771.54±0.4964.75±0.5777.12±0.68
Mi-AP87.47±0.1387.21±0.1480.40±0.1690.65±0.11
LRAP91.25±0.1091.05±0.0886.99±0.0693.10±0.07

[1] Inductive representation learning on large graphs. In NeurIPS 2017.

[2] Graph attention networks. In ICLR 2018.

[3] Dualformer: Dual graph transformer. In ICLR 2025.

AuthorsOpenReview ↗

Datasets diversity analysis.

To assess the generalization of our model, we conduct experiments on two additional networks: DBLP (a citation network) and BlogCatalog (a social network).

The experimental results in Table 2 show that our method also performs well on these datasets, proving that our method can generalize to different graph structures and application domains.

We have included these GNN baselines and datasets in the refined version.

Table 2. Comparative results on DBLP and BlogCatalog datasets.

DatasetsMetricsGraphSAGEGATMLGWML-GCNLANCSMLGDualFormerLIPCorGCNOurs
DBLPRL8.29±0.258.21±0.348.99±0.147.43±0.147.88±0.1218.91±0.766.38±0.166.10±0.166.84±0.285.61±0.15
HL11.83±0.2911.68±0.3211.71±0.0611.25±0.1011.18±0.0920.41±0.6010.17±0.1210.35±0.2810.04±0.298.29±0.21
Ma-AUC92.51±0.1792.73±0.3293.15±0.0392.56±0.0692.87±0.2179.13±0.8394.65±0.1295.30±0.1894.59±0.1895.79±0.25
Mi-AUC93.21±0.1493.41±0.2893.63±0.0493.29±0.0593.51±0.1582.71±0.7995.16±0.1395.54±0.1895.05±0.1896.13±0.22
Ma-AP84.54±0.4785.00±0.6085.36±0.0485.49±0.2185.78±0.5558.70±0.9487.82±0.2889.29±0.4288.07±0.2990.70±0.35
Mi-AP87.74±0.3088.01±0.4787.86±0.1188.51±0.1488.83±0.2067.33±0.9590.58±0.2890.92±0.4690.67±0.2992.54±0.35
LRAP91.04±0.2991.17±0.3990.24±0.0791.79±0.1791.61±0.1480.69±0.8192.98±0.2293.14±0.2092.66±0.3494.12±0.11
BlogCatalogRL25.94±0.3131.58±20.0225.45±0.4825.68±0.2525.48±0.1249.42±3.5625.74±0.2539.56±4.2525.42±0.2325.44±0.47
HL3.59±0.0113.72±14.823.60±0.013.58±0.033.55±0.033.80±0.433.60±0.0236.51±7.703.58±0.033.52±0.03
Ma-AUC49.08±0.6939.45±19.7950.34±0.4850.94±1.1753.29±0.5249.44±1.2450.08±0.9253.71±1.5654.48±0.5255.66±0.99
Mi-AUC73.58±0.1748.59±26.9558.07±11.4173.85±0.1774.11±0.0650.80±3.1268.10±0.4159.88±3.4174.15±0.2174.23±0.62
Ma-AP3.96±0.103.17±1.593.95±0.084.16±0.145.07±0.204.09±0.104.09±0.155.06±0.324.61±0.155.43±0.65
Mi-AP9.03±0.135.01±3.509.57±0.519.42±0.1511.71±0.555.47±0.657.19±0.214.58±0.599.65±0.2111.18±1.66
LRAP27.20±0.3515.43±10.4128.04±1.0227.87±0.1828.94±1.0020.35±1.8927.64±0.2213.63±2.6428.32±0.2929.00±1.28

Dear Reviewer DUzD,

With the discussion period nearing its conclusion, we would like to express our sincere appreciation for your suggestions and recognition of this work. We hope our experiments and clarifications have addressed the concerns you raised. If there are any remaining questions, please do not hesitate to let us know. We would be glad to provide further explanations.

Thank you again for your time and consideration.

Best regards,

Authors

Reviewer_m2nROpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper explores how to effectively integrate homogeneous and heterogeneous neighborhood information for multi-label node classification to enhance representation learning. DFNet employs a dual-frequency (high and low) fusion strategy, complemented by a joint representation-optimization consistency mechanism, to maximize the complementarity between these two types of information. Experimental results demonstrate that this approach offers distinct performance advantages.

Critical strengths
  1. Both homogeneous and heterogeneous neighborhood information are significant; the latter, in particular, often contains crucial discriminative features. Fusing both allows for a more comprehensive characterization of nodes.
  2. The Related Work section is well-structured and clearly articulates the paper’s position within the broader field, specifically regarding its relationship to multi-label learning.
Critical weaknesses
  1. The overall structure of the paper requires further refinement. There are numerous modules, yet their interrelationships are not clearly articulated. For instance, Figure 1 appears overly complex and cluttered, which significantly hinders the paper's readability.
  2. The extraction of high-frequency information relies on the discrepancy between two encoders; however, is this relationship strictly one-to-one? The authors should exercise greater rigor in their use of these concepts.
Questions for authors
  1. Have the authors considered the robustness of the proposed method? In multi-label datasets, obtaining complete and accurate labels is often challenging. If label information is missing or contains noise, it may lead to discrepancies between nodes and their neighbors. How does the model mitigate the impact of such erroneous signals?
  2. The model relies heavily on neighborhood information. If the number of informative neighbors is insufficient, or if the information density is too low, can the model still learn effectively?
  3. Given the complexity of the method’s design, which involves extensive mining and utilization of various relationships, how efficient is the approach in practice?
10 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1.

Response to W1:

Regarding the structural clarity, we have outlined the model workflow in the Introduction section and provided a brief overview at the beginning of the Methodology section in the initial manuscript. To further articulate the interrelationships among the modules and facilitate understanding, we restate the streamlined and unified flow of DF$^2$Net concisely:

(1) Dual-Frequency Representation Extraction extracts the necessary frequency representations;

(2) Label-Aware Frequency Fusion dynamically combines the information from these frequency representations;

(3) Joint Representation-Optimization Consistency includes an LFR module to disentangle node features and improve the representation quality for frequency fusion, and a GCC module to calibrate the fusion weight to enhance the fusion process.

To reduce the reading burden of Figure 2 (we assume that the Figure 1 you referred to may correspond to Figure 2 in our manuscript), we have modified the diagram in the refined manuscript as follows:

  1. We only retain the forward information flow from the Dual-Frequency Representation Extraction to the Label-Aware Frequency Fusion.

  2. For the Joint Representation-Optimization Consistency module, we have removed the intersecting lines that represent data dependencies from other modules and adjusted the colors of the corresponding symbols to match their originating modules. This color-coding strategy allows readers to easily and intuitively trace the source of the data used in other modules.

We believe these refinements help enhance the readability of the overall architecture.

AuthorsOpenReview ↗

Response to Critical Weakness 2 (Part 1/2).

Response to W2:

We clarify that the discrepancy captures the high-frequency information as follows:

1. Design Motivation:

In our work, we introduce graph signal processing theory to demonstrate that the discrepancy between an anchor node and its aggregated neighborhood can capture high-frequency signals. The theory serves as the guidance that motivates our model design: construct the discrepancy between the neighborhood representation and the intrinsic representation to extract high-frequency information.

2. Discrepancy Captures High-Frequency Signals:

To explain why our method extracts high-frequency information with two encoders, we decompose the predictions Zself\mathbf{Z}_{\mathrm{self}} and Zlow\mathbf{Z}_{\mathrm{low}} after applying the shared classification head W\mathbf{W}.

Since the input X=A^X+(I−A^)X\mathbf{X} = \hat{\mathbf{A}}\mathbf{X} + (\mathbf{I}-\hat{\mathbf{A}})\mathbf{X} consists of a smooth low-frequency component A^X\hat{\mathbf{A}}\mathbf{X} and a high-frequency component (I−A^)X(\mathbf{I}-\hat{\mathbf{A}})\mathbf{X}, Zself\mathbf{Z}_{\mathrm{self}} can be decomposed as:

Zself=A^XWMLPW+(I−A^)XWMLPW,\mathbf{Z}_{\mathrm{self}} = \hat{\mathbf{A}}\mathbf{X}\mathbf{W}_{\mathrm{MLP}}\mathbf{W} + (\mathbf{I}-\hat{\mathbf{A}})\mathbf{X}\mathbf{W}_{\mathrm{MLP}}\mathbf{W},

where WMLP\mathbf{W}_{\mathrm{MLP}} and WGCN\mathbf{W}_{\mathrm{GCN}} denote the weights of the two encoders. Then, the discrepancy Hhigh\mathbf{H}_{\mathrm{high}} can be decomposed as:

Hhigh=XWMLP−A^XWGCN=(I−A^)XWMLP+A^X(WMLP−WGCN),\mathbf{H}_{\mathrm{high}} = \mathbf{X}\mathbf{W}_{\mathrm{MLP}} - \hat{\mathbf{A}}\mathbf{X}\mathbf{W}_{\mathrm{GCN}} = (\mathbf{I}-\hat{\mathbf{A}})\mathbf{X}\mathbf{W}_{\mathrm{MLP}} + \hat{\mathbf{A}}\mathbf{X}(\mathbf{W}_{\mathrm{MLP}}-\mathbf{W}_{\mathrm{GCN}}),

where the first term is the high-frequency component and the second term is a redundant term. To suppress this redundant term, we apply a shared classification head W\mathbf{W} to project both streams into a unified label space:

Zhigh=Zself−Zlow=(I−A^)XWMLPW+A^X(WMLP−WGCN)W.\mathbf{Z}_{\mathrm{high}} = \mathbf{Z}_{\mathrm{self}} - \mathbf{Z}_{\mathrm{low}} = (\mathbf{I}-\hat{\mathbf{A}})\mathbf{X}\mathbf{W}_{\mathrm{MLP}}\mathbf{W} + \hat{\mathbf{A}}\mathbf{X}(\mathbf{W}_{\mathrm{MLP}}-\mathbf{W}_{\mathrm{GCN}})\mathbf{W}.

Here, the first term is the high-frequency component, while the second term is a low-frequency redundant term. Since both Zself\mathbf{Z}_{\mathrm{self}} and Zlow\mathbf{Z}_{\mathrm{low}} use the shared W\mathbf{W} to fit the same ground-truth labels, both streams are trained to produce consistent semantic predictions for the common low-frequency component A^X\hat{\mathbf{A}}\mathbf{X}. As a result, the outputs A^XWMLPW\hat{\mathbf{A}}\mathbf{X}\mathbf{W}_{\mathrm{MLP}}\mathbf{W} and A^XWGCNW\hat{\mathbf{A}}\mathbf{X}\mathbf{W}_{\mathrm{GCN}}\mathbf{W} align closely, reducing the redundant term and leaving the high-frequency component to dominate Zhigh\mathbf{Z}_{\mathrm{high}}.

AuthorsOpenReview ↗

Response to Critical Weakness 2 (Part 2/2).

3. Empirical Verification via Rayleigh Quotient:

To quantitatively verify that our method captures genuine high-frequency signals, we introduce the Rayleigh Quotient:

R(H)=Tr(H⊤L~H)Tr(H⊤H)=∑kλk∥h^k∥22∑k∥h^k∥22,R(\mathbf{H}) = \frac{\mathrm{Tr}(\mathbf{H}^\top\tilde{\mathbf{L}}\mathbf{H})}{\mathrm{Tr}(\mathbf{H}^\top\mathbf{H})} = \frac{\sum_{k}\lambda_{k}\Vert\hat{\mathbf{h}}_{k}\Vert_{2}^{2}}{\sum_{k}\Vert\hat{\mathbf{h}}_{k}\Vert_{2}^{2}},

where λk\lambda_{k} denotes the kk-th eigenvalue of the normalized graph Laplacian L~\tilde{\mathbf{L}}, and h^k\hat{\mathbf{h}}_{k} is the corresponding spectral coefficient vector. R(⋅)R(\cdot) calculates the Rayleigh Quotient that measures the energy-weighted average frequency of a representation. A smaller Rayleigh Quotient indicates that signal energy is concentrated in smooth, namely a low-frequency component, whereas a larger Rayleigh Quotient indicates a high-frequency component.

In view of this, we train our method on the HumLoc dataset 5 times, and record the mean Rayleigh Quotient with respect to Hlow\mathbf{H}_{\mathrm{low}}, Hhigh\mathbf{H}_{\mathrm{high}}, Zlow\mathbf{Z}_{\mathrm{low}}, and Zhigh\mathbf{Z}_{\mathrm{high}}. The results are shown in the Table below, where Low-Frequency records the Rayleigh Quotient of low-frequency signals, Discrepancy records the Rayleigh Quotient of discrepancy signals, and High-to-Low Ratio records the Rayleigh Quotient ratio between discrepancy signals and low-frequency signals.

According to the results, R(Hhigh)=0.502R(\mathbf{H}_{\mathrm{high}})=0.502 is higher than R(Hlow)=0.467R(\mathbf{H}_{\mathrm{low}})=0.467 with a High-to-Low Ratio of 1.0751.075, verifying that the discrepancy captures high-frequency information. More importantly, after projecting through the shared W\mathbf{W}, the Rayleigh Quotient of the low-frequency stream R(Zlow)R(\mathbf{Z}_{\mathrm{low}}) is reduced, while the discrepancy R(Zhigh)R(\mathbf{Z}_{\mathrm{high}}) is further boosted, elevating the High-to-Low Ratio to 1.3881.388. This empirical observation that the shared W\mathbf{W} simultaneously lowers the low-frequency Rayleigh Quotient and amplifies the high-frequency Rayleigh Quotient is consistent with the aforementioned analysis, confirming that sharing W\mathbf{W} reduces redundancy and facilitates the discrepancy to extract high-frequency information.

Representation SpacesLow-FrequencyDiscrepancyHigh-to-Low Ratio
H\mathbf{H}R(Hlow)=0.467R(\mathbf{H}_{\mathrm{low}})=0.467R(Hhigh)=0.502R(\mathbf{H}_{\mathrm{high}})=0.502R(Hhigh)/R(Hlow)=1.075R(\mathbf{H}_{\mathrm{high}})/R(\mathbf{H}_{\mathrm{low}})=1.075
Z=HW\mathbf{Z}=\mathbf{H}\mathbf{W}R(Zlow)=0.402R(\mathbf{Z}_{\mathrm{low}})=0.402R(Zhigh)=0.558R(\mathbf{Z}_{\mathrm{high}})=0.558R(Zhigh)/R(Zlow)=1.388R(\mathbf{Z}_{\mathrm{high}})/R(\mathbf{Z}_{\mathrm{low}})=1.388

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.