acceptodds
Withdrawn from ICLR 2027

ASPIRE: Make Spectral Graph Collaborative Filtering Great Again via Adaptive Filter Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph filter design is central to spectral collaborative filtering, yet most existing methods rely on manually tuned hyperparameters rather than fully learnable filters. We show that this challenge stems from a bias in traditional recommendation objectives, which induces a spectral phenomenon termed low-frequency explosion, thereby fundamentally hindering the effective learning of graph filters. To overcome this limitation, we propose a novel adaptive spectral graph collaborative filtering framework (ASPIRE) based on a bi-level optimization objective. Guided by our theoretical analysis, we disentangle the filter learning objective, which in turn leads to excellent recommendation performance, spectral adaptivity, and training stability in practice. Extensive experiments show our learned filters match the performance of carefully engineered task-specific designs. Furthermore, ASPIRE is equally effective in LLM-powered collaborative filtering. Our findings demonstrate that graph filter learning is viable and generalizable, paving the way for more expressive graph neural networks in collaborative filtering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spectral Graph Collaborative Filtering, Recommender Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KjgiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the instability of learnable spectral graph filters in collaborative filtering and attributes it to the low-frequency bias of conventional recommendation objectives. To address this issue, ASPIRE combines pre-filter normalization with a bi-level optimization framework to learn adaptive polynomial graph filters. Experiments across different graph settings and LLM-assisted recommendation tasks demonstrate improved training stability and competitive recommendation performance.

Critical strengths
  1. The theoretical analysis of low-frequency explosion provides useful insights into the instability of learnable graph filters under conventional recommendation objectives.
  2. The experiments cover diverse graph structures, filter initializations, and LLM-assisted recommendation settings, providing relatively comprehensive empirical evidence.
Critical weaknesses
  1. ASPIRE uses an additional 5% of observed interactions from the validation set to update filter parameters, whereas conventional baselines generally reserve validation data for model selection. I am concerned that this difference in supervision budgets may affect the fairness of the comparison, and the paper should further clarify potential information leakage and its impact on the reported performance gains.
  2. Although the paper compares ASPIRE with Naive-Learnable and handcrafted filters, the contributions of the learnable polynomial filter and the validation-guided bi-level optimization remain insufficiently disentangled. Further experiments incorporating alternative spectral filtering mechanisms, such as those used in JGCF [1] and SimGCF [2], would help assess whether the proposed optimization strategy generalizes across different filter designs.

[1] Guo et al. On manipulating signals of user-item graph: A jacobi polynomial-based graph collaborative filtering. SIGKDD'23.

[2] Liu et al. How do graph signals affect recommendation: Unveiling the mystery of low and high-frequency graph signals? SIGKDD'26.

Questions for authors
  1. ASPIRE uses an additional 5% of validation interactions to update filter parameters. Could the authors clarify whether this introduces information leakage or affects the fairness of comparisons with baselines?

  2. Would the proposed bi-level optimization remain effective when combined with spectral filtering mechanisms from JGCF or SimGCF? I would like to see this ablation to clarify whether the improvements mainly come from the filter design or the optimization strategy.

Reviewer_EtjkOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ASPIRE, a spectral graph collaborative filtering method that learns adaptive graph filters. The authors observe that directly learning a polynomial graph filter with BPR loss can cause low-frequency explosion, where the filter increasingly amplifies low-frequency components and recommendation performance deteriorates. They also provide a theoretical analysis of the BPR gradients to explain this behavior.

To address this problem, ASPIRE introduces pre-filter normalization and bi-level optimization. Specifically, user and item embeddings are learned using the training set, while the polynomial filter coefficients are learned using the validation set. This separates embedding learning from filter learning and aims to prevent low-frequency explosion.

Experiments on several recommendation datasets show that ASPIRE improves recommendation performance over existing baselines. The authors also provide ablation studies and analyses of the learned filters to examine the effectiveness of the proposed components.

Critical strengths
  • S1. The paper identifies an interesting problem in learnable graph filters. Directly optimizing the filter with BPR loss can lead to low-frequency explosion and degraded recommendation performance. The theoretical analysis provides useful insights into why this behavior may occur.

  • S2. The proposed method, ASPIRE, addresses this problem through pre-filter normalization and bi-level optimization. The experiments show consistent performance improvements across different datasets and graph settings. The ablation studies and analyses of learned filters further support the effectiveness of the proposed method.

Critical weaknesses
  • W1. Theorem 1 relies on strong assumptions, such as perfect uniformity, which may not hold in practice. More importantly, the theorem analyzes gradients with respect to individual spectral responses, while the actual method learns polynomial filter coefficients shared across all frequencies.

  • W2. The analysis focuses only on BPR loss. Does low-frequency explosion also occur with other recommendation losses, such as Sampled Softmax (SSM)?

    • On the Effectiveness of Sampled Softmax Loss for Item Recommendation (TOIS 2024)
  • W3. The motivation for using validation-based bi-level optimization could be clearer. Both training and validation use the same BPR loss, so it is not clear why optimizing the filter on validation data avoids the low-frequency explosion observed when optimizing it on training data.

  • W4. If BPR loss causes low-frequency explosion, a more straightforward solution might be to modify the loss function itself. The paper does not sufficiently explain why bi-level optimization is a better choice than directly addressing the problem through the training objective.

  • W5. It is unclear whether the performance gains come from preventing low-frequency explosion or simply from making training more stable. If training stability is the main issue, BPR itself may not be the root cause, and simpler stabilization techniques might achieve similar improvements.

  • W6. ASPIRE uses validation data to directly learn the filter coefficients, while other methods generally use validation data only for hyperparameter tuning. This raises a question about the fairness of the comparison, since ASPIRE may benefit from additional supervision that the baselines do not receive.

  • W7. Standard deviations or significance tests would make the empirical results more convincing.

Questions for authors
  • Q1. Could the authors clarify how the assumptions in Theorem 1 relate to practical training? How does the analysis of individual spectral responses explain the behavior of the learned polynomial filter coefficients?

  • Q2. Does low-frequency explosion also occur with other recommendation losses, such as Sampled Softmax (SSM)? Could the authors further examine whether their observations and theoretical analyses hold for SSM as well?

  • Q3. Why does optimizing the filter on validation set prevent low-frequency explosion when both training and validation use BPR loss? Is there any evidence that the two objectives produce different spectral gradients?

  • Q4. Could the authors explain why bi-level optimization is preferable to simpler alternatives, such as regularizing the filter coefficients or constraining the spectral response, etc.?

  • Q5. To what extent do the performance improvements come from preventing low-frequency explosion rather than simply improving training stability?

  • Q6. Could the authors clarify whether the baselines have access to comparable supervision, given that ASPIRE directly uses validation interactions to learn its filter coefficients? Could the authors include results for baselines that uses both training and validation sets for optimization?

  • Q7. Could the authors provide standard deviations or statistical significance tests for the main results?

Reviewer_jpZwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies why learnable spectral graph filters fail under standard recommendation objectives in collaborative filtering (CF). It identifies a training pathology termed "low-frequency explosion": under BPR (and, in the appendix, cross-entropy) supervision, the learned filter's low-frequency response grows without bound and ranking metrics oscillate or collapse. Theorem 1 argues that, under a Perfect Uniformity condition (Wang et al., 2022) and a negligibility assumption on a deviation term δ, the BPR gradient with respect to the spectral response g(λ_k) is proportional to λ_k·g(λ_k)·‖Ẽ_k‖² near an optimum. The proposed remedy, ASPIRE, combines (i) row-wise ℓ₂ normalization of embeddings before filtering, (ii) a bi-level formulation in which the filter coefficients θ are optimized against a validation loss while embeddings E are optimized on the training split, and (iii) a first-order approximation that drops the implicit gradient term. Experiments cover four datasets (Amazon Baby, Electronics; Ciao, LastFM) across multimodal and social graph scenarios, comparing six fixed/learnable filter baselines, plus an extension to LLM-initialized CF.

Critical strengths
  1. A crisp, well-illustrated phenomenon. Figure 1 and the appendices convincingly document the divergence of the low-frequency response and the accompanying metric collapse for a naive learnable filter, under both BPR and CE losses. The quantitative |g_max| trajectories are a useful diagnostic, and the phenomenon is a genuine practical obstacle confirmed by independent prior reports (e.g., He et al., KDD 2025).
  2. An explicit gradient-sign analysis. Theorem 1 (and its CE analogue) is a concrete attempt to attribute the pathology to the loss rather than to architecture or optimization quirks; the sign structure λ_k·g(λ_k) is a plausible mechanistic explanation for the observed low-frequency drift.
  3. A simple, cheap remedy with practical value. The fix adds essentially no capacity (θ has L+1 = 4 scalar parameters) and no architectural complexity; the reported gains over manually designed filters under a unified configuration are, if correct, practically useful.
Critical weaknesses
  1. The theoretical analysis does not cover the regime it is meant to explain. Theorem 1 is stated only for a nearly optimized model under a Perfect Uniformity assumption, but the low-frequency explosion it is invoked to explain is visible from the first training epochs, and the theorem's own hypothesis excludes the high-frequency directions the paper claims are under-optimized. The assumptions are never validated empirically. This is consequential because the theorem is the paper's stated motivation for the method; with the explanatory chain broken, the contribution rests entirely on empirical evidence.

  2. The empirical evidence cannot attribute the gains to the proposed mechanism. After the implicit term is dropped, the "bi-level" method reduces to periodic validation-supervised training of only 4 filter parameters, yet there is no random/grid-search control at matched budget and no ablation isolating the l2 normalization from the learned filter; all baselines are re-implemented by the authors with no external anchor, and no standard deviations are reported. This is consequential because the central claim — that learning the filter beats designing it — is indistinguishable from a trivial search or the normalization alone at the evidence level provided.

Questions for authors
  1. Theorem 1 is stated for a nearly optimized model under Perfect Uniformity, but the low-frequency explosion appears from the first epochs. Can you either provide an analysis that covers the early training regime, or show measured evidence (uniformity statistics, gradient magnitudes vs. the predicted lambda_k * g(lambda_k) * ||E_k||^2 profile along the trajectory) that the assumptions hold when the explosion begins? If neither is possible, my concern about the theory-to-method link stands.

  2. The filter theta has only 4 scalar parameters. How does ASPIRE compare to random or grid search over theta tuned on the same validation half at a matched evaluation budget? This is the decisive control for the claim that learning the filter beats designing it.

  3. What is the per-component contribution of the l2 pre-filter normalization versus the learned theta? Please provide the 2x2 ablation (normalization x {fixed handcrafted filter, learned theta}) on all four datasets.

  4. In Table 11, replacing validation supervision with training-set supervision ("w/o OS") drops N@20 from .0380 to .0254 on Baby. Does the method's benefit derive mainly from access to clean held-out interactions? What is the sensitivity to the 50/50 split and to the update interval T=5, and what exactly is the checkpoint selection criterion for the reported test numbers?

  5. Please report per-dataset standard deviations over your five runs for Tables 3 and 5. Which improvements over Lin-C and Jacobi remain statistically significant?

  6. Claim inconsistency: the abstract states "excellent recommendation performance," the contribution list states "comparable performance with handcrafted filters," and Table 3 shows ASPIRE losing to handcrafted filters in several cells (e.g., Baby-Homogeneous R@10 .0550 vs Jacobi .0566) with negative gains on Electronics-Heterogeneous. Please align the claims with the table.

  7. In the LLM-extension (Table 5, Whitening-Init), the naive learnable filter matches or beats the fixed Avg-P filter (SFR-Emb: .0968/.0430 vs .0961/.0424), which appears to contradict the universality of the pathology that motivates the paper. Footnote 1 also notes test and validation rankings disagree. Under what conditions does low-frequency explosion actually occur?

  8. The Level rubric ("75% reach Level 2") depends on a custom rule that can mark a result Level 0 even when it satisfies a higher level. Please report plain win/tie/loss counts per baseline alongside it.

  9. Evaluation is limited to side-information settings with self-implemented baselines. Could you clarify whether the pathology and the fix transfer to standard ID-based spectral CF benchmarks?

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

Related papers

Discussion (0)

Sign in to comment.