acceptodds
Under review as a conference paper at ICLR 2027

MM-DAGNN: Modality-Disentangled Adaptive Graph Neural Networks for Multimodal E-commerce Recommendation

Abstract

We proposed MM-DAGNN, a Modality-Disentangled Adaptive Graph Neural Network designed to address key challenges in multimodal e-commerce recommendation systems. Specifically, MM-DAGNN tackles the limitations of traditional models that assume uniform modality propagation, static fusion, and entangled projections. The proposed model integrates (i) modality-specific receptive fields Kv , Kt, Kid, (ii) a per-item adaptive fusion gate, and (iii) a contrastive disentanglement regularizer to improve the handling of heterogeneous modalities, against seven competitive baselines on the public AliExpress dataset. We treat country aggregated implicit feedback as pseudo-users and apply 5-core filtering, yielding a 177 × 5,073 bipartite graph with 105,419 edges. Three findings stand out. (F1) On Recall@20 and on every secondary metric, MM-DAGNN is competitive. MMGCN (0.0238) and LATTICE (0.0236) achieve the best Recall@20, while MM-DAGNN is the strongest method on Recall@10, NDCG, and MAP@20, a pattern consistent with recent reproducibility findings in collaborative filtering. (F2) A 4×4 grid over (Kv , Kt) shows a near-isotropic accuracy surface: the Recall@20 maximum sits at the diagonal (Kv =1, Kt=1) and the NDCG@2016 maximum at (Kv =2, Kt=2), contradicting the recent claim that visual and textual modalities require systematically different propagation depths on this dataset. (F3) The adaptive gate produces a high Modality Balance Index (MBI = 0.70) without harming accuracy, indicating that the gate genuinely uses all three streams rather than collapsing onto one. By solving the problem of modality imbalance and improving the adaptation across different streams, we contribute to a more flexible and accurate approach to multimodal recommendation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Recommender Systems, Graphs, Neural Networks, Data Fusion, Multimodal

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_yisyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MM-DAGNN, which combines modality-specific graph depths, shared/specific linear projections, a per-item softmax fusion gate, and a contrastive/orthogonality regularizer. It evaluates the model on a transformed AliExpress review dataset in which countries are treated as pseudo-users. The paper's main empirical conclusion is largely negative: the proposed components do not improve recommendation accuracy, modality-specific propagation depths do not outperform uniform depths, and a frozen item-item graph helps more than the proposed changes. The authors instead emphasize a proposed Modality Balance Index.

Critical strengths
  • The paper reports negative ablation results instead of hiding them.
Critical weaknesses
  • The proposed techniques are fairly simplistic, and, ironically, the ablation results indicate that some of them are actually detrimental.
  • The experimental evaluation is difficult to take seriously: it is conducted on only one dataset, which lacks a genuine visual modality, while relying mostly on baselines from several years ago.
  • The writing is poor and unfocused. The abstract alone contains a considerable amount of irrelevant detail instead of clearly presenting the key motivation, method, and findings.
Questions for authors

None

Reviewer_2986OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Modality-Disentangled Adaptive Graph Neural Networks (MM-DAGNN) for multimodal e-commerce recommendation. Specifically, it combines modality-specific graph propagation depths, shared/specific linear projections, a per-item fusion gate, and a contrastive disentanglement regularizer. Experiments on the AliExpress dataset show that MM-DAGNN performs relatively well on several ranking metrics.

Critical strengths
  1. Separating the treatment of visual, textual, and ID information is sensible, and the propagation-depth sweep provides a useful empirical diagnostic.
  2. The disclosed pseudo-user construction, and the provided public code link make the study easier to scrutinize.
Critical weaknesses
  1. Figure 1 seems too schematic to explain the complete method. It depicts three GCN branches and a gate but omits the fixed-weight user representation, the contrastive/orthogonality losses, and how shared versus specific features enter training.
  2. Several references appear to contain incorrect information. For example, the AAAI 2025 model MIG-GT is incorrectly referred to as MIGT, and its author list does not match the published record. Similar author inconsistencies are also found in the references to COHESION (SIGIR 2025), and MENTOR (AAAI 2025)
  3. The paper acknowledges that the interaction graph is unusually dense and that the visual features partially overlap with the country information in the graph. Therefore, the results in Figures 3 and 4 may not be sufficient to challenge the benefits of modality-specific propagation depths reported in previous studies.
  4. Section 4.1 aggregates individual users into 177 country pseudo-users, and Section 4.2 constructs the visual modality from review statistics and product categories rather than actual images. These settings may limit the generalizability of the findings to personalized multimodal recommendation.
  5. In Table 3, removing modality-specific depths improves both Recall@20 and NDCG@20. Removing the adaptive gate or disentanglement loss also improves NDCG@20, and adding an item-item graph yields a larger gain. These results raise some questions about the contributions of the three proposed components.
  6. It seems that only one transformed dataset is used, 157 pseudo-users are included in the held-out evaluation, and each result is based on two random seeds. Hence, the experimental evidence may rather narrow.
Questions for authors
  1. Could the authors further validate the conclusions about modality-specific propagation depths on sparser datasets with genuine visual and textual features?
  2. What motivates the use of country-level pseudo-users and review-derived visual features? How well do these settings reflect real-world personalized multimodal recommendation?
  3. Could the authors further clarify the contributions of the three proposed components?
  4. Could the authors conduct experiments on additional datasets with more random seeds to better demonstrate the robustness and generalizability of the findings?
Reviewer_Z1cKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript is labeled as a NeurIPS 2026 submission despite being submitted for ICLR review, raising serious concerns about compliance with ICLR submission requirements. I recommend that the Area Chair investigate this discrepancy and consider desk rejection if a violation is confirmed.

Critical strengths

None

Critical weaknesses

None

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.