acceptodds
Desk rejected at ICLR 2027

CSA-RIT: Learning Noise-Resilient Multimodal Recommendation via Coarse Semantic Alignment and Robust Interest Tree

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph-based multimodal recommenders typically build an item-item graph per modality and aggregate neighbor representations by graph convolution. Two things decide the outcome of this aggregation. The first is what the modality features being averaged express, and the second is which items enter a neighborhood. Both are settled before the first propagation step, and repeated smoothing spreads any error into every representation. On the feature side, an item's visual and textual features are typically extracted independently and often inconsistent. We call this content noise. On the graph side, existing methods often introduce noisy links when building item-item graphs. We call this structural noise. We propose CSA-RIT, a noise-resilient framework that treats content and structural noise as two sources of a single pre-propagation error and repairs both. Coarse Semantic Alignment (CSA) places a coarse discrete bottleneck ahead of graph propagation and aligns the same item across modalities only in a shared coarse code, so that the signals to be averaged share a common semantic basis. Robust Interest Tree (RIT) augments each frozen modality graph with a short, order-aware co-occurrence tree, supplying a stable prior that resists this structural noise. On three standard MMRec benchmarks spanning catalog sizes from 7k to 63k items, CSA-RIT is best on every metric, improving over the strongest baseline by 8.9% on average and by 13.7% on the largest split. Ablations confirm complementary gains, and RIT keeps its recall under injected structural noise. Our implementation is available at https://github.com/paper-ai212/CSA-RIT.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multimodal Recommendation, Interest Tree, graph convolution

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9TiyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CSA-RIT for noise-resilient multimodal recommendation. Coarse Semantic Alignment introduces a shared encoder–decoder and finite scalar quantization before graph propagation, with instance-level cross-modal consistency. Robust Interest Tree augments frozen modality graphs using co-occurrence-based neighborhoods and parameter-free fusion. Experiments on three Amazon datasets report improved ranking performance, supported by module ablations, structural corruption tests, and alignment trajectories.

Critical strengths

1, A relevant and coherent motivation. Distinguishing errors in the features being aggregated from errors in neighborhood selection provides an intuitive framework for improving graph-based multimodal recommendation. The two modules address complementary aspects of this aggregation process.

2, Promising empirical results and useful diagnostics. The model improves over the listed baselines across all reported metrics, including the larger Electronics dataset. The structural corruption experiments directly examine robustness, while the alignment trajectories acknowledge that increasing cross-modal similarity beyond the validation optimum does not necessarily improve recommendation.

Critical weaknesses

1, The incremental contribution over closely related work is insufficiently established. TAMER (Meng et al., MM 2025) already combines the same multi-component feature preprocessing with interest-tree-based modality graph augmentation. However, it is cited for preprocessing without a substantive methodological comparison or an explicit baseline entry. Its published Recall@20 scores are 0.1123/0.1272/0.0808 on Baby/Sports/Electronics, compared with 0.1122/0.1279/0.0820 here. Although matched reproductions are necessary, these results make the omission consequential. Furthermore, MoToRec already investigates discrete multimodal tokenization, and AlignRec addresses cross-modal misalignment. Controlled comparisons are needed to establish the specific benefits of the proposed coarse bottleneck and revised tree construction.

2, The experiments do not sufficiently isolate the claimed mechanisms or ensure comparison fairness. The module-level ablations do not separate quantization from continuous encoder–decoder alignment, or tree search from simpler co-occurrence augmentation. Structural corruption affects modality graphs while leaving the co-occurrence source intact, so resilience may partly reflect access to an uncorrupted information source; an equally informed control is needed. These tests use one seed and only Baby. Main-table uncertainty is also missing. Additionally, the claim of a common frozen-feature protocol requires clarification: the original MoToRec and AlignRec use different feature encoders. The paper should distinguish reproduced results from published numbers and explicitly confirm training-only construction of co-occurrence graphs.

Questions for authors

How does RIT differ from TAMER’s interest-tree augmentation, and how do the methods compare under identical preprocessing, splits, and tuning budgets?

Can continuous alignment and simple co-occurrence fusion controls isolate the contributions of quantization and order-aware tree search?

Which baseline results were rerun with identical features? Please provide multi-seed uncertainty and confirm that graph construction excludes validation/test interactions.

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

This paper studies multimodal recommendation based on frozen modality graphs. The authors categorize the errors affecting graph propagation into two types: "content noise" arising from inconsistencies across different modality features, and "structural noise" caused by erroneous neighbors introduced during modality similarity graph construction. The authors propose CSA-RIT, which separately addresses features and adjacency relations before propagation: Coarse Semantic Alignment (CSA) maps images and two text views into coarse-grained discrete representations through a shared encoder, a finite scalar quantizer, and a decoder, while imposing instance-level cross-view consistency constraints. Robust Interest Tree (RIT) constructs a short-range interest tree from the item co-occurrence graph, preserving first-order neighbors and applying beam search to higher-order nodes, then fuses the interest tree support with the frozen modality graph. The model is jointly trained with a BPR recommendation loss, a quantization reconstruction loss, and a consistency loss. On Amazon Baby, Sports, and Electronics, the full model achieves the best results across all reported metrics. The paper also provides component ablation studies, graph edge perturbation experiments, and training trajectory analyses of coarse-grained representation alignment versus validation set performance.

Critical strengths

S1. The results cover three datasets of varying scales, with consistent overall improvements. The model outperforms the compared methods in the table across all twelve reported metrics on the three datasets. After selecting the best baseline for each metric according to Table 2, the average relative improvement across the twelve metrics is approximately 8.94%; the average improvement across the four metrics for Electronics is approximately 13.68%, consistent with the abstract. The retention of approximately 63,000 items in Electronics also enhances the experiment's reference value for larger catalog scenarios. S2. The structural robustness experiments directly perturb graph edges, which is more convincing than inferring noise resistance solely from accuracy on clean data. Figure 2 employs random reconnection and low-similarity reconnection, respectively, maintaining row degrees and subjecting the compared methods to the same perturbed graphs. Under these settings, RIT's performance degradation is notably smaller than that of Backbone. This experiment supports a specific conclusion: the incorporation of the co-occurrence interest tree reduces the model's sensitivity to errors in modality graph edges. This constitutes the empirical part of the paper that holds greater value than mere module combination.

Critical weaknesses

W1. The comparison with the closest work and the definition of the backbone are insufficient, making it difficult to assess the novel contributions of RIT. The paper cites TAMER (ACM MM 2025), which has already proposed using interest trees to enhance modality graphs, capture high-order relationships, and reduce noisy connections—this is very close to RIT's research objectives. However, the current manuscript only cites this work in the feature preprocessing section, without adequately explaining: what specific changes RIT's tree construction, pruning, and fusion introduce relative to TAMER; whether the Backbone in Table 3 includes TAMER or any of its components; and how much improvement the new design achieves over the original interest tree enhancement scheme under identical features and training conditions. This is particularly important because the Backbone's Recall@20 on Electronics has already reached 0.0788, higher than the best comparison method's 0.0718 in Table 2; adding the full CSA-RIT further improves it to 0.0820, representing an approximate 4.06% improvement over the internal Backbone. Therefore, the overall advantage highlighted in the abstract cannot be directly equated with the contributions of the two new modules. It is necessary to clarify the backbone's source and conduct controlled comparisons with nearest-neighbor methods to reliably evaluate the paper's novelty and the magnitude of progress. W2. The existing experiments have not yet sufficiently isolated the mechanisms of "coarse-grained quantization repairing content noise" and "interest trees repairing structural noise." For CSA, Table 3 only compares the presence or absence of the entire module, without distinguishing whether the gains come from the shared encoder, reconstruction regularization, instance consistency loss, or discrete quantization itself. There is a lack of control conditions such as "continuous representation with the same consistency loss," "quantization only," and "alignment only." The correlation between alignment degree and Recall reported in Section 4.5 is also insufficient for mechanism attribution: both may increase with training progress, and the authors have observed that recommendation performance no longer improves as alignment degree continues to rise. It supports alignment degree as a training diagnostic, but cannot independently prove that content noise is eliminated. The manuscript also does not directly test performance under content feature perturbations or cross-modal mismatches. For RIT, the perturbation experiments retain additional co-occurrence information sources. Therefore, more stable results may also stem from increased reliance on the unperturbed auxiliary graph. Without direct co-occurrence graph fusion, first-order neighbor enhancement, and alternative schemes under the same edge budget, it is difficult to determine how much of the advantage comes from the specific tree structure and beam search. These issues do not negate the effectiveness of the complete system, but they limit the assessment of the necessity of the core designs.

Questions for authors

Q1. Please clarify what the Backbone in Table 3 is, and its relationship with TAMER. Could you provide a comparison of the original TAMER, Backbone, adding CSA, and adding RIT under the same protocol, and explain the algorithmic differences of RIT item by item? Q2. When network capacity, reconstruction loss, and training conditions are kept consistent, what are the respective results for continuous representations with ICR, quantized representations without ICR, and full CSA? Are there content perturbation experiments supporting "content noise repair," rather than merely general representation regularization? Q3. Compared with directly fusing the co-occurrence graph, adding only first-order co-occurrence neighbors, and high-order augmentation with matched edge counts, how much do the interest tree and beam search each contribute? Are these baselines equally stable under structural perturbations?

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

Traditional recommendation methods face two main problems: the visual and textual features of an item may be inconsistent, and the selected neighboring items may not actually be relevant. the authors propose CSA-RIT. CSA aligns the visual and textual features of the same item to make them more consistent, while RIT builds an item co-occurrence tree based on user interaction records and combines it with the original Top-K item similarity graphs. Finally, the model uses these combined signals to make recommendations, reducing the impact of noisy information.

Critical strengths

Addresses two types of noise: CSA reduces inconsistencies between visual and textual features, while RIT mitigates errors in item-item graphs.

Critical weaknesses

The Figure 1 appear to be potentially AI-generated, which raises concerns about the reliability and authenticity of the visual presentation. The core components—cross-modal feature alignment and co-occurrence-based graph enhancement—largely build on existing techniques. The paper fails to establish substantial methodological advances beyond their combination, particularly given the lack of direct comparison with TAMER, a closely related interest-tree-based recommender.

Questions for authors

Please refer to the weaknesses mentioned above.

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

Related papers

Discussion (0)

Sign in to comment.