acceptodds
Under review as a conference paper at ICLR 2027

Beyond Superficial Cues: Mitigating Popularity-Amplified Presentation Bias in Multimodal Recommendation

Abstract

Multimodal recommendation systems significantly enhance user experience by leveraging rich multimodal features. However, their effectiveness is often undermined by presentation bias, a prevalent issue where user interactions are disproportionately influenced by superficial presentation cues (e.g., flashy thumbnails, sensational titles) rather than intrinsic content. Such cues are often highly correlated with item popularity, leading to a popularity-amplified presentation bias that compromises recommendation fairness. To address this fundamental challenge, this paper introduces a novel Intrinsic **CO**ntent **RE**presentation (**CORE**) framework. We formalize presentation bias within a Structural Causal Model (SCM), revealing that conditioning on fully coupled multimodal features acts as a collider, inducing spurious correlations between intrinsic content and presentation manner. CORE mitigates this collider bias by disentangling an item's multimodal representation into two independent components: a content-aware representation and a bias-aware representation. This is achieved through a carefully designed automated contrastive learning scheme and enforced by an independence constraint. By performing recommendations based solely on the purified intrinsic content representation, our model effectively performs a causal intervention to mitigate the spurious effects. Extensive experiments on three real-world multimodal datasets demonstrate that CORE achieves significant and consistent improvements in both recommendation accuracy and fairness.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Recommendation, Causal Inference, Representation Disentanglement, Contrastive Learning, Debiasing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LZsuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes CORE, short for COntent REpresentation, a novel approach for multimodal recommendation. The main rationale of the methodologies lays in the presence of two principal signals in multimodal recommendation, namely, multimodal content and presentation bias; in this respect, the authors claim the latter is directly related to popularity bias of items and largely influences recommendation performance. To better formalize the scenario, the authors leverage causal graphs, and propose to learn two disentangled representations of content and presentation for each item. To begin with, the model encodes the visual and textual modalities with a learnable multi-layer perceptron preceded by a concatenation of the two modalities. After that, it uses two separate encoders (again, implemented as multi-layer perceptrons) for the content and presentation representations, and establish a twofold contrastive learning paradigm: the first one accounts for the content, and ensures items with same content but different presentation should be close in the latent space; the second one accounts for the presentation, and ensures items with same popularity but different content are close in the latent space. Then, an independence loss built around the Hilbert-Schmidt Independence Criterion (HSIC) ensures the independence of the two learned representations. Finally, the methodology adopts a LightGCN predictor to compute the user-item scores, and train a BPR-like loss function augmented with the above-described loss components. Results on three recommendation datasets and against nine state-of-the-art recommendation models spanning different families of approaches demonstrate the efficacy of the CORE methodology. Additional ablation studies and hyper-parameter sensitivity analyses further confirm the goodness of the architectural choices, while a visual embedding representation demonstrates the different learning capabilities of the content and presentation embeddings.

Critical strengths

I think the paper has the merits to identify a critical problem in multimodal recommendation, namely, how presentation bias might largely drive recommendation at the detriment of multimodal content. This is also supported by related literature, and the authors nicely formalize this under the causal inference paradigm. The methodology is clearly explained and quite simple, which I identify as another strong point in favor of the paper.

Critical weaknesses

I think that one of the main critical points of the paper lays in the design of the heuristics to select items with similar content but different presentations, especially as described in Section 4.3.1. I believe that using the textual modality to ensure two items are similar content-wise, but with different variations of the same image to say the two items are dissimilar presentation-wise, is a too much simplistic assumption. First, while I understand the visual modality might be one of the main driving factors for products on e-commerce, this might not be always the case in other domains, such as music recommendation. Additionally, I think that also the textual content might sometimes lead to presentation bias, for instance with items' descriptions that enhance the item's true content with exaggerated characteristics and details. While I acknowledge the authors provide a preliminary discussion on these aspects in Section A.2 from the appendix, I would expect further clarifications on this point, that is crucial for the whole design of the paper.

Experiments-wise, the paper lacks further multimodal baselines, where VBPR and MMGCN are quite outdated methods. I would suggest adding other more recent baselines to the benchmark, such as PGL [1], SMORE [2], and COHESION [3] to ensure the timeliness of the proposed model and its empirical evaluation.

References

[1] Penghang Yu, Zhiyi Tan, Guanming Lu, Bing-Kun Bao: Mind Individual Information! Principal Graph Learning for Multimedia Recommendation. AAAI 2025: 13096-13105

[2] Rongqing Kenneth Ong, Andy W. H. Khong: Spectrum-based Modality Representation Fusion Graph Convolutional Network for Multimodal Recommendation. WSDM 2025: 773-781

[3] Jinfeng Xu, Zheyu Chen, Wei Wang, Xiping Hu, Sang-Wook Kim, Edith C. H. Ngai: COHESION: Composite Graph Convolutional Network with Dual-Stage Fusion for Multimodal Recommendation. SIGIR 2025: 1830-1839

Questions for authors

My main questions are related to the above concerns.

  1. Can the authors elaborate further on the choice of the heuristics to flag items according to presentation? Additionally, did the authors try to run the same empirical pipeline of CORE, by considering another modality as the one used for presentation? So, instead of images, using text for the contrastive learning part.

  2. Did the authors try to assess the performance of CORE on datasets with more than 2 modalities? What are the main methodological and performance implications? As an example, authors might refer to this recent dataset

∗*

.

References

∗*

Y. Ni, Y. Cheng, X. Liu, J. Fu, Y. Li, X. He, Y. Zhang, and F. Yuan, “A content-driven micro-video recommendation dataset at scale,” in CIKM, 2025.

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

This paper studies presentation bias in multimodal recommendation, where user interactions may be driven by superficial presentation cues rather than intrinsic item content. The authors formulate the observed multimodal feature as a collider between latent intrinsic content and presentation manner and propose CORE to separate a fused multimodal representation into a content representation and a bias representation.

Critical strengths
  1. The paper highlights an important and underexplored problem. Multimodal features can encode superficial presentation signals as well as semantic content, and these signals may reinforce popularity concentration. The proposed representation-level treatment is intuitive, relatively simple, and potentially compatible with existing recommender architectures.

  2. The reported evaluation is reasonably broad. The paper evaluates three datasets, several accuracy and exposure-diversity metrics, multiple classes of baselines, component ablations, sensitivity analyses, and popularity-group performance. The reported improvements are consistent across the three datasets and suggest that the proposed objectives may be useful as practical regularizers.

Critical weaknesses

The causal interpretation is not sufficiently justified. Enforcing independence between zcz_c and zsz_s does not identify the latent variables CC and SS, and excluding zsz_s at inference does not necessarily estimate P(Y∣do(C))P(Y \mid do(C)). The causal graph and structural equations are also inconsistent regarding X→YX \to Y.

The experiments do not clearly distinguish presentation bias from popularity bias. Popularity is used as a proxy for presentation style, but the datasets contain no controlled presentation changes or presentation labels. Some implementation details are ambiguous, and the results lack uncertainty estimates.

Questions for authors
  1. Is the content-positive sample a cross-modal synthetic item or another complete item selected by Algorithm 1?
  2. What justifies interpreting prediction with zcz_c as an approximation to P(Y∣do(C))P(Y \mid do(C))?
  3. Can the authors provide direct evidence that zcz_c removes presentation information rather than general popularity information?
  4. Please report standard deviations and clarify the inconsistent percentage comparisons in Table 4.
Reviewer_RBjWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework to reduce presentation bias in multimodal recommendation. It fuses visual and textual features and separates them into content and bias representations using contrastive learning and an independence penalty. A graph neural network then predicts user preferences using only the content representation. The paper reports top accuracy and fairness scores across multiple datasets.

Critical strengths
  1. Relevant problem motivation. Separating item content from superficial appeal in multimodal recommendation is a practical problem.
  2. Simple representation learning recipe. Using category labels and popularity statistics to guide representation branches with independence regularization is easy to implement without manual label.
Critical weaknesses
  1. Flawed causal argument and contradictory method description. Making representations independent via regularization does not perform a true causal intervention. The model still trains on observed clicks affected by presentation bias. Furthermore, the method actually targets popularity rather than presentation, and its positive-pair construction contradicts itself between the main text and the algorithm.
  2. Mathematically impossible results and unclear baselines. Group-level accuracy changes cannot mathematically yield the reported overall gain. Reported coverage and diversity metrics violate basic mathematical bounds. Dataset statistics match raw unfiltered data despite claims of filtering.
Questions for authors
  1. How can overall accuracy gains exceed the gains of every individual subgroup? How do the reported coverage and diversity metrics satisfy basic mathematical bounds?
  2. Why do dataset statistics match raw unfiltered data? Did your implementation swap images or select whole items for positive pairs?
  3. How were social and baseline models adapted without proper inputs? Can you explain the numerous misattributed references?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.