From ID-based to ID-free: Rethinking ID Effectiveness in Multimodal Collaborative Filtering Recommendation

Guohao Li, Jing Li, Jia Wu, Xuefei Li, Kai Zhu, Yue He

TL;DR

This paper argues that traditional ID-based features in multimodal recommendation systems offer limited benefits and can even hinder performance. They introduce IDFREE, an ID-free recommendation baseline that replaces ID embeddings with rich multimodal features and positional encodings, achieving significant performance gains.

What Problem Does This Paper Address?

Multimodal Collaborative Filtering Recommendation (MCFRec) aims to improve recommendations by leveraging not only user-item interactions but also content from various modalities like text and images. Most existing MCFRec methods rely heavily on user and item IDs as primary features, using multimodal content as supplementary information. However, this paper questions the necessity and effectiveness of these ID features, suggesting they might lack semantic richness, impede generalization to new data, and even cause representation shifts. The core problem is to determine if MCFRec can achieve superior performance by completely abandoning ID features and fully embracing multimodal content.

Methods

The proposed solution, IDFREE, is an ID-free MCFRec baseline that tackles the problem in several key ways:

  • ID-free Embedding: Instead of using user/item IDs, IDFREE uses multimodal features (text and image) as initial embeddings. To maintain unique identifiers, it incorporates Positional Encoding (PE), inspired by Transformer architectures, to generate semantically rich, ID-free embeddings for users and items.
  • Adaptive Similarity Graph (ASG): To dynamically capture user-user and item-item relationships based on multimodal features, IDFREE proposes an ASG module. This module constructs user-user and item-item graphs based on multimodal embeddings, allowing for more adaptive and context-aware graph structures compared to static, pre-computed graphs.
  • Augmented User-Item Graph Encoder (AGE): This module fuses user and item multimodal features and then enhances the user-item interaction graph using the dynamically generated graphs from ASG. It employs a LightGCN-like architecture for effective encoding of users and items.
  • ID-free Loss: IDFREE utilizes a two-pronged loss function:
  • Multimodal Alignment Loss: This uses InfoNCE loss to ensure semantic alignment between different modalities (text and image) for both users and items, promoting a unified understanding.
  • Recommendation Loss: This employs Softmax loss, which the authors found to be more effective than traditional BPR loss in their experiments.

Datasets

The experiments were conducted on three public Amazon review datasets:

  • Baby
  • Sports
  • Clothing

These datasets were pre-processed to extract 4096-dimensional visual features and 384-dimensional textual features.

Key Insights and Ideas

  • ID features are a bottleneck, not a boost: The paper convincingly argues that while ID features are effective, their benefits in MCFRec are limited. They can hinder generalization and might not be as semantically rich as multimodal content.
  • Multimodal content is king: By replacing ID features with rich multimodal information (text and images) and using positional encodings, IDFREE achieves superior performance, demonstrating the power of truly multimodal representations.
  • Dynamic graphs are crucial: The Adaptive Similarity Graph module is key to capturing evolving relationships between users and items based on their multimodal content, leading to more accurate recommendations.
  • Semantic richness matters: Visualizations show that ID-free embeddings are more semantically rich and have broader distributions than ID-based embeddings, allowing for better differentiation and personalized recommendations.
  • Alignment is key: The multimodal alignment loss is vital for ensuring that different modalities speak the same semantic language, which is crucial for effective fusion and recommendation.

Key Results

  • Significant Performance Gains: IDFREE outperforms existing ID-based MCFRec methods by an average of 72.24% across standard metrics (Recall@K and NDCG@K for K={5, 10, 20, 50}) on the three datasets.
  • Module Effectiveness: Ablation studies show that all proposed modules (PE, ASG, AGE, Lalign) significantly contribute to IDFREE's performance, with PE having the most impact.
  • Hyperparameter Sensitivity: The model shows sensitivity to hyperparameters like k (graph augmentation), p (denoising ratio), and τ (temperature parameter), with optimal values identified for each.
  • Generalizability: IDFREE demonstrates strong generalizability, effectively alleviating generalization problems compared to ID-based methods.
  • Cold Start Performance: IDFREE significantly outperforms SOIL in item cold-start scenarios, highlighting the advantage of semantically rich, ID-free embeddings.
  • Modality Contribution: Textual features were found to contribute more to recommendation performance than visual features, though both are essential.

Limitations and Future Work

  • Training Time: While IDFREE has fewer parameters than some baselines, its training and inference times are slightly longer. The authors attribute this to the more complex graph construction and encoding modules.
  • Recommendation Length: While IDFREE excels at various recommendation lengths, the performance improvement slightly diminishes for longer recommendation sequences, suggesting room for improvement in recommendation coverage.
  • Future Work: The paper implicitly suggests future work could focus on optimizing the computational efficiency of the ID-free embedding and graph modules, and further exploring the interplay and optimal fusion strategies for different multimodal features. The success of IDFREE also opens avenues for applying ID-free principles to other recommendation tasks beyond MCFRec.

Summary updated

Abstract

Most existing multimodal collaborative filtering recommendation (MCFRec) methods rely heavily on ID features and multimodal content to enhance recommendation performance. However, this paper reveals that ID features are effective but have limited benefits in multimodal collaborative filtering recommendation. Therefore, this paper systematically deconstruct the pros and cons of ID features: (i) they provide initial embedding but lack semantic richness, (ii) they provide a unique identifier for each user and item but hinder generalization to untrained data, and (iii) they assist in aligning and fusing multimodal features but may lead to representation shift. Based on these insights, this paper proposes IDFREE, an ID-free multimodal collaborative Filtering REcommEndation baseline. IDFREE replaces ID features with multimodal features and positional encodings to generate semantically meaningful ID-free embeddings. For ID-free multimodal collaborative filtering, it further proposes an adaptive similarity graph module to construct dynamic user-user and item-item graphs based on multimodal features. Then, an augmented user-item graph encoder is proposed to construct more effective user and item encoding. Finally, IDFREE achieves inter-multimodal alignment based on the contrastive learning and uses Softmax loss as recommendation loss. Basic experiments on three public datasets demonstrate that IDFREE outperforms existing ID-based MCFRec methods, achieving an average performance gain of 72.24% across standard metrics (Recall@5, 10, 20, 50 and NDCG@5, 10, 20, 50). Exploratory and extended experiments further validate our findings on the limitations of ID features in MCFRec. The code is released at https://github.com/G-H-Li/IDFREE.
BibTeX
@misc{li2025from,
  author = {Guohao Li and Jing Li and Jia Wu and Xuefei Li and Kai Zhu and Yue He},
  title = {{From ID-based to ID-free: Rethinking ID Effectiveness in Multimodal Collaborative Filtering Recommendation}},
  year = {2025},
  eprint = {2507.05715},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.05715},
}
Figure