Enterprise AI Analysis
CROSS: Feedback-Oriented Multi-Modal Dynamic Alignment in Recommendation Systems
Authors: Yang Li, Junpeng Du, Chenzhan Wang, Zunlong Liu, Xiaomin Zhu, Chen Lin
Affiliations: Xiamen University, Chongqing University, Shandong Normal University, Academy of Military Sciences, National Institute for Data Science in Health and Medicine.
Published: March 2026
Executive Impact
This research introduces CROSS, a plug-and-play framework for multi-modal recommendation systems that significantly boosts performance by dynamically aligning modalities and incorporating collaborative signals. It addresses critical challenges in conventional recommendation systems and prior multi-modal approaches.
Aligning the multi-modal content and ID embeddings is crucial in multi-modal recommendation systems. Existing solutions typically adopt a bidirectional alignment paradigm. Our prior work, FETTLE, challenges this paradigm by proposing a one-way directional alignment at the item level, thus reducing the negative impact of low-quality modalities. However, FETTLE leaves two open questions: (1) when is one-way directional alignment optimal, and (2) how to incorporate collaborative signals to enhance alignment? We present CROSS (feedbaCk-oRiented multi-mOdal alignment in recommendation SyStem), a plug-and-play framework that extends FETTLE by introducing three major advancements. First, we introduce Dynamic Item-Level Alignment, which dynamically calibrates the “strength” of each modality via a variance-based compensation mechanism, mitigating the risk of overshadowing weaker modalities in the early stages of training. Second, we develop Multi-grained Collaborative Alignment, which introduces a medium-granularity alignment strategy based on neighboring items that share similar user feedback profiles. This neighbor-level alignment effectively balances noisy user interactions and excessive smoothing across items. Third, we conduct extensive experiments on more real-world datasets and show that CROSS significantly boosts the performance of both collaborative filtering (CF) models and multi-modal recommendation (MRS) approaches, achieving 21.52%-70.78% average improvement on CF backbones and 8.70%-20.73% on MRS backbones. Compared with FETTLE, CROSS achieves additional improvements of 3.82%-5.24%.
0%
Avg Improvement on CF Models (up to)
0%
Avg Improvement on MRS Models (up to)
0%
Additional Improvement vs. FETTLE (up to)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Adaptive Modality Alignment
CROSS introduces Dynamic Item-Level Alignment, which dynamically calibrates the contribution 'strength' of each modality. This variance-based compensation mechanism ensures optimal alignment strategies throughout the training process, transitioning from bidirectional to one-way alignment as confidence in modality strength increases. This mitigates the risk of weaker modalities being overshadowed in early stages and resolves conflicting directional signals via Multi-Modal Alignment.
Enterprise Process Flow: Dynamic Modality Alignment
Early Training Stage: High Uncertainty, Low Scores
→
Bidirectional Alignment Enabled
→
Later Training Stage: Low Uncertainty, High Scores
→
One-Way Directional Alignment Optimized
→
Optimal Multi-Modal Fusion
Adaptive
Modality Strength Calibration
CROSS dynamically adjusts modality 'strength' based on estimated user feedback variance. This prevents weaker modalities from being overshadowed in early training phases and evolves to optimal one-way alignment as the model matures, ensuring balanced information utilization.
Leveraging Collaborative Signals
CROSS significantly enhances alignment by incorporating Multi-grained Collaborative Alignment. This approach operates at the item-neighbor and item-cluster levels, allowing the model to balance noisy user feedback with more stable, aggregated signals. By identifying robust neighbors and enforcing uniform cluster distributions, CROSS mitigates the impact of irrelevant interactions and addresses popularity bias.
Case Study: Filtering Noise with Neighbor-Level Alignment
Consider User 46 from the Amazon Baby dataset, who interacted with both relevant items like booster seats and rocking chairs, but also unrelated diaper accessories (Figure 5 from the paper). Such noisy interactions can adversely affect item-level alignment decisions.
CROSS's Item Neighbor-level Alignment addresses this by identifying robust co-occurring neighbors for Item 7049, such as night inserts, cloth diapers, and steam sterilization bags. By aligning modalities among these stable neighbors, CROSS filters out the noise from User 46's isolated misclick, thereby capturing the genuine interests of the user community and enhancing the robustness of the alignment.
| Approach | Key Mechanism | User Feedback Involvement | Robustness to Noise | Effect on Popularity Bias |
|---|---|---|---|---|
| Item-Level Alignment (FETTLE) | One-way directional alignment per item | Direct (can be noisy) | Lower | Minimal |
| Neighbor-Level Alignment (CROSS) | Aligns items with robust co-occurrence neighbors | Indirect (filtered by neighbors) | Higher | Mitigates for niche items |
| Item Cluster-Level Alignment (CROSS) | Aligns items to cluster prototypes via Sinkhorn algorithm | Implicit (cluster prototypes) | High | Significant (enforces uniform distribution) |
| User Cluster-Level Alignment (CROSS) | Aligns users to cluster prototypes | Implicit (abstract user interests) | High | Helps cold-start item recommendation |
Empirical Performance & Robustness
Extensive experiments on four real-world datasets (Amazon Baby, Sports, Clothing, TikTok) demonstrate CROSS's superior performance. It consistently boosts both collaborative filtering (CF) and multi-modal recommendation (MRS) models. Crucially, CROSS exhibits enhanced robustness in challenging scenarios such as missing/noisy multi-modal content and noisy user feedback, and effectively mitigates popularity bias, particularly for cold-start items.
| Category | Avg R@10 Improvement | Avg R@20 Improvement | Avg N@10 Improvement | Avg N@20 Improvement |
|---|---|---|---|---|
| CF Backbones (Overall Avg) | 39.71% | 36.55% | 37.32% | 36.15% |
| MRS Backbones (Overall Avg) | 14.22% | 12.55% | 15.88% | 14.71% |
| Vs. FETTLE (Overall Avg) | 3.82-5.24% | 3.32-4.23% | 4.71-7.90% | 3.94-7.59% |
Stronger
Performance with Noisy Data
CROSS exhibits superior robustness against missing/noisy multi-modal content and noisy user feedback. It boosts average R@10 by 1.85% compared to FETTLE in scenarios with noisy user feedback, demonstrating its resilience.
36.14%
R@10 Improvement for Tail 1 Items
The Multi-grained Collaborative Alignment significantly improves recommendations for cold-start (Tail 1) items, achieving a substantial 36.14% R@10 improvement, effectively addressing the long-tail phenomenon in recommendation systems.
Calculate Your Potential AI ROI
Estimate the efficiency gains and cost savings your enterprise could achieve by implementing advanced AI solutions based on this research.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A typical phased approach to integrating advanced recommendation AI into your enterprise systems.
Phase 01: Discovery & Strategy
Assess current recommendation systems, data infrastructure, and business objectives. Define clear KPIs and a tailored strategy for integrating CROSS's dynamic multi-modal alignment.
Phase 02: Data Integration & Preprocessing
Integrate multi-modal data (images, text) with existing ID embeddings. Implement projection layers to ensure dimensional consistency and prepare data for alignment mechanisms.
Phase 03: Dynamic Alignment Framework Deployment
Deploy CROSS's Dynamic Item-Level Alignment and Multi-grained Collaborative Alignment. Configure variance-based compensation and neighbor/cluster-level modules. Fine-tune hyperparameters for optimal performance.
Phase 04: Model Training & Evaluation
Train the integrated recommendation system with real-world datasets. Continuously evaluate performance against baselines using metrics like R@K and N@K, ensuring robustness against noise and bias mitigation.
Phase 05: Monitoring & Optimization
Implement continuous monitoring of model performance and data quality. Iteratively optimize alignment strategies and integrate new data sources to maintain high accuracy and adaptability.
Ready to Transform Your Recommendations?
Leverage the power of dynamic multi-modal alignment and collaborative intelligence to build more accurate, robust, and personalized recommendation systems for your enterprise.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat