Multimodal Gen Rec多模态生成式推荐

Overview

Traditional recommendation systems struggle with two fundamental challenges: long-tail items (rarely seen products) and cold-start users (users with limited interaction history). These systems typically rely on candidate-ranking pipelines that favor popular items, failing to surface the vast majority of available content.

We introduce CLAIRO (Collaborative Learning with Action-aware Image-text Representation Optimization), a multimodal generative recommendation system that reformulates recommendation as an autoregressive generation task. By extending ActionPiece’s context-aware tokenization framework with visual features, CLAIRO learns to merge co-occurring textual and visual patterns, creating richer item representations that improve recommendation accuracy while maintaining computational efficiency.

Duration: May 2025 - present

Institution: New York University Shanghai

Advisor: Prof. Hongyi Wen

Team Members: Zhaodong Liu, Yuquan Hu, Tuoye Liu

View Full Report for detailed methodology, comprehensive experimental results, and in-depth analysis.

Key Innovation

CLAIRO’s core innovation is multimodal token merging - extending ActionPiece’s BPE-inspired algorithm to jointly learn from both visual and textual features. Unlike existing approaches that treat modalities separately, CLAIRO discovers cross-modal co-occurrence patterns during vocabulary construction, enabling the model to capture semantic relationships that emerge when visual patterns align with textual descriptions.

Baselines

We compare CLAIRO against two categories of state-of-the-art models:

Text-only Baseline:

• ActionPiece: Context-aware tokenization using collaborative token merging for sequential recommendation

Multimodal Baseline:

• MQL4GRec: Uses RQ-VAE to discretize multimodal features into separate token sequences

Methodology

System Architecture

CLAIRO extends ActionPiece’s tokenization framework through a streamlined pipeline:

1. Multimodal Feature Extraction

• Visual: CLIP ViT-L/14 extracts image embeddings
• Textual: SentenceT5 generates sentence embeddings
• PCA compression unifies dimensionality (384-dim each → 768-dim fused)

2. Optimized Product Quantization (OPQ)

• FAISS-based quantization decomposes fused embeddings into discrete semantic codes
• Key insight: Skipping final-stage PCA before OPQ yields 37-42% improvement by preserving fine-grained retrieval information

3. Collaborative Token Merging

Unlike traditional approaches that process modalities separately, CLAIRO’s token merging algorithm jointly clusters visual and textual features based on co-occurrence frequency. This enables the vocabulary to capture cross-modal semantic patterns - for example, when album cover aesthetics consistently align with music genre descriptions.

Dataset

We evaluate on Amazon Review Data (2018) across four diverse categories:

• Arts, Crafts and Sewing
• CDs and Vinyl
• Musical Instruments
• Sports and Outdoors

The dataset provides 233.1M reviews with rich multimodal information (product images, descriptions, user interaction histories).

Results

Performance Highlights

CLAIRO achieves significant improvements over both text-only and multimodal baselines across multiple datasets:

View Full Results with detailed metrics (Recall@5/10, NDCG@5/10) and statistical analysis.

Why Performance Varies Across Categories

Our analysis reveals that visual features contribute differently depending on product type:

🎵 CDs and Vinyl (Highest gains: +135.3% over MQL4GRec)

• Album covers contain semantically rich visual information (artistic style, genre cues, emotional tone)
• Strong visual-textual correlation enables effective co-occurrence learning
• Visual patterns provide highly complementary signals to text descriptions

⚽ Sports and Outdoors (Mixed results)

• Complex contextual information (athletes, scenes, usage environments)
• Rich visual diversity helps vs text-only baseline (+47.2%)
• But misalignment with text descriptions harms vs multimodal baseline (-22.5%)
• Highlights the importance of proper multimodal alignment

🎨 Arts and 🎸 Instruments (Marginal gains: ~2-3%)

• Visually heterogeneous products (raw materials, tools, plain backgrounds)
• Visual features play complementary rather than dominant role
• Demonstrates that adding visual data alone isn’t sufficient

Key Technical Insights

1. Skip Final-stage PCA → +37-42% improvement

• PCA’s global dimensionality reduction discards fine-grained retrieval information
• OPQ already optimizes for quantization error; additional PCA is detrimental

2. Visual Features are Complementary, Not Dominant

• Text-only variant performs similarly to full model
• Visual-only variant fails due to limited embedding diversity
• Best results come from proper alignment of both modalities

Main Contributions

1. Cross-modal Token Merging: First work to extend ActionPiece’s collaborative tokenization to jointly learn from visual and textual features, capturing co-occurrence patterns across modalities

2. Efficient Fusion Strategy: Discovered that skipping final-stage PCA before OPQ improves performance by 37-42% while maintaining computational efficiency

3. Category-specific Analysis: Provided comprehensive empirical evidence showing how visual feature effectiveness varies by product type, with insights on when multimodal integration is most beneficial

4. State-of-the-art Results: Achieved up to 135.3% improvement over existing multimodal baseline (MQL4GRec) on semantically rich visual categories

Future Directions

1. Additional Modalities

• Incorporate video and audio features (especially promising for music and video game recommendations)
• Leverage temporal dynamics and acoustic patterns to enrich item representations

2. Adaptive Modality Weighting

• Dynamically adjust visual/textual contribution based on category characteristics
• Emphasize informative modalities while suppressing noisy ones

3. Enhanced Encoders

• Explore more powerful visual encoders beyond CLIP ViT-L/14
• Investigate domain-specific fine-tuning to improve visual embedding discriminability

4. Cross-domain Generalization

• Evaluate on diverse datasets: MovieLens (films), Steam (games), Yelp (restaurants)
• Study how visual semantic richness affects cross-modal learning across different domains

5. Dynamic Vocabulary Expansion

• Enable incremental learning of new token patterns without full retraining
• Adapt to evolving user behaviors and item attributes over time

Technical Skills

• Deep Learning: PyTorch implementation, transformer architectures
• Multimodal Learning: Vision-language fusion, CLIP, ViT, SentenceT5
• Recommendation Systems: Collaborative filtering, sequential recommendation, generative retrieval
• Quantization: Product quantization (OPQ), vector quantization, FAISS
• Data Processing: Large-scale dataset handling (233M reviews), feature extraction pipeline
• Research: Baseline reproduction, ablation studies, performance analysis

📄 Resources:

• Full Technical Report
• ActionPiece Paper
• MQL4GRec Paper