TL;DR
This paper introduces EasyRec, a novel recommendation system that leverages language models (LMs) to create text-based user and item profiles. By integrating collaborative filtering signals into LM training, EasyRec achieves state-of-the-art performance in zero-shot recommendation scenarios and offers efficient adaptation to evolving user preferences.
What Problem Does This Paper Address?
Traditional recommender systems heavily rely on unique user and item IDs. While effective, this ID-based approach suffers from a significant limitation: poor generalization to new users or items, especially in "zero-shot" scenarios where there's no overlap between training and testing data. This makes it challenging for systems to adapt to new domains or handle situations with sparse interaction data. Existing methods that try to leverage language models often use general-purpose LMs or are computationally expensive, hindering practical deployment. The authors aim to bridge this gap by developing a recommendation system that is both effective in zero-shot settings and computationally efficient, by specifically tailoring LMs for recommendation tasks using text profiles and collaborative signals.
Methods
EasyRec's core innovation lies in its collaborative language modeling approach. Instead of relying solely on IDs, it generates rich textual profiles for users and items.
- Collaborative Profiling: For items, profiles are generated using LMs by combining their titles, categories, and descriptions, and optionally user reviews. For users, profiles are created by leveraging their interaction history with items, using their profiles and reviews to capture collaborative signals. This process is inspired by Chain-of-Thought prompting to improve interpretability and reliability.
- Profile Embedder with Collaborative LM: A bidirectional Transformer encoder (specifically RoBERTa) is used to encode these textual profiles into embeddings. This encoder is trained not just on text but also incorporates contrastive learning with collaborative filtering signals. This means the model learns to align text-based semantic representations with user behavior patterns.
- Contrastive Collaborative LM Learning: The training objective combines a contrastive loss (to pull similar user-item embeddings together and push dissimilar ones apart) with a masked language modeling (MLM) loss to stabilize training and enhance generalization.
- Profile Diversification: To further improve generalization, especially for limited datasets, EasyRec employs LLMs to diversify user and item profiles. This involves rephrasing existing profiles multiple times to create semantically similar but distinct versions, effectively augmenting the training data.
Datasets
The authors evaluated EasyRec on several datasets:
- Amazon Review Data: Used for training, covering categories like Arts, Crafts and Sewing, Movies and TV, Video Games, Home and Kitchen, Electronics, and Tools.
- Test Datasets:
- Sports and Outdoors (Sports): From Amazon.
- Cross-platform datasets: Steam and Yelp. These datasets were processed to include interactions, ratings (≥ 3), and a 10-core filtering for sparsity.
Key Insights and Ideas
- Textual profiles are powerful: By moving beyond IDs and leveraging rich textual descriptions of users and items, recommender systems can achieve better generalization.
- Tailoring LMs for recommendation is key: Using general-purpose LMs as text encoders is less effective than training LMs specifically for recommendation tasks, incorporating collaborative signals.
- Contrastive learning bridges semantics and behavior: Aligning text-based semantic embeddings with collaborative filtering signals through contrastive learning is crucial for capturing nuanced user preferences.
- Profile diversification boosts performance: Augmenting training data with diversified text profiles significantly improves model performance, especially for larger models and in scenarios with limited data.
- EasyRec is adaptable: The system can efficiently adapt to shifts in user preferences by simply updating user profiles, without requiring retraining.
Key Results
- State-of-the-art in Zero-Shot Recommendation: EasyRec consistently outperformed various state-of-the-art language models (BERT, BART, SimCSE, GTR, BGE, BLaIR) in text-based zero-shot recommendation across Sports, Steam, and Yelp datasets. For instance, on the Steam dataset, EasyRec achieved a Recall@20 of 0.0225, significantly higher than other models.
- Effective Text-enhanced CF: When integrated with backbone CF models like LightGCN, EasyRec improved performance, demonstrating the value of incorporating text modality. On the Steam dataset, EasyRec with LightGCN achieved a Recall@20 of 0.0888, outperforming other text-enhanced methods.
- Scalability: EasyRec exhibits a scaling law, meaning its performance improves with increased model size (from EasyRec-Small to EasyRec-Large).
- Efficiency: EasyRec achieves a computational cost of approximately 0.01 seconds per prediction, making it practical for large-scale recommendation tasks, unlike some other LLM-based methods that take up to a second per prediction.
- Profile Diversification Impact: Profile diversification led to a notable increase in recommendation diversity, as measured by Intra-List Similarity (ILS). For example, on the Steam dataset, diversification increased ILS@20 from 0.7573 to 0.7727 for EasyRec-Large.
Limitations and Future Work
- Data Modality Diversity: The current EasyRec primarily relies on textual data. The authors acknowledge that items can have richer modalities like images and videos, which could convey aesthetic preferences and cultural trends that textual data might miss. Future work could explore integrating multimodal data processing techniques to enhance personalization and predictive accuracy.
- General-purpose LLMs vs. Recommendation-tailored LMs: While EasyRec demonstrates the benefit of tailored LMs, the paper notes that many existing approaches still rely on general-purpose encoders. Future research could further explore the development and application of LMs specifically designed for recommendation tasks.
Summary updated