Leveraging Language Semantics for Collaborative Filtering with TextGCN and TextGCN-MLP: Zero-Shot vs In-Domain Performance

TL;DR

The paper tackles leveraging textual semantics from Large Language Models to improve collaborative filtering, contrasting a parameter-free TextGCN that diffuses LLM-derived item embeddings on a user-item graph with TextGCN-MLP, which adds a trainable MLP and contrastive learning for in-domain refinement. TextGCN delivers state-of-the-art zero-shot performance by exploiting language semantics and graph diffusion, whereas TextGCN-MLP achieves superior in-domain accuracy at the cost of reduced cross-domain generalization. Key contributions include the introduction of a k-positive contrastive loss and a two-tower architecture to adapt language-informed representations to downstream data, along with comprehensive ablations and efficiency analyses. The findings highlight a fundamental generalization-specialization trade-off in foundation-model-inspired recommender systems and offer a scalable, training-free alternative that excels in zero-shot settings.

Abstract

In recent years, various approaches have been proposed to leverage large language models (LLMs) for incorporating textual information about items into recommender systems. Existing methods primarily focus on either fine-tuning LLMs to generate recommendations or integrating LLM-based embeddings into downstream models. In this work, we follow the latter direction and propose \textbf{TextGCN}, which applies parameter-free graph convolution layers directly over LLM-based item-title embeddings, instead of learning ID-based embeddings as in traditional methods. By combining language semantics with graph message passing, this architecture achieves state-of-the-art zero-shot performance, significantly outperforming prior approaches. Furthermore, we introduce \textbf{TextGCN-MLP}, which extends TextGCN with a trainable multilayer perceptron trained using a contrastive loss, achieving state-of-the-art in-domain performance on recommendation benchmarks. However, the zero-shot performance of TextGCN-MLP remains lower than that of TextGCN, highlighting the trade-off between in-domain specialization and zero-shot generalization. We release our code on github at \href{https://github.com/ChernovAndrey/TFCE}{github.com/ChernovAndrey/TFCE}.

Figures (10)

• Figure 1: Proposed pipeline for obtaining training-free collaborative embeddings (TextGCN) in the language space. Modules with snowflake are frozen (no trainable parameters).
• Figure 2: TextGCN-MLP training head using k positive contrastive loss (KCL). Modules with flame allow gradient flow and parameter updates.
• Figure 3: Comparison of TextGCN-MLP with AlphaRec architecture
• Figure 4: TextGCN-MLP partial dependence on the number of GCLs.
• Figure 5: TextGCN-MLP partial dependence on the MLPs output size.