Augmenting the User-Item Graph with Textual Similarity Models

Federico López, Martin Scholz, Jessica Yung, Marie Pellat, Michael Strube, Lucas DixonarXiv 2021 · arXiv

Introduction

Recommender systems (RS) are fundamental to modern digital platforms, yet they consistently struggle with the cold-start problem. When new users or items enter a system without historical interaction data, traditional collaborative filtering methods fail. While textual metadata such as product descriptions and reviews are often available, standard deep learning models frequently fail to extract significant performance gains from this unstructured data.

Researchers from the Heidelberg Institute for Theoretical Studies and Google address this in their paper, "Augmenting the User-Item Graph with Textual Similarity Models." They propose an unsupervised framework that transforms textual information into explicit structural edges within a knowledge graph (KG), providing a stronger inductive bias for recommendation models.

Methodology: From Text to Structure

The proposed pipeline moves beyond treating text as a simple feature vector, instead using it to define the "map" of the item space.

  1. Mining Semantic Relations: The authors utilize the Universal Sentence Encoder (USE) to generate high-dimensional embeddings for items based on metadata and reviews. By calculating cosine similarity between all item pairs, they identify the top-$k$ most similar items.
  2. Graph Augmentation: These semantic links are injected into the standard user-item interaction graph as a new relation type: item_A has_semantic_similarity item_B. This process transforms a sparse bipartite graph into a dense, multi-relational Knowledge Graph.
  3. Geometric Embedding: The augmented graph is evaluated using various KG embedding models.
Visualization of the augmented knowledge graph structure.

The Hyperbolic Advantage

A critical contribution of this work is the analysis of graph curvature. Using Ollivier-Ricci curvature analysis, the researchers found that as semantic relations are added, the graph increasingly exhibits a "tree-like" or hierarchical structure. Such structures are mathematically better suited for Hyperbolic space (e.g., the Poincaré ball) than flat Euclidean space.

Curvature analysis of the augmented graph.

By embedding the graph in hyperbolic space, models like MuRP can capture latent hierarchies more efficiently, achieving superior performance with significantly fewer dimensions than Euclidean counterparts. The prevalence of negative curvature (indicated by red in the visualizations) confirms the suitability of non-Euclidean embedding spaces.

Experimental Results

The framework was tested on three subsets of the Amazon Product Dataset: Musical Instruments, Video Games, and Arts, Crafts and Sewing.

  • Accuracy Boost: The structural augmentation led to improvements in Hit Rate (HR@10) of up to 32.6% for certain models (e.g., RotatE on Musical Instruments).
  • Cold-Start Mitigation: In scenarios involving users with minimal interaction history, semantic relations were the most impactful addition, yielding a 15.4% performance increase in the Musical Instruments category.
  • Metadata Efficiency: Interestingly, using concise product metadata (names and descriptions) often proved as effective as using the full text of thousands of reviews, suggesting that high-quality, condensed text provides the cleanest semantic signals.

Conclusion

This research demonstrates that explicitly defining relationships through textual similarity is a powerful way to regularize recommender systems. By shifting the focus from "learning features" to "augmenting structure," and leveraging the natural hyperbolic geometry of these augmented graphs, developers can build more robust systems that perform well even in data-sparse environments.

Summary updated

BibTeX
@misc{lopez2021augmenting,
  author = {Federico López and Martin Scholz and Jessica Yung and Marie Pellat and Michael Strube and Lucas Dixon},
  title = {{Augmenting the User-Item Graph with Textual Similarity Models}},
  year = {2021},
  eprint = {2109.09358},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.09358},
}
Figure