An Automatic Graph Construction Framework based on Large Language Models for Recommendation

📝 Paper Summary

Graph Construction for Recommendation LLM-enhanced Recommender Systems

AutoGraph constructs recommendation graphs by using LLMs to generate semantic vectors and vector quantization to extract latent factor nodes, enabling global-view learning with linear complexity.

Core Problem

Existing LLM-based graph construction methods rely on pairwise comparisons, which lack a global view due to context window limits and are computationally inefficient ($O(N^2)$).

Why it matters:

Graph structure quality fundamentally limits GNN performance; poor graphs lead to suboptimal recommendations.
Pairwise assessment is prohibitively expensive for industrial systems with millions of users/items.
Local-view comparisons overlook global context (e.g., feature frequency), leading to inaccurate similarity assessments.

Concrete Example: Consider items $i_1, i_2, i_3$. Locally, $i_1$ and $i_2$ might seem most similar due to many shared features. However, if $i_1$ and $i_3$ share a globally rare/informative feature $f_8$, they should be strongly linked. Pairwise LLM methods miss this global distribution context.

Key Novelty

Quantization-based Graph Construction

Uses Vector Quantization (VQ) to map LLM-enriched user/item vectors to discrete latent codes.
Treats these latent codes as new graph nodes, effectively creating 'hubs' that link users/items based on global semantic patterns rather than just local pairwise similarity.
Decouples semantic inference (Pointwise LLM call) from structure learning, reducing complexity from quadratic to linear.

Architecture

The complete AutoGraph framework, illustrating the two-stage process: Quantization-based Graph Construction and Graph-enhanced Recommendation.

Evaluation Highlights

+2.69% improvement on RPM (Revenue Per Mille) in online A/B testing on Huawei's advertising platform.
+7.31% improvement on eCPM (effective Cost Per Mille) in online A/B testing.
Deployed as the main traffic model serving hundreds of millions of users.

Breakthrough Assessment

8/10

Significantly improves efficiency of LLM-based graph construction (linear vs quadratic) while adding a mechanism for global context, validated by large-scale industrial deployment.

⚙️ Technical Details

Problem Definition

Setting: Graph-enhanced Recommendation

Inputs: User set $\mathcal{U}$, Item set $\mathcal{I}$, interaction sequences $\mathcal{S}^u$, user/item features $\mathcal{F}$

Outputs: Predicted preference score $\hat{y}_{ui}$

Pipeline Flow

Semantic Vector Generation: LLM → Semantic Vectors
Latent Factor Extraction: Vectors → Residual Quantization → Discrete Codes
Graph Construction: Nodes + Codes → Global Graph
Recommendation: Graph → Metapath GAT → Enhanced Embeddings

System Modules

Semantic Vector Generator (Quantization-based Graph Construction)

Infers user preference and item knowledge from profiles

Model or implementation: Large Language Model (specific architecture not named in text)

Latent Factor Extractor (Quantization-based Graph Construction)

Extracts global latent factors via quantization

Model or implementation: Residual Quantization (RQ) Encoder-Decoder

Graph Constructor (Quantization-based Graph Construction)

Builds heterogeneous graph with latent factor nodes

Model or implementation: Deterministic Graph Builder

Metapath Aggregator

Aggregates semantic and collaborative info

Model or implementation: Graph Attention Network (GAT)

Novel Architectural Elements

Latent Factor Nodes: Treating quantization codes as explicit graph nodes to bridge users/items.
Quantization-based Graph Construction: A pipeline replacing pairwise similarity checks with shared codebook membership.

Modeling

Base Model: Model-agnostic framework (compatible with various backbones)

Training Method: Joint optimization of Quantization and Recommendation

Objective Functions:

Purpose: Reconstruct semantic vectors and cluster residuals.

Formally: $\mathcal{L}_{rq} = \mathcal{L}_{rec} + \beta \mathcal{L}_{com}$ (Reconstruction + Commitment loss).
Purpose: Optimize recommendation task (e.g., CTR).

Formally: Standard recommendation loss (e.g., Log Loss) using graph-enhanced embeddings.

Training Data:

Yelp, Amazon-Book, Movielens-1M (mentioned as datasets used)

Compute: Not reported in the paper

Comparison to Prior Work

vs. Pairwise LLM methods: AutoGraph is $O(N)$ vs $O(N^2)$ and captures global view via quantization.
vs. Rule-based: AutoGraph captures in-depth semantics via LLMs rather than simple interactions.

Limitations

Dependency on the quality of LLM-generated semantic vectors.
Complexity of managing heterogeneous graph structures with added latent nodes.
Requires training distinct quantization parameters for users and items.

Reproducibility

No code provided. Prompts for LLM semantic inference are mentioned to be in Appendix C (not included in input). Specific LLM model and hyperparameters not detailed in the snippet.

📊 Experiments & Results

Evaluation Setup

Online A/B testing on industrial platform; Offline evaluation on public datasets.

Benchmarks:

Huawei Advertising Platform (Online CTR/Revenue Prediction)
Yelp (Recommendation)
Amazon-Book (Recommendation)
Movielens-1M (Recommendation)

Metrics:

RPM (Revenue Per Mille)
eCPM (effective Cost Per Mille)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Online A/B test results on Huawei advertising platform show significant revenue improvements.
Huawei Advertising Platform	RPM	Not reported in the paper	Not reported in the paper	+2.69%
Huawei Advertising Platform	eCPM	Not reported in the paper	Not reported in the paper	+7.31%

Experiment Figures

Comparison of local vs. global view in graph construction.

Main Takeaways

AutoGraph significantly boosts revenue metrics (RPM, eCPM) in a large-scale industrial setting, validating its effectiveness.
The framework effectively combines LLM semantics with collaborative signals via the constructed graph.
The linear complexity $O(N)$ makes it feasible for deployment to hundreds of millions of users.

📚 Prerequisite Knowledge

Prerequisites

Graph Neural Networks (GNNs)
Vector Quantization (VQ)
Large Language Models (LLMs) for semantic embedding

Key Terms

RPM: Revenue Per Mille—estimated revenue per 1,000 impressions.

eCPM: effective Cost Per Mille—a measure of ad revenue generated per 1,000 impressions.

Vector Quantization: A technique to map continuous vectors to a finite set of discrete codebook vectors (latent factors).

Metapath: A sequence of node types and edge types (e.g., User-Item-User) used to define semantic relationships in a heterogeneous graph.

Residual Quantization: A multi-stage quantization method where residuals from one level are quantized by the next, refining the representation.

GAT: Graph Attention Network—a neural network architecture that computes node representations by attending to neighbors.