Explain What You Mean: Intent Augmented Knowledge Graph Recommender Built With LLM

📝 Paper Summary

Knowledge Graph (KG) Recommendation LLM-Augmented Recommendation Cold-Start Recommendation

IKGR constructs an intent-centric knowledge graph by using a RAG-guided LLM to extract explicit intent nodes, densifying connections between users and long-tail items without requiring cross-graph fusion.

Core Problem

Existing methods either rely on coarse category-level commonsense that misses fine-grained user intents or use latent vector representations that are non-auditable and bounded by sparse base graphs.

Why it matters:

Enterprise search systems struggle with heterogeneous vocabularies and domain jargon, making standard retrieval ineffective for long-tail content
Cross-graph fusion methods (aligning LLM graphs with interaction graphs) introduce structural noise and entity resolution errors that are hard to debug
Pure LLM-as-recommender approaches introduce high inference latency and often hallucinate or drift, making them unstable for production service-level agreements

Concrete Example: Two users clicking the same 'camera' page may have different intents: 'low-light astrophotography' vs. 'lightweight travel kit'. Category-level edges fail to capture this, while latent methods struggle to connect these users to relevant niche items if the interaction graph is sparse.

Key Novelty

Intent-Centric Knowledge Graph Recommender (IKGR)

Treats 'intent' as a first-class explicit node in the graph, extracted via a tuning-free RAG-guided LLM pipeline that normalizes noisy text into canonical entities
Densifies the graph through a 'mutual-intent' strategy that connects users and items to shared intent nodes, creating paths for information flow even when direct user-item interactions are missing

Architecture

The IKGR framework: Intent Extraction, Graph Construction, and Inference.

Evaluation Highlights

+22% MRR improvement (0.0125 to 0.0153) on the Enterprise Search dataset by adding intent edges
Consistently outperforms strong baselines (KGIN, LLMRec) across public datasets (Books, Beauty, Steam, Yelp), particularly on cold-start slices
Achieves state-of-the-art performance on long-tail edge sets (users/items with ≤3 interactions) in the Books dataset

Breakthrough Assessment

7/10

Solid practical contribution for enterprise/cold-start scenarios. While the GNN component is standard, the explicit grounding of intent via RAG to solve graph sparsity without cross-graph fusion is a smart architectural choice.

⚙️ Technical Details

Problem Definition

Setting: Top-N recommendation using implicit feedback on a user-item bipartite graph augmented with knowledge graph entities

Inputs: User set U, Item set I, Interaction graph G, Item descriptions, User profiles (optional)

Outputs: Ranked list of items for a target user

Pipeline Flow

Input Processing: Knowledge Base Construction → Item/User Text Expansion
Intent Extraction: Round 1 (Exact Intent) → Round 2 (Related Intent via RAG)
Graph Construction: Merge Intent Nodes → Densify User-Intent-Item Graph
Inference: Intent-Prior GNN → Recommendation Scores

System Modules

Knowledge Base (KB) Constructor

Curate key-value pairs of domain-specific abbreviations and explanations to ground LLM generation

Model or implementation: Lookup Table / Simple Retrieval

Intent Extractor (Round 1) (Intent Extraction)

Extract explicit intent entities directly mentioned in item descriptions or user profiles

Model or implementation: Llama-3.1-8B

Intent Densifier (Round 2) (Intent Extraction)

Connect users/items to additional 'related' intent nodes from the existing pool to improve connectivity

Model or implementation: Llama-3.1-8B with RAG

Intent-Prior GNN

Learn node embeddings and predict user-item affinity using the densified intent graph

Model or implementation: Translation-based GNN (similar to KTUP)

Novel Architectural Elements

Two-round intent extraction topology: Round 1 creates the node pool (Exact), Round 2 densifies connections (Related) using the Round 1 pool as a retrieval target
Mutual-intent connectivity densification: Structure that explicitly links users and items via shared 'intent' nodes without fusing separate graphs

Modeling

Base Model: Llama-3.1-8B (for extraction), Translation-based GNN (for ranking)

Training Method: In-context learning for extraction; Supervised learning for GNN

Objective Functions:

Purpose: Maximize link prediction accuracy on the augmented graph.

Formally: maximize sum over dev set of log p(e_ij | G_train, Theta)

Training Data:

8:1:1 split for train/dev/test edges
Datasets: Enterprise Search, Amazon Books, Amazon Beauty, Steam, Yelp

Key Hyperparameters:

retrieval_k: 100
gnn_layers: Not explicitly reported in the paper
text_encoder: all-mpnet-base-v2

Compute: LLM inference is offline/batch; Online component is lightweight GNN (low latency)

Comparison to Prior Work

vs. KGIN: Explicit, readable intent nodes vs. latent/vector-based intent mixtures
vs. CSRec: Densifies a single intent-centric graph vs. fusing separate commonsense and interaction graphs (avoids alignment errors)
vs. LLMRec: Structural augmentation via edges vs. synthetic interaction generation (avoids popularity bias)
+ 1 more
vs. UDITSR [not cited in paper]: IKGR uses RAG to ground intents in external knowledge, whereas UDITSR relies primarily on query-translation mechanisms

Limitations

Reliance on high-quality item descriptions; performance drops when descriptions are poor or absent
Requires pre-computation of intent graph, making real-time updates for strictly new items (without descriptions) challenging
Enterprise results depend on a proprietary dataset, limiting full reproducibility of that specific claim

Reproducibility

Code availability is not provided. Datasets (Amazon, Steam, Yelp) are public; Enterprise Search dataset is proprietary. Prompt templates are provided in Appendix A (referenced in text). Llama-3.1-8B and all-mpnet-base-v2 are open models.

📊 Experiments & Results

Evaluation Setup

Top-N recommendation ranking task using implicit feedback

Benchmarks:

Enterprise Search (Dataset Search / Re-ranking) [New]
Amazon Books (Product Recommendation)
Amazon Beauty (Product Recommendation)
Steam (Game Recommendation)
Yelp (Business Recommendation)

Metrics:

Hit Ratio (HR@1, 5, 10)
NDCG@1, 5, 10
Mean Reciprocal Rank (MRR)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Ablation study on the Search dataset demonstrating the impact of each IKGR component.
Enterprise Search	MRR	0.0125	0.0153	+0.0028
Enterprise Search	NDCG@10	0.0242	0.0278	+0.0036
Cold-start performance on Amazon Books (tail edge set where nodes have <= 3 interactions).
Amazon Books (Tail Slice)	NDCG@10	0.0385	0.0432	+0.0047
Amazon Books (Tail Slice)	NDCG@10	0.0401	0.0432	+0.0031

Main Takeaways

Explicit intent modeling consistently improves performance over implicit/latent methods (KGIN) and pure LLM augmentation (LLMRec), especially in sparse scenarios.
The 'related intent' densification step is critical; removing it significantly drops performance, proving that simply extracting exact entities is insufficient for connectivity.
Hyperparameter sensitivity analysis shows performance saturates at k=100 for retrieved intents, balancing context length and recall.
Performance gains are most pronounced on the Search dataset which has the largest knowledge gap, suggesting the RAG-based grounding is particularly effective for domain-specific tasks.

📚 Prerequisite Knowledge

Prerequisites

Knowledge Graph Convolutional Networks (KGCN)
Retrieval-Augmented Generation (RAG)
Graph Neural Networks (GNN)
Recommender Systems (Collaborative Filtering)

Key Terms

RAG: Retrieval-Augmented Generation—AI systems that answer questions or generate content by first searching for relevant documents

LLM: Large Language Model—a deep learning algorithm that can recognize, summarize, translate, predict, and generate text

KG: Knowledge Graph—a structured representation of data using nodes (entities) and edges (relationships)

GNN: Graph Neural Network—a class of neural networks designed to perform inference on data described by graphs

Cold-Start: The problem of recommending items to users who have little to no interaction history, or recommending new items with no consumption history

MRR: Mean Reciprocal Rank—a statistic measure for evaluating any process that produces a list of possible responses to a sample of queries

NDCG: Normalized Discounted Cumulative Gain—a measure of ranking quality that takes into account the position of relevant items in the list

HR: Hit Ratio—the fraction of users for which the correct answer is included in the recommendation list (equivalent to Recall@k for single ground truth)

NER: Named Entity Recognition—a subtask of information extraction that seeks to locate and classify named entities mentioned in unstructured text

KGIN: Knowledge Graph-based Intent Network—a baseline method that models intent as latent mixtures over relations

TransH: A knowledge graph embedding model that translates entities in a hyperplane specific to the relation

BM25: Best Matching 25—a ranking function used by search engines to estimate the relevance of documents to a given search query

kNN: k-Nearest Neighbors—an algorithm used here to retrieve the top-k most similar intent candidates