Enriching Semantic Profiles into Knowledge Graph for Recommender Systems Using Large Language Models

📝 Paper Summary

Knowledge-aware recommendation LLM-enhanced recommendation

SPiKE improves recommendation by generating semantic profiles for all entities using LLMs and propagating them through a Knowledge Graph to capture both attribute-level and structural preferences.

Core Problem

Acquiring rich personal data for recommendation is difficult due to privacy and sparsity, while current profiling methods either lack structural context (LLM-only) or semantic depth (KG-only).

Why it matters:

Profiles solely relying on KGs miss rich semantic details available in textual metadata (descriptions, reviews).
LLM-only profiles are often limited to local context and cannot effectively propagate influence to distant, sparsely connected items.
Existing hybrid approaches often confine profile signals to local user-item modeling rather than propagating them throughout the entire graph.

Concrete Example: A KG-only model links a user to a movie but misses why they liked it (e.g., 'dark humor'). An LLM might infer 'dark humor' preference but fails to link it to other movies with that attribute if they aren't directly adjacent in the text prompt context.

Key Novelty

Semantic Profiles into Knowledge Graphs (SPiKE)

Generates text profiles for *all* entities (items, users, and auxiliary nodes like genres/directors) using LLMs to capture specific attribute preferences.
Injects these text profiles into the Knowledge Graph as additive signals during aggregation, allowing semantic nuances to propagate structurally.
Uses a ResNet-like mechanism where profiles are added during message passing but removed afterwards to preserve the original KG embedding space.

Architecture

Overview of the SPiKE framework, detailing the three main components: Entity Profile Generation, Profile-aware KG Aggregation, and Preference Matching.

Evaluation Highlights

Consistently outperforms state-of-the-art KG-based (e.g., KGIN, KGAT) and LLM-based recommenders across three real-world benchmarks.
Demonstrates effective handling of sparse data by leveraging auxiliary entity profiles to bridge gaps between users and items.
Ablation studies confirm the necessity of profiling all entity types (users, items, auxiliary), not just users or items alone.

Breakthrough Assessment

7/10

Solid hybrid approach addressing the specific limitations of LLM vs. KG profiling. While the components (LLM summarization, KG aggregation) are established, the 'profiling all entities' + 'inject-then-remove' propagation strategy is a clever integration.

⚙️ Technical Details

Problem Definition

Setting: PKG-aware recommendation (Profile-Enriched Knowledge Graph)

Inputs: User u, Item v, Knowledge Graph G (entities E, relations R), textual metadata (descriptions, reviews)

Outputs: Predicted preference score y_uv in [0, 1]

Pipeline Flow

Group 1: Entity Profile Generation (LLM-based)
Group 2: Profile-aware KG Aggregation (GNN-based)
Group 3: Preference Matching (Training Objective)

System Modules

Item Profiler (Entity Profile Generation)

Generate text summaries for items using descriptions, reviews, and KG neighbors

Model or implementation: LLM (Specific model not detailed in snippet, typically frozen)

Auxiliary Profiler (Entity Profile Generation)

Generate summaries for non-user/non-item entities (e.g., genres) based on related items

Model or implementation: LLM

User Profiler (Entity Profile Generation)

Generate user preference summaries based on interaction history

Model or implementation: LLM

Text Encoder (Entity Profile Generation)

Convert text profiles into vector embeddings

Model or implementation: Pre-trained text encoder (e.g., BERT/Sentence-BERT, specific model not detailed)

Profile-Injector (Profile-aware KG Aggregation)

Project and add profile embeddings to node embeddings before aggregation

Model or implementation: 2-layer MLP (M)

KG Aggregator (Profile-aware KG Aggregation)

Propagate information across the graph using attention mechanism

Model or implementation: Graph Attention Network component

Novel Architectural Elements

Bottom-up profiling pipeline: Item -> Auxiliary -> User (dependency chain for prompt construction)
Inject-then-Remove aggregation: Semantic profiles are added to embeddings specifically for the message-passing step to influence neighbor weights, then removed to keep the final representation in the structural KG space

Modeling

Base Model: LLM for profiling (frozen), GNN for recommendation (trainable)

Training Method: End-to-end training of the GNN components (LLM is used offline/frozen for profiling)

Objective Functions:

Purpose: Encourage similar profiles to be embedded close together.

Formally: Pairwise profile preference matching (implied contrastive or matching loss, exact formula cut off in text)

Adaptation: None (LLM is used as a fixed profiler)

Trainable Parameters: GNN weights, Attention parameters W1/W2, Profile projection MLP M, Entity/Relation embeddings

Training Data:

KG triplets (h, r, t)
User-Item interactions
Textual metadata (reviews, descriptions) for prompt construction

Key Hyperparameters:

scaling_factor_lambda_p: Controls influence of profile embedding

Compute: Not reported in the paper

Comparison to Prior Work

vs. KG-only (KGAT, KGIN): SPiKE injects rich semantic text profiles (generated by LLMs) into the aggregation, capturing attribute-level preferences missing in pure structural models.
vs. LLM-only (ChatRec): SPiKE uses the KG to propagate profiles, extending influence beyond the LLM's local context window and avoiding scalability issues of direct LLM inference.
vs. Hybrid (CoLLM) [not cited in paper]: Unlike methods that distill LLM knowledge into user/item embeddings only, SPiKE profiles *all* entities (including auxiliary ones like genres) and actively uses them in the message-passing calculation.

Limitations

Dependency on the quality of the LLM: if the LLM hallucinates in profiles, noise propagates through the KG.
Computationally expensive offline profiling step: generating text profiles for every entity in a large KG is costly.
Snippet does not detail how the 'removal' of profiles after aggregation affects non-linearities in deep GNNs.

Reproducibility

Code: https://doi.org/10.5281/zenodo.18041669

Code is publicly available at https://doi.org/10.5281/zenodo.18041669. The paper snippet does not specify exact LLM or Text Encoder models used (likely detailed in full experimental section not provided).

📊 Experiments & Results

Evaluation Setup

Top-K Recommendation on real-world datasets

Benchmarks:

Not explicitly named in snippet (Recommendation)

Metrics:

Recall
NDCG
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

SPiKE consistently outperforms state-of-the-art methods (KG-based, LLM-based, and hybrid) across three real-world benchmarks.
The 'Profile-aware KG aggregation' mechanism effectively propagates semantic signals to distant nodes, addressing the limited impact range of LLM-only methods.
Profiling auxiliary entities (not just users/items) is crucial for aligning item characteristics with user tastes at a granular level.

📚 Prerequisite Knowledge

Prerequisites

Knowledge Graph (KG) structure (triplets, entities, relations)
Graph Neural Networks (GNNs) / Message Passing
Large Language Models (LLMs) for text summarization
Contrastive Learning (for embedding alignment)

Key Terms

PKG: Profile-Enriched Knowledge Graph—an extension of a standard KG where each entity is augmented with a semantic text profile generated by an LLM.

Auxiliary Entities: Nodes in a KG that are neither users nor target items, but provide context (e.g., 'Fantasy' genre node, 'Steven Spielberg' director node).

1-hop / 2-hop neighborhood: Direct connections (1-hop) and connections-of-connections (2-hop) in a graph, used here to gather context for LLM profiling.

Message Passing: A GNN mechanism where nodes update their representations by aggregating information from their neighbors.

ResNet-like coupling: A technique where a signal (here, the profile) is added to a representation for an operation and then subtracted or removed to maintain the original latent space.