Introduction
In the landscape of modern recommender systems, "sparsity" remains a primary challenge. When users have minimal interaction history or items belong to the "long-tail" (niche products with few purchases), traditional collaborative filtering algorithms often fail. While Large Language Models (LLMs) have emerged as a potential solution due to their vast semantic knowledge, they are frequently too computationally expensive for real-time inference or lack the domain-specific nuance required for enterprise applications.
Researchers at Capital One have introduced IKGR (Intent Knowledge Graph Recommender), a framework that leverages a tuning-free LLM to transform abstract user needs into explicit "intent nodes" within a knowledge graph. By elevating the why behind a user's action to a first-class data entity, IKGR bridges the gap between sparse interactions and rich semantic meaning.
The Challenge: Sparsity and the Knowledge Gap
Recommender systems typically struggle with two specific hurdles:
- Sparsity: In enterprise environments, users and items often have very few direct connections. For instance, a developer searching for a specific internal dataset may have no prior history with that data type, leaving the model with no "anchor" for a recommendation.
- The Knowledge Gap: Standard models often fail to interpret domain-specific jargon. While a general LLM might recognize "MLP" as a neural network, in a specific corporate context, it might refer to a "Marketing Launch Plan."
Existing approaches often attempt to solve this by using LLMs to generate synthetic interactions or by relying on broad categories. However, these methods frequently miss the granular nuance of specific user intent.
Methodology: The IKGR Framework
The IKGR framework operates through a three-stage pipeline designed for offline processing to maintain production-level latency:
RAG-Guided Intent Extraction
Rather than allowing an LLM to generate intents in a vacuum, IKGR utilizes Retrieval-Augmented Generation (RAG). The system provides the LLM with a curated knowledge base of domain-specific terms. The LLM then extracts precise "intent entities" from user profiles and item descriptions (e.g., identifying "physics-based gameplay" from a raw game description).
Mutual-Intent Connectivity Densification
To address the sparsity problem, the framework performs a second round of extraction to identify "related intents." This process creates a denser web of connections, effectively shortening the semantic path between a user and potentially relevant long-tail items.
Intent Prior GNN
The final stage employs a lightweight Graph Neural Network (GNN). This model uses an "embedding translation" layer (inspired by TransH) to treat the relationship "user consumes item" as a translation in vector space, guided by shared intent nodes. This ensures that if a user and an item share an underlying intent, they are pulled closer together in the embedding space.
Experimental Results
The researchers evaluated IKGR across five diverse datasets: Enterprise Search (proprietary), Amazon Beauty, Amazon Books, Steam, and Yelp2022.
IKGR consistently outperformed state-of-the-art baselines, including KGIN and LLMRec, across all tested metrics.
Key Performance Insights
- Accuracy Gains: On the Enterprise Search dataset, which is characterized by high jargon and high sparsity, IKGR improved the Mean Reciprocal Rank (MRR) by 22% compared to versions of the model without intent nodes.
- Cold-Start Excellence: The model demonstrated its greatest strength on the "tail edge"—users or items with three or fewer interactions. On the Steam dataset, IKGR achieved a Hit Ratio (HR@10) of 0.3574, significantly outperforming the next best baseline.
- Production Readiness: By keeping the heavy LLM processing offline and utilizing a lightweight GNN for online scoring, the system maintains the low latency required for real-world search and recommendation engines.
Conclusion and Future Work
IKGR demonstrates that making intent explicit—rather than treating it as a latent variable—significantly improves recommendation quality and model interpretability. By grounding LLM extraction in a RAG-based knowledge base, the system overcomes the common pitfalls of hallucination and domain ignorance.
Future research will focus on handling scenarios where textual descriptions are poor or user profiles are empty, as well as adapting the framework to handle rapidly evolving, real-time user intents in dynamic environments.
Summary updated