Knowledge Graph Context-Enhanced Diversified Recommendation

📝 Paper Summary

Recommendation Diversity Knowledge Graph-based Recommendation

KG-Diverse enhances recommendation diversity by pushing user embeddings away from their interaction history using knowledge graph contexts, while maintaining accuracy through item alignment based on shared entities.

Core Problem

Traditional recommender systems prioritize accuracy based on historical interactions, leading to 'echo chambers' where users only see familiar content types, while existing diversity metrics (like category coverage) are too coarse-grained.

Why it matters:

Users trapped in filter bubbles miss novel items of interest, reducing long-term engagement.
Existing diversity metrics fail to distinguish between items in the same broad category (e.g., two different phones vs. two models of the same phone), ignoring rich contextual differences.
Simply optimizing for diversity often hurts accuracy; achieving a better trade-off is a persistent challenge in the field.

Concrete Example: A user watches 'The Avengers'. A standard system recommends 'Captain America 3' (same genre/actors). A category-diverse system might pick a random non-action movie. KG-Diverse recommends 'Dolittle' (comedy, but same actor R. Downey Jr.) or 'Before We Go' (different genre, same actor Chris Evans as director), offering diversity through specific KG connections rather than just broad genres.

Key Novelty

KG-Diverse (Knowledge Graph-based Diversified Recommendation)

Introduces Entity Coverage and Relation Coverage metrics to measure diversity at a granular, semantic level rather than just item categories.
Uses a Diversified Embedding Learning (DEL) module that explicitly pushes user representations away from their past interactions and towards unexplored regions of the embedding space.
Employs 'Conditional Alignment and Uniformity' to ensure items sharing KG entities remain close in vector space, preserving semantic consistency while diversifying.

Architecture

The overall architecture of KG-Diverse, illustrating the flow from KG propagation to user/item representation learning.

Evaluation Highlights

Outperforms baselines on diversity metrics (Entity Coverage and Relation Coverage) while maintaining comparable accuracy (Recall/NDCG) across three datasets (MovieLens-1M, Last.FM, Book-Crossing).
Achieves better trade-off between accuracy and diversity compared to state-of-the-art diversified recommendation methods like DGCN and MMR.
Ablation studies confirm that the Diversified Embedding Learning module is the primary driver for increased diversity scores.

Breakthrough Assessment

7/10

Offers a novel, granular approach to diversity using KGs instead of coarse categories. The proposed metrics and embedding strategy are theoretically sound and empirically effective, though the scope is limited to KG-enhanced scenarios.

⚙️ Technical Details

Problem Definition

Setting: Top-k Recommendation on a User-Item Bipartite Graph augmented with a Knowledge Graph

Inputs: User-item interaction graph G, Knowledge Graph G_K (triplets of head, relation, tail)

Outputs: List of top-k recommended items for each user

Pipeline Flow

KG Propagation: Item/Entity embedding aggregation
Diversified Embedding Learning: Generate diversity-aware user embeddings
Conditional Alignment: Refine item embeddings based on shared entities
Collaborative Propagation: LightGCN layers on User-Item graph
Prediction: Dot product of final user and item embeddings

System Modules

KG Aggregator

Aggregates neighbor information in the KG to form item representations

Model or implementation: GNN-based aggregator with relation modeling (Equation 2)

Diversified Embedding Learning (DEL)

Creates user embeddings that are distinct from the mean of their interacted items

Model or implementation: Distance-based loss + Embedding shift (Equation 5)

Conditional Alignment Module

Aligns embeddings of items that share entities in the KG

Model or implementation: Conditional Alignment Loss (Equation 8) + Uniformity Loss

LightGCN Encoder

Captures collaborative signals from user-item interactions

Model or implementation: LightGCN layers (Equation 10)

Novel Architectural Elements

DEL module: Specifically modifies user embeddings to be distant from the centroid of their history
Conditional Alignment: Incorporates shared KG entities explicitly into the alignment loss for items

Modeling

Base Model: LightGCN backbone augmented with KG propagation

Training Method: End-to-end training with multi-task loss

Objective Functions:

Purpose: Optimize recommendation accuracy (ranking).

Formally: BPR Loss L_BPR = sum ln(sigma(y_ui - y_uj))
Purpose: Enforce diversity in user representations.

Formally: Implicitly handled via the DEL embedding generation process and end-to-end optimization
Purpose: Align similar items based on KG overlaps.

Formally: L_align = sum ||(e_i1 - e_overlap) - (e_i2 - e_overlap)||^2
Purpose: Distribute item embeddings uniformly.

Formally: L_uniformity = log E[e^(-2||x-y||^2)]

Key Hyperparameters:

KG_propagation_layers: 3
LightGCN_layers: 3
embedding_dimension: 64
+ 5 more
batch_size: 2048
learning_rate: 0.0001
lambda_1 (alignment weight): 0.01
lambda_2 (uniformity weight): 0.01
lambda_3 (regularization): 1e-5

Compute: Not reported in the paper

Comparison to Prior Work

vs. KGIN: KG-Diverse explicitly pushes user embeddings away from history for diversity, whereas KGIN focuses on intent modeling.
vs. MMR: KG-Diverse learns diverse representations during training (pre-processing/in-model), whereas MMR is a post-processing re-ranking step.
vs. KGAT: KG-Diverse uses conditional alignment for items sharing entities, whereas KGAT uses attention mechanisms for aggregation.

Limitations

Reliance on high-quality Knowledge Graphs; performance depends on KG completeness.
Increased computational complexity due to calculating entity overlaps for conditional alignment.
Evaluation limited to offline metrics; no user study or online A/B test reported.

Reproducibility

Code: https://github.com/Xiaolong-Liu-bdsc/KG-diverse

Code is publicly available at https://github.com/Xiaolong-Liu-bdsc/KG-diverse. The paper provides hyperparameter ranges and specific values used for reported results.

📊 Experiments & Results

Evaluation Setup

Top-20 Recommendation on three public datasets

Benchmarks:

MovieLens-1M (Movie Recommendation)
Last.FM (Music Recommendation)
Book-Crossing (Book Recommendation)

Metrics:

Recall@20
NDCG@20
Entity Coverage (EC@20)
Relation Coverage (RC@20)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Performance comparison on MovieLens-1M showing KG-Diverse achieves competitive accuracy with superior diversity.
MovieLens-1M	Recall@20	0.2078	0.1983	-0.0095
MovieLens-1M	Entity Coverage@20	44.68	57.48	+12.80
MovieLens-1M	Relation Coverage@20	13.67	18.32	+4.65
Ablation study on Last.FM demonstrating the contribution of the Diversified Embedding Learning (DEL) module.
Last.FM	Entity Coverage@20	39.12	65.34	+26.22
Last.FM	Recall@20	0.0882	0.0890	+0.0008

Experiment Figures

Performance comparison (Recall vs. Diversity) under different K (top-k) settings.

Main Takeaways

KG-Diverse consistently improves Entity Coverage and Relation Coverage across all datasets compared to state-of-the-art baselines like KGIN and KGAT.
The accuracy trade-off (Recall/NDCG drop) is minimal or non-existent in some cases, validating the effectiveness of the Conditional Alignment module in preserving utility.
The DEL module is the critical component for diversity; without it, the model behaves like a standard accuracy-focused recommender.

📚 Prerequisite Knowledge

Prerequisites

Graph Neural Networks (GNNs)
Knowledge Graph Embeddings (TransE/TransR)
Contrastive Learning (Alignment and Uniformity)
Recommender Systems basics (BPR Loss, Recall, NDCG)

Key Terms

_comment: REQUIRED: Define ALL technical terms, acronyms, and method names used ANYWHERE in the entire summary. After drafting the summary, perform a MANDATORY POST-DRAFT SCAN: check every section individually (Core.one_sentence_thesis, evaluation_highlights, core_problem, Technical_details, Experiments.key_results notes, Figures descriptions and key_insights). HIGH-VISIBILITY RULE: Terms appearing in one_sentence_thesis, evaluation_highlights, or figure key_insights MUST be defined—these are the first things readers see. COMMONLY MISSED: PPO, DPO, MARL, dense retrieval, silver labels, cosine schedule, clipped surrogate objective, Top-k, greedy decoding, beam search, logit, ViT, CLIP, Pareto improvement, BLEU, ROUGE, perplexity, attention heads, parameter sharing, warm start, convex combination, sawtooth profile, length-normalized attention ratio, NTP. If in doubt, define it.

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

DEL: Diversified Embedding Learning—a module proposed in this paper to generate user embeddings that are distant from their historical interaction embeddings to promote diversity

CAU: Conditional Alignment and Uniformity—a proposed technique to align item embeddings based on shared KG entities and scatter them uniformly in the vector space

LightGCN: Light Graph Convolution Network—a simplified GCN architecture for recommendation that removes feature transformation and non-linear activation

BPR Loss: Bayesian Personalized Ranking Loss—an optimization objective that maximizes the difference between the scores of observed (positive) and unobserved (negative) items

Entity Coverage: A metric measuring the number of unique Knowledge Graph entities linked to the recommended items

Relation Coverage: A metric measuring the number of unique Knowledge Graph relations linked to the recommended items

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

Recall: The fraction of relevant items that are successfully retrieved

LGC: Light Graph Convolution layer—the propagation layer used in LightGCN