MOPI-HFRS: A Multi-objective Personalized Health-aware Food Recommendation System with LLM-enhanced Interpretation

📝 Paper Summary

Health-aware Food Recommendation Graph Neural Networks for Recommendation LLM-based Interpretation

MOPI-HFRS creates a personalized food recommendation system that balances user preference, health needs, and diversity by refining graph structures and using LLMs to explain why recommendations are healthy.

Core Problem

Existing food recommendation systems prioritize user preference over health, lack personalization based on medical conditions (like hypertension or diabetes), and fail to explain why a recommendation is healthy.

Why it matters:

Unhealthy diets contribute to 11 million deaths annually and 42.4% obesity rates in the US, yet platforms like Yelp ignore health outcomes.
A 'healthy' food varies by individual (e.g., high-fiber is good for some but bad for others), meaning generic health scores are insufficient.
Without interpretability, users cannot understand or trust why a recommended substitution is better for their specific condition.

Concrete Example: A user with high blood pressure needs a low-sodium diet, but a standard recommender might suggest high-sodium options based on taste history. Current health-aware systems might hard-code a 'health score' post-filtering, failing to balance this with the user's taste preferences or explain the medical reasoning.

Key Novelty

Multi-Objective Personalized Interpretable Health-aware Food Recommendation System (MOPI-HFRS)

Constructs a signed bipartite graph where edges represent not just interactions but 'healthy' vs 'unhealthy' relationships based on user medical data and food nutrient profiles.
Uses a structure learning module that dynamically rewires the graph based on feature similarity and health compatibility (balance theory) during training to reduce noise.
Employs Pareto optimization to simultaneously maximize three conflicting objectives: user preference (accuracy), personalized healthiness, and nutritional diversity.

Architecture

The overall MOPI-HFRS framework, illustrating the parallel processing of raw, feature-based, and health-based graphs, their fusion, and the multi-objective optimization targets.

Evaluation Highlights

Outperforms state-of-the-art baselines in Health metric by significant margins (e.g., +15-20% relative improvement) while maintaining competitive Recall/NDCG.
Achieves higher diversity (coverage) compared to preference-only models like LightGCN.
LLM-enhanced interpretation module provides verifiable medical reasoning for recommendations, bridging the gap between numerical scores and user understanding.

Breakthrough Assessment

7/10

Strong contribution in constructing the first large-scale personalized health-aware benchmark using real medical data (NHANES). The methodology effectively combines graph learning with multi-objective optimization, though the LLM component is primarily for interpretation rather than the core recommendation logic.

⚙️ Technical Details

Problem Definition

Setting: Multi-objective Top-K Recommendation on a Bipartite Graph

Inputs: User set U, Food set F, Interaction edges E, User features X_U (demographics + medical), Food features X_F (nutrients + text)

Outputs: Ranked list of foods for each user optimizing Preference, Healthiness, and Diversity simultaneously

Pipeline Flow

Data Processing: Tagging users and foods based on medical data and nutrient standards
Graph Structure Learning: Refining the graph using feature similarity and health compatibility
Encoder: GNN-based embedding generation
Multi-Objective Optimization: Pareto optimization for Preference, Health, and Diversity
LLM Interpretation: Generating explanations for recommendations

System Modules

Health-aware Graph Structure Learning

Refine noisy interaction graph into a cleaner structure for recommendation

Model or implementation: Custom Graph Learning Module

Backbone Encoder

Generate user and food embeddings

Model or implementation: LightGCN (as base encoder)

Pareto Optimization

Optimize multiple conflicting losses simultaneously

Model or implementation: Multi-Gradient Descent Algorithm (MGDA)

LLM Interpreter

Explain why a food is recommended based on health tags

Model or implementation: Not specified (Generic LLM via prompting)

Novel Architectural Elements

Health-aware Graph Structure Learning module that fuses semantic health graphs (signed edges based on medical compatibility) with feature similarity graphs using a channel attention mechanism
Structure pooling layer that aggregates multiple learned graph views (raw, feature-based, health-based) into a single optimized adjacency matrix end-to-end

Modeling

Base Model: LightGCN (backbone for recommendation)

Training Method: End-to-end training with Pareto Multi-Objective Optimization

Objective Functions:

Purpose: Maximize user preference matching.

Formally: Bayesian Personalized Ranking (BPR) loss: L_pref = -sum(ln(sigmoid(y_ui - y_uj)))
Purpose: Minimize recommendation of unhealthy foods.

Formally: Health loss (margin ranking loss pushing healthy items above unhealthy ones)
Purpose: Maximize nutritional diversity.

Formally: Diversity loss (Distance Correlation between recommended items)

Training Data:

Constructed from NHANES (2003-2020) data
Users tagged with medical conditions (e.g., 'low_sodium' for hypertension)
Foods tagged with nutrient levels (e.g., 'high_sodium', 'low_fat')

Key Hyperparameters:

learning_rate: 1e-3
embedding_size: 64
batch_size: 2048
+ 2 more
layers: 3 (LightGCN)
epoch: 100

Compute: Not reported in the paper

Comparison to Prior Work

vs. LightGCN: MOPI-HFRS adds health and diversity objectives and graph structure learning.
vs. H-GCN: MOPI-HFRS personalizes health (based on user medical tags) rather than using a global health score, and optimizes via Pareto rather than weighted sum.
vs. KGAT: MOPI-HFRS explicitly models 'healthy/unhealthy' signed edges rather than generic knowledge graph relations.

Limitations

Relies on accuracy of self-reported or clinically measured NHANES data which may have noise.
Requires explicit mapping of medical conditions to nutrient rules (e.g., Hypertension -> Low Sodium), which might be an oversimplification.
Computational cost of multi-objective Pareto optimization is generally higher than scalarization methods.

Reproducibility

Code: https://github.com/ZheyuanZhang1/MOPI-HFRS

Code and benchmark datasets are publicly available at https://github.com/ZheyuanZhang1/MOPI-HFRS. The paper uses public NHANES data but the specific processing scripts to generate the signed graph are part of the contribution.

📊 Experiments & Results

Evaluation Setup

Top-K Recommendation (K=20) on two NHANES-derived benchmarks

Benchmarks:

NHANES-Macro (Food Recommendation) [New]
NHANES-Full (Food Recommendation (Macro + Micro nutrients)) [New]

Metrics:

Recall@K
NDCG@K
Health@K (Proportion of healthy items in top K)
Diversity (ILD - Intra-List Distance)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
MOPI-HFRS achieves the best balance of health and accuracy on the NHANES-Macro dataset.
NHANES-Macro	Health@20	0.3842	0.8953	+0.5111
NHANES-Macro	Recall@20	0.0521	0.0543	+0.0022
NHANES-Macro	NDCG@20	0.0418	0.0432	+0.0014
Results on NHANES-Full (including micro-nutrients) show similar trends, validating robustness.
NHANES-Full	Health@20	0.3512	0.8876	+0.5364
NHANES-Macro	Recall@20	0.0498	0.0543	+0.0045

Experiment Figures

Performance comparison (Recall, NDCG, Health, Diversity) on NHANES-Macro and NHANES-Full datasets.

Ablation study removing different components (e.g., w/o Structure Learning, w/o Pareto).

Main Takeaways

MOPI-HFRS significantly improves the healthiness of recommendations (often doubling the health score) without sacrificing user preference (Recall/NDCG).
Structure learning is critical; removing it leads to drops in both accuracy and health metrics, validating the need to denoise the raw interaction graph.
Multi-objective Pareto optimization successfully balances conflicting goals (Preference vs. Health) better than linear scalarization methods.
The method generalizes well across both Macro-nutrient only and Full-nutrient settings.

📚 Prerequisite Knowledge

Prerequisites

Graph Neural Networks (GNNs)
Multi-objective Optimization (Pareto optimality)
Bipartite Graphs
Large Language Models (for explanation)

Key Terms

NHANES: National Health and Nutrition Examination Survey—a program of studies designed to assess the health and nutritional status of adults and children in the United States

Pareto optimization: An optimization approach that finds solutions where no objective can be improved without degrading another

Signed bipartite graph: A graph where edges between two sets of nodes (users and foods) have signs (positive/healthy or negative/unhealthy)

Balance theory: A social network theory suggesting that 'the friend of my friend is my friend' (positive cycle) and 'the enemy of my enemy is my friend'—used here to aggregate embeddings for healthy/unhealthy relationships

Structure learning: Ideally learning the optimal graph connectivity matrix rather than just using the raw, noisy input adjacency matrix

Macro/Micro-nutrients: Macro (fat, protein, carbs) are needed in large amounts; Micro (vitamins, minerals like sodium) are needed in smaller amounts but are critical for specific conditions