PaperlayerView exact revision on arXiv

Source-linked AI summary

Eating for a Sustainable Planet: Personalized Sustainable Diet Recommendation via Constraint-Aware Decision-Making Modeling

Ying Jin, Weiqing Min, Mingyu Huang, Shuqiang Jiang

arXiv:2608.24274v1cs.AI

TL;DR

The paper addresses the difficulty of translating population-level sustainable-diet guidance into recommendations that accommodate heterogeneous individual preferences and acceptability. It formulates recommendation as constraint-aware decision-making with learnable user-specific sustainability constraints, evaluates the approach using SusDiet, and reports more sustainable choices with competitive preference accuracy.

  • Problem

    Population-level sustainability models and preference-focused recommenders do not adequately address heterogeneous individual dietary behavior and the acceptability of required changes.

  • Method

    The framework jointly learns user preferences and user-specific sustainability constraint thresholds, using sustainability representations and interaction data to rank recipes.

  • Results

    The method produces substantially more sustainable recommendations while maintaining competitive recommendation accuracy on SusDiet.

  • Takeaways & Limitations

    The framework provides a basis for personalized sustainable dietary decision-making and quantitative evidence relevant to sustainable-diet interventions and policy-making.

  • Takeaways & Limitations

    Constraints inferred from historical online recipe interactions may only partially reflect users’ true sustainability tolerances, and offline evaluation leaves real-world behavioral impact unvalidated.

Abstract

from arXiv · show

A sustainable diet represents a multi-dimensional synergy among four essential pillars: nutrition adequacy, economic affordability, cultural acceptability, and environmental respect. Despite the prevalence of population-level sustainability modeling, practical implementation relies on effective individual-level adoption. This transition is often hindered by inter-individual heterogeneity, posing a formidable challenge in aligning sustainable diet requirements with individual preferences. To address this issue, we propose a personalized sustainable diet recommendation model based on a constraint-aware decision-making mechanism, where sustainability is incorporated through learnable constraints rather than modeled as user preferences. To systematically evaluate the proposed approach, we construct a sustainable diet dataset named SusDiet with about 150k recipes, characterized by broad coverage of sustainability indicators. Experimental results on this dataset show that our method promotes more sustainable choices without compromising individual preference. This work establishes a framework for aligning individual dietary choices with planetary health, offering quantitative evidence to guide future sustainable diet interventions and policy-making for sustainable development.

1. Introduction

Sustainable diets must jointly address health, environmental, economic, safety, and cultural considerations, yet population-level guidance often overlooks individual heterogeneity and acceptability. The paper proposes a constraint-aware personalized recommendation framework, introduces SusDiet, and reports more sustainable recommendations with competitive accuracy.

  • Sustainable diets combine individual health and well-being with low environmental impact, affordability, safety, and cultural acceptability.
  • Population-level optimization treats sustainability as an explicit objective or hard constraint, limiting its application to heterogeneous individual dietary behavior.
  • Major changes to staple foods, meat, and dairy can exceed individual acceptability, reducing willingness to change dietary behavior.
  • The framework models sustainability through learnable, user-specific constraint thresholds while jointly learning preferences in a unified decision-making process.Its components include an MoE-based sustainability model, multi-interest preference learning, and personalized sustainability constraints.
  • SusDiet contains approximately 150k recipes annotated across nutrition, environmental impact, economic cost, and animal welfare, aligned with user-recipe interactions.
  • Experiments show that user-specific sustainability constraints yield more sustainable recommendations while maintaining competitive recommendation accuracy.

2. Related Work

Sustainable-diet research spans macro-level optimization for populations and micro-level recommendation for individuals. Macro approaches provide policy evidence, while individual-level systems offer personalization but have focused mainly on limited sustainability dimensions.

  • Sustainable-diet research divides broadly into macro-level optimization for policy formulation and micro-level modeling for individual behavioral interventions.
  • Macro-level Modeling and Optimization: Macro-level methods design population or national dietary patterns that satisfy nutrition while considering environmental impact, cost, or cultural constraints.
  • Macro-level Modeling and Optimization: These policy-centric approaches provide evidence about sustainable food-system feasibility and inform global strategies, but produce macro-level optimal patterns rather than individual decision models.
  • Individual-level Recommendations: Individual-level recommendation systems can deliver tailored dietary guidance by learning preferences at scale, but recent sustainability-oriented work primarily addresses nutritional and environmental dimensions.

3. Task Formulation

The task is to recommend recipes that match user tastes while respecting user-specific sustainability boundaries. It learns sustainability constraints and trade-off weights, then maximizes a combined preference-and-constraint score to select each user’s top-K recipes.

  • Personalized sustainable diet recommendation seeks recipes satisfying user taste preferences while adhering to unique boundaries across sustainability metrics.
  • Each recipe is represented by an N-dimensional sustainability vector, with one indicator for each sustainability criterion.
  • For each user, the model defines a constraint vector whose nth bound specifies the sustainability-indicator value that user is willing to tolerate.
  • All indicators are normalized into a unified minimization form, so lower values represent more sustainable outcomes.
  • The objective learns user-specific sustainability constraints and trade-off weights, recommending the top-K recipes that maximize F(u, r).

4. Methodology

The framework combines sustainability representation, preference learning, and personalized constraint-aware trade-offs in a unified decision-making model. It jointly optimizes sustainability, rating, and ranking objectives while modeling user-specific sustainability constraints.

  • Framework Overview: The unified framework combines recipe sustainability representation, preference learning, and personalized constraint-aware trade-offs.Recipe embeddings, preference scores, and user-specific parameters are synthesized into the decision function.
  • Recipe Sustainability Representation: A MoE encoder maps ingredient compositions and quantities into sustainability-aware recipe embeddings covering heterogeneous sustainability indicators.The encoder uses four Transformer-based experts and a gating network to capture diverse ingredient patterns.
  • Recipe Sustainability Representation: The sustainability module jointly predicts multiple indicators from recipe embeddings using task-specific heads and uncertainty-weighted multi-task learning.Learnable parameters balance task gradients according to their respective noise levels.
  • Preference Learning: Preference learning estimates user–recipe scores from historical interactions using chronological encoding, self-attention, and multi-interest attention.Identity, ingredient, and regional recipe representations are processed separately before being combined into the final user representation.
  • Preference Learning: The final user representation aggregates identity, ingredient, and regional interests before producing a preference score through an inner product.This score is trained with rating accuracy and ranking consistency objectives based on positive–negative recipe pairs.
  • Constraint-aware Trade-off: A user-specific sustainability penalty uses learnable constraint parameters, and minimizing it reduces violations of personalized sustainability constraints.The overall objective also includes ℓ2 regularization on constraint-related parameters for stable optimization.

5. Construction of Dataset

SusDiet integrates recipes from multiple public sources with standardized ingredients, country labels, multidimensional sustainability indicators, and user–recipe interactions. GPT-based processing supports ingredient structuring and indicator mapping, with human evaluation reporting high acceptability across the three processing tasks.

  • Dataset Overview: SusDiet aligns recipes with nutritional, environmental, economic, and animal-welfare indicators and large-scale user interaction data.The dataset is designed to support personalized decision-making under sustainability constraints.
  • Data Collection: The recipe collection integrates recipeDB, Yummly-28k, Yummly-66k, WorldCuisines, and World Wide Recipe sources.Each recipe is characterized by a standardized ingredient list and a country label.
  • Data Processing: GPT performs component phrase analysis, quantity estimation, and sustainability indicator mapping to structure ingredient descriptions.The processing aligns ingredient information with food product categories used by sustainability indicator sources.
  • Data Processing: 97%, 91%, and 94% of samples were rated acceptable for component phrase analysis, weight estimation, and sustainability indicator mapping, respectively.The paper reports these human-evaluation results as evidence on the reliability of the three GPT-supported tasks.
  • Interaction Alignment: Recipe records are linked to explicit user ratings through name matching and metadata verification, producing clean user–recipe–rating tuples.The interaction data comes from HUMMUS and the dataset introduced by Majumder et al.
  • Dataset Integration: The final dataset integrates recipes, standardized ingredients, multidimensional sustainability indicators, country labels, and user interactions.Table 1 summarizes the resulting dataset statistics.

6. Empirical Evaluations

The evaluation measures preference alignment, multidimensional sustainability, predictive quality, ablation effects, coverage, personalization of constraints, and a user-level case study. Results show competitive recommendation accuracy alongside balanced sustainability improvements, with explicit constraints contributing substantially.

  • Evaluation Setup: The framework is evaluated on personalized top-K recipe recommendation at K = 5, 10, and 20 using accuracy and four sustainability dimensions.NDCG@K and Recall@K assess learned preference alignment, while nutrition, environment, economy, and animal welfare are averaged over recommended recipes.
  • Overall Results: Our method achieves competitive NDCG and Recall, reaching the best or near-best results across several cutoff values.The comparison includes general-purpose and food recommendation baselines.
  • Overall Results: The method produces high Nutrition scores, low Economy and Animal Welfare values, and a balanced sustainability profile despite not achieving the lowest Environment score.It consistently outperforms historical user interactions across all sustainability dimensions.
  • Sustainability Prediction: The multitask sustainability model achieves low MAE and RMSE with consistently high R2 scores across four dimensions.These predictions provide sustainability estimates for the recommendation framework.
  • Ablation Study: Removing the personalized constraint-aware trade-off substantially worsens sustainability indicators, while removing interest attention lowers NDCG and Recall.Removing the MoE sustainability representation also degrades sustainability indicators, with only slight effects on ranking accuracy.
  • Coverage and Personalization: Sustainability constraints moderately reduce item coverage but retain coverage comparable to observed interactions, while learned environmental thresholds range from 46.11 to 92.83 across users.Less sustainable items are ranked lower rather than strictly excluded, and threshold variation indicates heterogeneous user boundaries.

7. Discussion

The framework reframes personalized sustainable diet recommendation as constraint-aware decision-making, combining user-specific sustainability constraints with preference modeling. Its scope remains bounded by historical interaction data, static assumptions, uncertain indicators, and offline evaluation.

  • Framework: The framework models sustainability as learnable, user-specific constraints rather than explicit preferences, integrating them with preference learning in a unified decision process.Sustainability signals dynamically penalize recipes exceeding individual constraint thresholds.
  • Framework: The method combines MoE-based sustainability representation learning, multi-interest preference modeling, and personalized constraint learning under one optimization objective.
  • Significance: The formulation is intended to accommodate heterogeneous dietary preferences while incorporating sustainability objectives into adaptive recommendation strategies.
  • Limitations: Historical online recipe interactions may only partially reflect users’ true sustainability tolerance because exposure bias and unobserved socioeconomic factors affect observed behavior.
  • Limitations: The framework assumes static preferences and constraints, uses potentially uncertain and simplified sustainability metrics, and has been evaluated only offline.Long-term adherence, trust, and unintended effects therefore remain unvalidated in online or user-centered settings.

Impact Statement

The work targets more informed, personalized dietary decision-making as part of broader sustainability and public-health challenges. It explores preference-aware food choices as a pathway toward wider adoption of sustainable eating habits.

  • Impact Statement: The research explores personalized, preference-aware dietary choices intended to support broader adoption of sustainable eating habits.The stated broader context includes sustainability, public health, and responsible consumption.
  • Impact Statement: At the consumer level, the work connects dietary decision-making with potential reductions in greenhouse-gas emissions and land-use pressures while supporting human health.

A. Extended Methodology

The methodology builds sustainability-aware recipe representations, learns multi-interest user preferences from sequential interactions, and combines both for personalized recommendation. Recipes are standardized from heterogeneous public sources before multi-task sustainability prediction and preference scoring.

  • Recipe Representation: Recipes are represented as fixed-length ingredient sequences retaining canonical ingredient tokens, quantities, and masks for valid positions.Variable-length ingredient lists are truncated or padded to length L.
  • Recipe Representation: Ingredient identity and quantity are embedded separately, concatenated, and linearly projected into the model’s representation space.
  • Sustainability Representation: A Mixture-of-Experts encoder with E = 4 Transformer-based experts uses masked pooling and learned gating to capture heterogeneous ingredient compositions.The final recipe embedding is a gated combination of expert outputs, allowing experts to specialize in distinct compositions.
  • Sustainability Prediction: The shared recipe encoder predicts K = 4 sustainability attributes through independent task-specific regression heads.Learned uncertainty-based task weighting balances gradients across tasks with different noise levels and scales.
  • Sustainability Prediction: Sustainability targets are direction-adjusted, log-transformed, quantile-clipped, and z-score normalized using training recipes.Training-set quantiles mitigate outliers, while training-only µ and σ support normalization.
  • Preference Learning: Multi-interest self-attention learns chronological, multi-view dietary preferences from user interaction sequences and target-recipe identity, ingredient, and region features.Position embeddings preserve sequence order, while attention produces context-aware preference representations and a user–recipe preference score.
  • Dataset Construction: The recipe corpus combines recipeDB, Yummly-28k, Yummly-66k, WorldCuisines, and World Wide Recipe, preserving country or regional labels alongside ingredient lists.Cleaning and deduplication address heterogeneous formatting, ingredient naming, and unit systems.

B.7. Human Evaluation of GPT-Based Processing

Human evaluation assessed GPT-based component analysis, sustainability mapping, and quantity estimation on sampled instances using 1–4 scores. GPT outputs were generally rated acceptable, while a two-stage parsing procedure outperformed direct raw-phrase estimation.

  • Evaluation Protocol: 500 instances per GPT-processing task were human-scored from 1–4 for correctness and semantic sufficiency.The evaluated tasks were component phrase analysis, sustainability indicator mapping, and quantity estimation.
  • Results: 3.776, 3.664, and 3.352 were the average human scores for component phrase analysis, sustainability indicator mapping, and quantity estimation, respectively.Figure 5 reports score distributions and percentage compositions for the three tasks.
  • Results: 97%, 94%, and 91% of samples received acceptable scores between 3 and 4 across the three GPT tasks, respectively.
  • Weight Estimation: 3.36 versus 2.85 was the average manual score for two-stage parsing followed by weight estimation versus direct raw-phrase estimation.The comparison used a 4-point scale on 200 samples per approach.
  • Scoring Criteria: Component phrase scores assess quantity, unit, and ingredient identification; mapping scores assess reference or category agreement; quantity scores assess numerical reasonableness.These task-specific criteria define what scores from 1 to 4 represent.
  • Dataset Characteristics: The dataset exhibits a long-tail ingredient distribution, broad but uneven geographic coverage, and variation in sustainability indicators across recipes and regions.

C.1. Implementation Details

The evaluation compares the proposed framework with conventional, LLM-based, and ablated alternatives using sustainability-aware recommendation settings. Results emphasize the value of structured sustainability modeling and explicit dimension-specific constraints.

  • Implementation Details: The study tunes learning rate, embedding size, batch size, attention heads, and loss trade-off coefficients over predefined search ranges.The reported settings include α = 0.5 and ranges for ranking, sustainability penalty, and L2 regularization weights.
  • Evaluation Setup: SusDiet evaluation includes structured recipe and user-interaction data, with LLM-based methods assessed through sustainability representations and QA-style sustainability constraints.The LLM evaluation uses structured recipe graphs, user behavior sequences, and sustainability-attribute constraints.
  • LLM Comparison: KERL slightly exceeds the proposed method on NDCG@10 and Recall@10 but performs substantially worse on Environment, Economy, and Animal Welfare metrics.The passage attributes this pattern to KERL’s lack of strict sustainability constraints and greater preference-oriented trade-offs.
  • LLM Comparison: The proposed method consistently outperforms LLM-based variants in recommendation accuracy and sustainability performance.Direct LLM sustainability outputs are described as lacking structured, reliable, fine-grained modeling of sustainability dimensions.
  • Ablation Study: Removing any single sustainability constraint degrades its corresponding sustainability dimension and can also harm other dimensions.The ablation isolates each sustainability constraint while keeping the rest of the framework unchanged.

D.4. Analysis on High-Conflict Scenarios

The boundary test examines users at opposite ends of historical environmental sustainability and compares the preference-only baseline with the full model. The full model improves sustainability outcomes even for high-conflict users while limiting accuracy loss.

  • Boundary Test: The test stratifies users into Top 10% least-sustainable and Bottom 10% most-sustainable groups using historical environmental scores.Recommendation accuracy and sustainability indicators are compared under the preference-only baseline and the full model.
  • Robustness: The results indicate that sustainability constraints remain effective at extreme preference–sustainability boundaries while preserving recommendation accuracy to a substantial extent.The authors frame this as robustness in resolving high-conflict dietary trade-offs.
  • High-Conflict Users: For high-conflict users, the full model improves all four sustainability dimensions versus the preference-only baseline with only a marginal recommendation-accuracy loss.These users have preferences most opposed to sustainability requirements.
  • Low-Conflict Users: The most-sustainable user group shows even greater gains in sustainability indicators under the full model.The reported comparison uses the Bottom 10% group as the low-conflict boundary.
Loading 2608.24274v1…