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

Ying Jin; Weiqing Min; Mingyu Huang; Shuqiang Jiang

cs.AIAug 25, 2026

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

Authors: Ying Jin, Weiqing Min, Mingyu Huang, Shuqiang Jiang

Organizations: School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, Beijing, China · State Key Laboratory of AI Safety, Institute of Computing Technology, Beijing, China · University of Chinese Academy of Sciences, Beijing, China

Abstract

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.

Figures & tables

Figure 1: The proposed framework for personalized sustainable diet recommendation. The architecture including three components: (1) The Recipe Sustainability Representation (left) utilizes a MoE encoder to transform ingredient tokens and quantities into a multi-dimensional sustainability embedding $\mathbf{z}_{r}$ . (2) The Preference Learning (right) leverages self-attention and interest attention to output the preference score $\hat{y}_{u,i}$ from user-recipe interactions. (3) The Personalized Constraint-aware Trade-off (bottom) synthesizes these outputs to instantiate the decision function via user-specific parameters $\boldsymbol{\lambda}_{u}$ and $\boldsymbol{\alpha}_{u}$ , as shown in the term $\boldsymbol{\alpha}_{u}^{\top}(\mathbf{z}_{r}-\boldsymbol{\lambda}_{u})$ . The entire system is optimized using a joint loss including sustainability, rating, and ranking objectives.

Category	Count
Recipe	149,036
Ingredients	7,340
Units	64
Ingredient Phrases	1,527,618
Countries	131
User	179,869

Table 1: Summary statistics of the SusDiet

Top-K	Metrics	Origin	KNN	ICLRec	MSSR	HAFR	FGCN	GRAPE	Ours
N=5	NDCG	-	0.0051	0.0119	0.0125	0.0058	0.0111	0.0105	0.0120
	Recall	-	0.0055	0.0162	0.0169	0.0075	0.0129	0.0152	0.0171
	Nutrition $\uparrow$	38.7363	34.9750	34.5969	35.6965	34.5918	38.7869	39.5836	40.7221
	Environment $\downarrow$	70.3760	67.6003	60.5654	63.4417	58.4069	62.7795	70.9837	69.9764
	Economy $\downarrow$	7.3932	8.3118	4.9420	6.8017	6.2700	5.6742	5.2464	4.9258
	Animal Welfare $\downarrow$	12.1631	16.1255	3.5907	9.7362	12.1456	6.3145	3.5327	2.7954

Table 2: Performances of different methods for Top-K recommendation. The best results are bold, and the second-best are underlined.

\uparrow

indicates that higher is better,

\downarrow

indicates that lower is better. Origin refers to the indicator values computed from users’ historically interacted recipes.

Task	MAE	RMSE	$R^{2}$
Nutrition	0.055	0.087	0.992
Environment	0.052	0.083	0.993
Economy	0.063	0.103	0.989
Animal Welfare	0.026	0.112	0.987

Table 3: Predictive performance of the multitask sustainability learning model across different dimensions.

MoE	IntAttn	Constraint	NDCG	Recall	Nutrition $\uparrow$	Environment $\downarrow$	Economy $\downarrow$	Animal Welfare $\downarrow$
$\times$	$\checkmark$	$\checkmark$	0.0155	0.0321	35.3628	68.6534	6.2455	4.9456
$\checkmark$	$\times$	$\checkmark$	0.0110	0.0259	36.6532	68.2435	5.3234	4.3553
$\checkmark$	$\checkmark$	$\times$	0.0173	0.0367	34.2454	70.8643	7.2345	5.9553
$\checkmark$	$\checkmark$	$\checkmark$	0.0159	0.0316	37.5927	66.2894	4.9445	3.0255

Table 4: Ablation of MoE, interest attention, and constraint-aware trade-off modeling on Top-20 recommendation performance. MoE denotes the mixture-of-experts sustainability encoder in recipe representation; IntAttn denotes the interest attention mechanism in the preference learning module; Constraint denotes the personalized constraint-aware decision mechanism. Best results are highlighted in bold.

\uparrow

indicates that higher is better,

\downarrow

indicates that lower is better.

Metric	History	Preference-only	Ours
IC@5	0.61	0.42	0.36
IC@10	0.61	0.55	0.45
IC@20	0.61	0.72	0.55

Table 5: Item Coverage at different top-

k

settings.

	mean	std	P10	P50	P90
$\lambda^{\text{Nutr.}}$	20.93	4.10	15.19	22.36	24.83
$\lambda^{\text{Envi.}}$	67.25	47.40	46.11	63.63	92.83
$\lambda^{\text{Econ.}}$	9.14	4.10	6.94	7.84	12.09
$\lambda^{\text{Anim.}}$	9.37	20.80	3.80	5.10	13.65

Table 6: Distribution of learned user-specific constraint thresholds

\boldsymbol{\lambda}_{u}

List	Nutr. $\uparrow$	Envi. $\downarrow$	Econ. $\downarrow$	Anim. $\downarrow$
Hist.	37.51	86.78	9.97	18.70
Pref.	38.06	87.22	10.34	14.78
Ours	39.69	71.59	6.99	12.10

Table 7: User case study comparing average sustainability indicators computed over historical interactions and Top-5 recommendations.

\uparrow

indicates that higher is better,

\downarrow

indicates that lower is better. Hist., Pref., and Ours denote the average indicators of the user’s historical interactions, preference-only recommendations, and our recommendations, respectively. Nutr., Envi., Econ., and Anim. denote nutrition, environment, economy, and animal welfare, respectively.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: Prompt of component phrase analysis

Figure 3: Prompt of sustainability indicator mapping

Figure 4: Prompt of quantity estimation

Dataset	Recipe	User	Interaction
( Bölz et al., 2023 )	507,335	302,412	1.916,424
( Majumder et al., 2019 )	231,637	226,570	1,132,367
SusDiet(Ours)	149,036	179,869	744,138

Appendix

Table 8: Statistics of the constructed user–recipe interaction dataset

Figure 5: Human evaluation results of GPT performance across (A) Component Phrase Analysis, (B) Sustainability Indicator Mapping, and (C) Quantity Estimation. Left: Score distributions for each task, with the horizontal line denoting the mean human score. Right: Percentage composition of scores for each task.

Figure 6: Statistical map of the distribution of the number of recipes in different countries

Figure 7: Frequency Distribution of the Top 20 Most Common Ingredients in the Recipe Dataset

Metric	Phi-3	KERL	Ours
NDCG@10	0.0035	0.0160	0.0136
Recall@10	0.0087	0.0243	0.0223
Nutrition@10 $\uparrow$	48.5782	43.5833	39.1429
Environment@10 $\downarrow$	125.0952	76.8827	67.8675
Economy@10 $\downarrow$	7.5879	9.5769	4.8684
Animal Welfare@10 $\downarrow$	4.3617	20.1999	2.9186

Appendix

Table 9: Performance of additional LLM-based baselines (Top-10).

Metrics	ICLRec(LLM)	MSSR(LLM)	HAFR(LLM)	FGCN(LLM)	GRAPE(LLM)	Ours(LLM)	Ours
NDCG	0.0124	0.0154	0.0092	0.0105	0.0137	0.0148	0.0159
Recall	0.0248	0.0201	0.0159	0.0224	0.0264	0.0293	0.0316
Nutrition $\uparrow$	33.6432	36.0231	36.6532	32.6542	36.4352	37.1464	37.5927
Environment $\downarrow$	63.3256	66.7433	64.6534	71.5634	68.6534	67.6534	66.2894
Economy $\downarrow$	6.6743	7.0234	7.7643	8.4534	6.4764	6.3462	4.9445
Animal Welfare $\downarrow$	10.4234	9.3454	12.3244	12.2345	8.9753	5.7434	3.0255

Appendix

Table 10: Top-20 recommendation performance of different methods using LLM for sustainability representations. The best results are in bold.

\uparrow

indicates that higher is better,

\downarrow

indicates that lower is better.

	Nutr. $\uparrow$	Envi. $\downarrow$	Econ. $\downarrow$	Anim. $\downarrow$
w/o Nutr.	35.3454	66.1432	4.9464	3.1235
w/o Envi.	37.6464	70.0443	6.5433	4.7532
w/o Econ.	36.1467	67.4213	7.0745	4.9742
w/o Anim.	37.4642	68.6434	6.3134	8.4523
Full Model	37.5927	66.2894	4.9445	3.0255

Appendix

Table 11: Ablation study on the sustainability dimensions.

\uparrow

\downarrow

indicate higher/lower is better.

Metric	History		Preference-only		Ours
Metric	Top 10%	Bottom 10%	Top 10%	Bottom 10%	Top 10%	Bottom 10%
NDCG@10	–	–	0.0123	0.0142	0.0122	0.0135
Recall@10	–	–	0.0256	0.0256	0.0230	0.0244
Nutrition@10	57.2219	22.4635	37.0862	34.5305	40.7569	35.3851
Environment@10	164.5703	19.6472	71.4178	54.7939	69.5293	46.2580
Economy@10	9.9947	4.3291	7.0616	6.1799	5.1901	4.0410

Appendix

Table 12: Performance on high-conflict (Top 10%) and low-conflict (Bottom 10%) user groups. History indicators are computed from users’ historically interacted recipes.

Historical interactions
Method	Recipe	Nutrition $\uparrow$	Environment $\downarrow$	Economy $\downarrow$	Animal Welfare $\downarrow$
History	Spicy Beef Noodle Soup	39.23	102.88	13.83	30.90
History	Mapo Tofu With Crispy Chinese Sausage	45.46	64.26	14.65	21.13
History	Kung Pao Chicken	35.14	89.33	7.63	12.43
History	Braised Pork Rice	38.91	106.96	9.32	18.88
History	Fried Rice w/ Egg & Ham	28.80	70.47	4.41	10.17

Appendix

Table 13: Detailed recipe-level results for User U1 with pork- and beef-oriented preferences.

\uparrow

\downarrow

indicate higher/lower is better.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 521 papers

ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning
Oct 6, 2026Miao Xie, Xiao Zhang, Yuan Wang +2Multi-Objective Optimization LLM Agent Planning
College of Information and Electrical Engineering, China Agricultural University · Key Laboratory of Precision Nutrition and Food Quality, China Agricultural University
An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations
May 11, 2026Yibin Wang, Yanjie Yang, Grace Melo Guerrero +2Retrieval-Augmented Generation Recommender Systems
Department of Biological and Agricultural Engineering, Texas A&M AgriLife Research, Dallas, TX, USA · Department of Agricultural Economics, Texas A&M University, College Station, TX, USA
NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines
Feb 20, 2026Junwei Wu, Runze Yan, Hanqi Luo +7Multi-Agent LLM Systems Clinical Decision Support
Emory University, Atlanta, GA, USA · Texas Woman’s University, Denton, TX, USA · Cecelia Health, New York, NY, USA

Oct 6, 2026J/K move · Enter open · S save

ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning

Miao Xie, Xiao Zhang, Yuan Wang, Ruixin Zhu+1

College of Information and Electrical Engineering, China Agricultural University · Key Laboratory of Precision Nutrition and Food Quality, China Agricultural University

Multi-Objective Optimization LLM Agent Planning

Open paper AskPDF arXiv

Dietary nutrition planning plays an important role in chronic disease management and maintaining a healthy body. In applications, it must simultaneously satisfy personalized constraints and reasonable multidimensional nutritional goals. These two aspects often conflict, and user constraints evolve with feedback, resulting in a substantial gap between generic guidelines and executable plans. To bridge this gap, we first propose the personalized fully quantified multiobjective dietary planning problem (MDP). To tackle MDP, we develop a nutrition agent, ShanLiangRen. The system first transforms dietary specifications, nutrient data, user attributes and natural language requirements into an individualized constrained planning instance. It then employs an exact retrieval-augmented generation method to shrink the feasible candidate set from a large scale ingredient and recipe space. Finally, it adopts a refinement guided by Pareto principles, where an LLM iteratively revises candidate plans under deterministic nutrition computation and feedback from constraint verification. The system outputs fully quantified meal plans with explicit ingredients and portion sizes, together with reports on nutrition compliance that show constraint satisfaction and nutrient interval attainment. We have released the system online as a WeChat Program, ShanLiangRen. A demo video is available at https://www.youtube.com/watch?v=652OtY5VlGA.

Abstract

Recommender Systems

Explore similar work

ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning

An LLM-RAG Approach for Healthy Eating Index-Informed Personalized Food Recommendations

NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines