GRAPE

This is the implementation for our AAAI2025 paper:

Bites of Tomorrow: Personalized Recommendations for a Healthier and Greener Plate

Environment

We use Python language and Pytorch library to establish our model.

For the detailed environment, please follow requirements.txt. (Updated at 2025 Jun. 10th)

Dataset

We leverage the dataset introduced in GreenRec: A Large-Scale Dataset for Green Food Recommendation

Download

The dataset is available at GreenFood Dataset

Leveraged File

We use valid_data.txt which contains the interaction records, and recipe_three_scores.csv which contains sustainability indicator information for all recipes.

Run GRAPE

Dataset Preprocessing

Please preprocess the dataset to capture train, validation, and test set for GRAPE with

python ./transfer.py

Model Training and Evaluation

Run GRAPE with

python ./run_model.py --dataset="Green_Rec" --hidden_size=256 --n_layers=3 --n_heads=2 --tau=1 --config_files="configs/Green_Rec.yaml" --priority=0

Experiment Results

Please refer to our new experiment results:

Main Performances

Performances of different methods for Top-N recommendation. The best results are bold, and the second-best are italicized. A lower EIS indicates greater environmental friendliness, whereas higher NIS and HMI values denote more nutritious and healthier food, respectively.

Top-N	Metrics	BPR	KNN	SHT	STOSA	ICLRec	NOVA	CAFE	FDSA-CL	MSSR	GRAPE($\mathcal{L}_{np}$)
	HR	0.0095	0.0107	0.0125	0.0113	0.0134	0.0137	0.0146	0.0173	0.0180	0.0159
	NDCG	0.0080	0.0095	0.0098	0.0096	0.0125	0.0128	0.0137	0.0157	0.0164	0.0151
5	EIS↓	89.58	83.70	96.43	113.92	97.78	89.67	88.16	116.43	116.33	109.35
	NIS↑	32.59	31.23	33.83	35.51	35.26	36.78	37.10	36.39	34.39	33.70
	HMI↑	44.56	45.94	40.56	35.38	35.31	44.69	43.50	43.01	45.99	45.82
	HR	0.0164	0.0188	0.0213	0.0207	0.0224	0.0243	0.0242	0.0269	0.0275	0.0285
	NDCG	0.0112	0.0132	0.0149	0.0141	0.0168	0.0176	0.0181	0.0202	0.0208	0.0210
10	EIS↓	88.82	84.90	86.32	88.56	104.54	88.38	79.43	92.43	83.79	83.08
	NIS↑	31.79	31.15	32.16	34.87	31.10	35.85	33.84	34.80	31.10	31.28
	HMI↑	42.10	42.91	43.98	43.88	43.46	43.36	44.11	43.30	44.05	44.26
	HR	0.0269	0.0296	0.0344	0.0335	0.0343	0.0372	0.0399	0.0426	0.0437	0.0472
	NDCG	0.0150	0.0171	0.0191	0.0187	0.0211	0.0222	0.0238	0.0259	0.0266	0.0279
20	EIS↓	85.17	80.93	85.63	89.24	80.82	81.89	70.86	78.48	76.98	76.42
	NIS↑	30.83	30.67	31.36	32.39	32.19	35.27	33.22	33.45	33.48	33.78
	HMI↑	43.91	42.75	42.95	43.62	42.26	43.19	42.38	43.05	44.36	44.67

Effectiveness of Prioritized Green Loss

Top-10 performance of GRAPE with Non-prioritized Green Loss ($\mathcal{L}{ng}$), and Prioritized Green Loss($\mathcal{L}{pg}$) applying different priority orders. ($1^{st}$) denotes the highest priority, ($2^{nd}$) denotes the second priority, and ($3^{rd}$) denotes the lowest priority. The best results for each evaluation metric are bold.

Model	HR	NDCG	EIS↓	NIS↑	HMI↑
	0.0288	0.0194	($1^{st}$) 79.47	($2^{nd}$) 30.38	($3^{rd}$) 42.57
	0.0276	0.0205	($1^{st}$) 75.2	($3^{rd}$) 31.85	($2^{nd}$) 41.44
	0.0264	0.0198	($2^{nd}$) 76.31	($1^{st}$) 33.4	($3^{rd}$) 42.41
GRAPE($\mathcal{L}_{pg}$)	0.0286	0.0201	($3^{rd}$) 89.67	($1^{st}$) 33.02	($2^{nd}$) 41.68
	0.0285	0.0202	($2^{nd}$) 79.99	($3^{rd}$) 32.56	($1^{st}$) 44.27
	0.0285	0.0206	($3^{rd}$) 81.41	($2^{nd}$) 30.92	($1^{st}$) 43.19
GRAPE($\mathcal{L}_{ng}$)	0.0285	0.0210	83.08	31.28	44.26

Acknowledgement

This repository is based on RecBole and MSSR.

Citation

@inproceedings{jing2025bites,
  title={Bites of Tomorrow: Personalized Recommendations for a Healthier and Greener Plate},
  author={Jing, Jiazheng and Zhang, Yinan and Miao, Chunyan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2025}
}