HODE-MDP

The repository for Hypergraph ODE-based Multi-aspect User Dynamic Preference Modeling for Next POI Recommendation.

Requirements

• Pytorch 1.13.0
• torchdiffeq 0.2.3
• scikit-learn 0.23.2
• scipy 1.10.1
• pandas 1.5.3
• numpy 1.24.3

Create a virtual environment with conda

conda create -n hode-mdp python=3.8

conda activate hode_mdp

Install PyTorch (adjust CUDA version as needed)

conda install pytorch==1.13.0 torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

Install other dependencies

pip install torchdiffeq==0.2.3 scikit-learn==0.23.2 scipy==1.10.1 pandas==1.5.3 numpy==1.24.3

Datasets

We evaluate on four real-world LBSN datasets:

Dataset	#Users	#POIs	#Check-ins	Sparsity
NYC	834	3,835	44,686	98.61%
TKY	2,173	7,038	308,566	97.82%
SH	10,251	11,535	303,635	99.74%
Gowalla	38,857	102,077	1,911,888	99.95%

Data Preprocessing

We provide the preprocessed NYC dataset, which is located in datasets/NYC/.

train_session_label.pkl: Training sessions with labels

valid_session_label.pkl: Validation sessions

test_poi_zero.txt: Test sessions

NYC_pois_coos_poi_zero.pkl: POI coordinates

Each user’s trajectory is divided into sessions, where consecutive check-ins within a 24-hour interval belong to the same session.

Running

run HODE-MDP on NYC: python run.py --dataset NYC --t1 7 --t2 14 --t3 7 --lambda_cl 0.4

run HODE-MDP on TKY: python run.py --dataset TKY --t1 7 --t2 5 --t3 7 --lambda_cl 0.3