This folder provides a reference implementation of DiffTrip, as described in the paper: "Replace and Refine: Faithful Trip Recommendation with Explicit Diffusion Guidance", which, submitted to KDD 2024 for anonymous review.
DiffTrip leverages Denoising Diffusion Probabilistic Models (DDPMs) to gradually align the generated trajectory with the tourist’s intent. The core idea stems from one of the characteristics of DDPM: the progressive data generation process. We propose & employ an explicit condition-injecting strategy during the inference stage to achieve the alignment. This strategy progressively substitutes the source & destination of the generated trajectory with the ground truth of the source & destination (from the tourist’s query), enabling the model to iteratively refine itself and ultimately produce realistic, intent-consistent trajectories.
We run the code on a computer with RTX3060, i5 12400F, and 16G memory. Please Install the dependencies via anaconda:
conda create -n DiffTrip python=3.9.18
conda activate DiffTrip
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia
conda install numpy pandas
pip install scikit-learn
| Folder Name | Description |
|---|---|
| asset | Metadata and preprocessing process |
| data | Preprocessed input data |
| results | Storage related experimental results |
| T-Base | Transformer-based model (Base) and the Base using clipping-merging strategy (Base-CM) |
| T-Base-WSE | Transformer-based model using weighted classification loss on start and end points (Weighted Source and Destination with Base, Base-WSE for short) |
| T-Diff | the source code of DiffTrip |
| T-Diff-EPS | DiffTrip predicts epsilon (EPS) |
| T-Diff-FS | DiffTrip using 4 steps fast sampling (FS) |
| README.md | -- |
| run.bat | Script file for running code |
The detailed operation mode and parameter settings of each model can be found in run.bat.
@echo off
REM Setting Python Interpreter Path
set python_path = Python location of your virtual environment
REM To Run T-Diff(DiffTrip)
REM The DiffTrip set default values T(step) in 32 and beta_t(max_noise) in 0.02, respectively. If you want to change it, try to add --step and --max_noise
python .\T-Diff\train_diffusion.py --dataset Osak --lr 0.01 --batch_size 4 --d_model 16
python .\T-Diff\train_diffusion.py --dataset Glas --lr 0.01 --batch_size 4 --d_model 64
python .\T-Diff\train_diffusion.py --dataset Edin --lr 0.01 --batch_size 16 --d_model 64
python .\T-Diff\train_diffusion.py --dataset Toro --lr 0.01 --batch_size 8 --d_model 64
REM To Run T-Diff-EPS(predict noise)
python .\T-Diff-EPS\train_diffusion.py --dataset Osak --lr 0.005 --batch_size 4 --d_model 32
python .\T-Diff-EPS\train_diffusion.py --dataset Glas --lr 0.005 --batch_size 4 --d_model 32
python .\T-Diff-EPS\train_diffusion.py --dataset Edin --lr 0.005 --batch_size 16 --d_model 32
python .\T-Diff-EPS\train_diffusion.py --dataset Toro --lr 0.005 --batch_size 8 --d_model 32
REM To Run T-Diff-FS(Fast Sampling)
python .\T-Diff-FS\train_diffusion.py --dataset Osak --lr 0.01 --batch_size 4 --d_model 16
python .\T-Diff-FS\train_diffusion.py --dataset Glas --lr 0.01 --batch_size 4 --d_model 64
python .\T-Diff-FS\train_diffusion.py --dataset Edin --lr 0.01 --batch_size 16 --d_model 64
python .\T-Diff-FS\train_diffusion.py --dataset Toro --lr 0.01 --batch_size 8 --d_model 64
REM To Run T-Base(Base)
REM max f1(pairs-f1) represent the results of Base-CM and total f1(pairs-f1) represent the results of Base
python .\T-Base\train_base.py --dataset Osak --lr 0.001 --batch_size 4 --d_model 128
python .\T-Base\train_base.py --dataset Glas --lr 0.001 --batch_size 4 --d_model 128
python .\T-Base\train_base.py --dataset Edin --lr 0.001 --batch_size 16 --d_model 128
python .\T-Base\train_base.py --dataset Toro --lr 0.001 --batch_size 8 --d_model 64
REM To Run T-Base-WSE(Weight of Start and End)
python .\T-Base-WSE\train_base.py --dataset Osak --lr 0.001 --batch_size 4 --d_model 128 --se_weight 5
python .\T-Base-WSE\train_base.py --dataset Glas --lr 0.001 --batch_size 4 --d_model 128 --se_weight 5
python .\T-Base-WSE\train_base.py --dataset Edin --lr 0.001 --batch_size 16 --d_model 128 --se_weight 5
python .\T-Base-WSE\train_base.py --dataset Toro --lr 0.001 --batch_size 8 --d_model 64 --se_weight 5
If your operating system is Windows, you can use the command In the working directory as
.\run.bat
to run this script file directly. You can also directly paste commands into the terminal to run the program just like
python .\T-Diff\train_diffusion.py --dataset Osak --lr 0.01 --batch_size 4 --d_model 16
% Hope such an implementation could help you on your projects. Any comments and feedback are appreciated.