ExGeo

2.Update log(23.10.27): We modified the prior probability parameter p=0.1 for Aib. It will be more obvious to use a smaller probability parameter if we want to reproduce the interpretability experiment. It is worth noting that different weights 'beta' will eventually make the learned Q(Aib/A) converge to the prior distribution P(A). However, by our careful observation of the convergence process, we can see that a larger weight 'beta' will retain fewer nodes with the same performance(mse,median,rmse), which is consistent with our experimental results.

1.Update log(23.10.19): We supplement code for print the mask_matrix in main.py which can show the change of matrix by change the number of epoch.

This folder provides a reference implementation of ExGeo.

Basic Usage

Requirements

The code was tested with python 3.8.13, pytorch 1.12.1, cudatoolkit 11.6.0, and cudnn 7.6.5. Install the dependencies via Anaconda:

# create virtual environment
conda create --name ExGeo python=3.8.13

# activate environment
conda activate ExGeo

# install pytorch & cudatoolkit
conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge
# install other requirements
conda install numpy pandas
pip install scikit-learn

Run the code

# Open the "ExGeo" folder
cd ExGeo

# data preprocess (executing IP clustering). 
python generateidx.py --dataset "New_York"
python generateidx.py --dataset "Los_Angeles"
python generateidx.py --dataset "Shanghai"

python preprocess.py --dataset "New_York"
python preprocess.py --dataset "Los_Angeles"
python preprocess.py --dataset "Shanghai"
# run the model ExGeo
python main.py --dataset "New_York" --dim_in 30 --lr 2e-3 --saved_epoch 10
python main.py --dataset "Los_Angeles" --dim_in 30 --lr 2e-3 --saved_epoch 10
python main.py --dataset "Shanghai" --dim_in 51 --lr 1e-3 --saved_epoch 10

# load the checkpoint and then test
python test.py --dataset "New_York" --dim_in 30 --lr 2e-3 --load_epoch 100
python test.py --dataset "Los_Angeles" --dim_in 30 --lr 2e-3 --load_epoch 100
python test.py --dataset "Shanghai" --dim_in 51 --lr 1e-3 --load_epoch 70

The description of hyperparameters used in main.py

Hyperparameter	Description
seed	the random number seed used for parameter initialization during training
model_name	the name of model
dataset	the dataset used by main.py
lambda_1	the trade-off coefficient of data perturbation in loss function
lambda_2	the trade-off coefficient of parameter perturbation in loss function
lr	learning rate
harved_epoch	when how many consecutive epochs the performance does not increase, the learning rate is halved
early_stop_epoch	when how many consecutive epochs the performance does not increase, the training stops.
saved_epoch	how many epochs to save checkpoint for the testing
dim_in	the dimension of input data
dim_med	the dimension of middle layers
dim_z	the dimension of vector representation
eta	magnitude of data disturbance
zeta	magnitude of parameter disturbance
step	times of gradient ascent in a single parameter disturbance
mu	inner learning rate of parameter disturbance
c_mlp	when predicting if use collaborative_mlp or not
epoch_threshold	when we start adding perturbation both in data and parameter

Folder Structure

└── ExGeo
	├── datasets # Contains three large-scale real-world street-level IP geolocation datasets.
	│	|── New_York # Street-level IP geolocation dataset collected from New York City including 91,808 IP addresses.
	│	|── Los_Angeles # Street-level IP geolocation dataset collected from Los Angeles including 92,804 IP addresses.
	│	|── Shanghai # Street-level IP geolocation dataset collected from Shanghai including 126,258 IP addresses.
	├── lib # Contains model implementation files
	│	|── layers.py # The code of the attention mechanism.
	│	|── model.py # The core source code of proposed RIPGeo
	│	|── sublayers.py # The support file for layer.py
	│	|── utils.py # Auxiliary functions
	├── asset # Contains saved checkpoints and logs when running the model
	│	|── log # Contains logs when running the model 
	│	|── model # Contains the saved checkpoints
        ├── generateidx.py # generate the idx of traget nodes and landmark nodes
	├── preprocess.py # Preprocess dataset and execute IP clustering the for model running
	├── main.py # Run model for training and test
	├── test.py # Load checkpoint and then test
	└── README.md

Dataset Information

The "datasets" folder contains three subfolders corresponding to three large-scale real-world street-level IP geolocation datasets collected from New York City, Los Angeles and Shanghai. There are three files in each subfolder:

• data.csv # features (including attribute knowledge and network measurements) and labels (longitude and latitude) for street-level IP geolocation
• ip.csv # IP addresses
• last_traceroute.csv # last four routers and corresponding delays for efficient IP host clustering

The detailed columns and description of data.csv in New York dataset are as follows:

New York

Column Name	Data Description
ip	The IPv4 address
as_mult_info	The ID of the autonomous system where IP locates
country	The country where the IP locates
prov_cn_name	The state/province where the IP locates
city	The city where the IP locates
isp	The Internet Service Provider of the IP
vp900/901/..._ping_delay_time	The ping delay from probing hosts "vp900/901/..." to the IP host
vp900/901/..._trace	The traceroute list from probing hosts "vp900/901/..." to the IP host
vp900/901/..._tr_steps	#steps of the traceroute from probing hosts "vp900/901/..." to the IP host
vp900/901/..._last_router_delay	The delay from the last router to the IP host in the traceroute list from probing hosts "vp900/901/..."
vp900/901/..._total_delay	The total delay from probing hosts "vp900/901/..." to the IP host
longitude	The longitude of the IP (as label)
latitude	The latitude of the IP host (as label)

PS: The detailed columns and description of data.csv in other two datasets are similar to New York dataset's.