mogicianxd/fmrt

A pytorch implementation of FMRT (Certifiable Robustness to Discrete Adversarial Perturbations for Factorization Machines) [SIGIR20]

FM-RT

An unofficial pytorch implementation of FM-RT:

Yang Liu, Xianzhuo Xia, Liang Chen, Xiangnan He, Carl Yang, Zibin Zheng. Certifiable Robustness to Discrete Adversarial Perturbations for Factorization Machines, in SIGIR 2020

Note

The paper used a FM as:

FMRT2.py is the corresponding implementation.

However, I cannot see similar metrics. Before FMRT, the avg-max $q$ is only 1.12, while the paper showed 1.60.

So I followed the formula Rendle adopted, which is

Then I use FMRT.py by default. It dismissed elements on the diagonal.

Here's the performance comparison:

Files in the folder

• src/
  • BaseModel.py: The base class encapsulates device initialization, train & test tasks for rank, classification and regression.
  • FMRT.py: Derived from BaseModel. Include training and evaling methods.
  • dataset.py: Derived from torch.util.data.Dataset
    • FeatureDataset: Dataset for FM.
• main.py: You can look for descriptions of args with -h
• ../benchmarks/datasets/

Acknowledgements

For FM, We use torchfm with few changes.

fmrt's Issues

will it work for third order categorical features interaction

Great code, thanks !

Plese help to understand
1
will it work for third order categorical features interaction ?
2
will it run on Windows computer ?

3
will it work for sparse data ?

reshape data in dataset.py

Hi,

in line 10 of dataset.py file, it reshapes the feature to (1+neg_num, feature.shape[-1]).
What is neg_num? and why there is a need to reshape feature?

data format

could you provide the data format?