snudatalab / info

This is a repository of "Fast and accurate interpretation of workload classification", submitted on PLOS ONE 2023.

Fast and accurate interpretation of workload classification

These are codes for "Fast and accurate interpretation of workload classification", published in Plos One 2023.

Dataset

We use a real-world workload dataset for classification. The detailed information about dataset is summarized in the table below.

dataset	# of classes	input size	# of train	# of test of known
SEC-seq	40	1735	586,885	293,444
Memtest86-seq	31	1599	433,334	216,696

We include the preprocessed data of Memtest86-seq dataset in this repository. Due to the size limit, we upload final_data folder on google drive. You can download the data from [link]. final_data folder contains the entire feature vector for workload classification. sequences folder includes n-gram sequences selected as CMD features. The detailed structures of these directories are as follows:

current directory
├── final_data
│   ├── 7-grams
│   ├── 11-grams
│   ├── 15-grams
│   ├── bank_access_counts
│   ├── data_split_ids
│   ├── test_ngram_zeros.npy
│   └── row_col_address_access_counts
└── sequences
    ├── 7-grams.top25_osr.pkl
    ├── 11-grams.top25_osr.pkl
    └── 15-grams.top25_osr.pkl

Model

We provide a pre-trained MLP model for workload classification in models directory.

• model_mlp.pt: a trained 2-layer MLP model

Code Description

All codes in this repository are implemented based on Python 3.7. This repository contains the code for INFO, INterpretable model For wOrkload classification.

• The information of codes implemented in this repository.
  • main.py: train and evaluate an interpretable model.
  • cluster.py: cluster features for workload classification to generate super features for interpretation.
  • train.py: train and evaluate a workload classification model.
  • model.py: implement MLP for classification.
  • dataloader.py: load workload sequence data and extract feature vectors for classification.

To interpret the classification results, you have to type the following command.

python main.py

You can train a workload classification model through the command below.

python train.py

These commands run the codes in the default setting.

Citation

@article{shim2023fast,
  title={Fast and accurate interpretation of workload classification model},
  author={Shim, Sooyeon and Kim, Doyeon and Jang, Jun-Gi and Chae, Suhyun and Lee, Jeeyong and Kang, U},
  journal={Plos one},
  volume={18},
  number={3},
  pages={e0282595},
  year={2023},
  publisher={Public Library of Science San Francisco, CA USA}
}