transcendent-multiclass-CDD__Wdis 1.1.1

Transcendent adaptation for multiclass problems

Transcendent Multiclass

This repository enables users to apply Transcendent-like concept drift detection to both binary and multiclass problems.

Modifications have been made specifically to the ICE (Inductive Conformal Evaluator) implementation, while the other solutions (i.e. TCE, CCE, etc.) are out of the scope.

This project adapts Transcendent for multiclass problems by implementing two Nonconformity Measures (NCM) for Random forest and LightGBM classifiers.

Prerequisites

• Setup the train/test split directory, which should contains the following files:

  time_split/
  ├── X_train.pkl
  ├── X_test.pkl
  ├── X_proper_train.pkl
  ├── X_cal.pkl
  ├── y_train.pkl
  ├── y_test.pkl
  ├── y_proper_train.pkl
  └── y_cal.pkl
  

• Make sure to have a running and active version of Docker.

Usage:

1. Clone the repository and change directory:

   git clone [e-mail]:w-disaster/transcendent-multiclass.git && cd transcendent-multiclass
   

2. Configure the env variables and Run Inductive Conformal Evaluator:

   PE_DATASET_NAME=<YOUR_PE_DATASET_NAME>
   SPLITTED_MPH_DATASET_PATH=<YOUR_PRE_SPLITTED_DATA>
   BEST_HYP_DIR=<YOUR_BEST_HYP_DIR> # Based on format produced by overfitting-analysis
   
   docker run -d \
   --name mph-feature-extraction-$PE_DATASET_NAME \
   -e BASE_DATASET_PATH=/usr/app/dataset/ \
   -e PE_DATASET_TYPE=${PE_DATASET_NAME}_mph \
   -e SPLITTED_MPH_DATASET_PATH=/usr/input_data/splitted_dataset/ \
   -e BEST_HYP_DIR=/usr/input_data/best_hyp/ \
   -e FEATURE_TYPE=dts \
   -v $BEST_HYP_DIR:/usr/input_data/best_hyp/ \
   -v $SPLITTED_MPH_DATASET_PATH:/usr/input_data/splitted_dataset/ \
   -v ./results_multiclass/:/usr/app/models/ \
   ghcr.io/malware-concept-drift-detection/transcendent-multiclass:main
   

   A results_multiclass/ directory will be locally created containing the credibility ($p$-values) and confidence scores for both calibration and testing sets.

3. Analysis post ICE:

   Check whether novel families in the testing set produce smaller $p$-values, and thus can be discriminated from seen families.