To SMOTE, or not to SMOTE?

Last update: Jan 03, 2022

Related tags

Deep Learning to-smote-or-not

Overview

To SMOTE, or not to SMOTE?

This package includes the code required to repeat the experiments in the paper and to analyze the results.

To SMOTE, or not to SMOTE?

Yotam Elor and Hadar Averbuch-Elor

Installation

# Create a new conda environment and activate it
conda create --name to-SMOTE-or-not -y python=3.7
conda activate to-SMOTE-or-not
# Install dependencies
pip install -r requirements.txt

Running experiments

The data is not included with this package. See an example of running a single experiment with a dataset from imblanaced-learn

# Load the data
import pandas as pd
import numpy as np
from imblearn.datasets import fetch_datasets
data = fetch_datasets()["mammography"]
x = pd.DataFrame(data["data"])
y = np.array(data["target"]).reshape((-1, 1))

# Run the experiment
from experiment import experiment
from classifiers import CLASSIFIER_HPS
from oversamplers import OVERSAMPLER_HPS
results = experiment(
    x=x,
    y=y,
    oversampler={
        "type": "smote",
        "ratio": 0.4,
        "params": OVERSAMPLER_HPS["smote"][0],
    },
    classifier={
        "type": "cat",  # Catboost
        "params": CLASSIFIER_HPS["cat"][0]
    },
    seed=0,
    normalize=False,
    clean_early_stopping=False,
    consistent=True,
    repeats=1
)

# Print the results nicely
import json
print(json.dumps(results, indent=4))

To run all the experiments in our study, wrap the above in loops, for example

for dataset in datasets:
    x, y = load_dataset(dataset)  # this functionality is not provided
    for seed in range(7):
        for classifier, classifier_hp_configs in CLASSIFIER_HPS.items():
            for classifier_hp in classifier_hp_configs:
                for oversampler, oversampler_hp_configs in OVERSAMPLER_HPS.items():
                    for oversampler_hp in oversampler_hp_configs:
                        for ratio in [0.1, 0.2, 0.3, 0.4, 0.5]:
                            results = experiment(
                                x=x,
                                y=y,
                                oversampler={
                                    "type": oversampler,
                                    "ratio": ratio,
                                    "params": oversampler_hp,
                                },
                                classifier={
                                    "type": classifier,
                                    "params": classifier_hp
                                },
                                seed=seed,
                                normalize=...,
                                clean_early_stopping=...,
                                consistent=...,
                                repeats=...
                            )

Analyze

Read the results from the compressed csv file. As the results file is large, it is tracked using git-lfs. You might need to download it manually or install git-lfs.

import os
import pandas as pd
data_path = os.path.join(os.path.dirname(__file__), "../data/results.gz")
df = pd.read_csv(data_path)

Drop nans and filter experiments with consistent classifiers, no normalization and a single validation fold

df = df.dropna()
df = df[
    (df["consistent"] == True)
    & (df["normalize"] == False)
    & (df["clean_early_stopping"] == False)
    & (df["repeats"] == 1)
]

Select the best HP configurations according to AUC validation scores. opt_metric is the key used to select the best configuration. For example, for a-priori HPs use opt_metric="test.roc_auc" and for validation-HPs use opt_metric="validation.roc_auc". Additionaly calculate average score and rank

from analyze import filter_optimal_hps
df = filter_optimal_hps(
    df, opt_metric="validation.roc_auc", output_metrics=["test.roc_auc"]
)
print(df)

Plot the results

from analyze import avg_plots
avg_plots(df, "test.roc_auc")

Citation

@misc{elor2022smote,
    title={To SMOTE, or not to SMOTE?}, 
    author={Yotam Elor and Hadar Averbuch-Elor},
    year={2022},
    eprint={2201.08528},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Security

See CONTRIBUTING for more information.

License

This library is licensed under the MIT-0 License. See the LICENSE file.

To SMOTE, or not to SMOTE?

Related tags

Overview

To SMOTE, or not to SMOTE?

Installation

Running experiments

Analyze

Citation

Security

License

Owner

Amazon Web Services

Wordle-solver - Wordle answer generation program in python

This repo contains research materials released by members of the Google Brain team in Tokyo.

Portfolio Optimization and Quantitative Strategic Asset Allocation in Python

Multi-Scale Aligned Distillation for Low-Resolution Detection (CVPR2021)

Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation [3DV 2021 Oral]

PyTorch implementation of "A Simple Baseline for Low-Budget Active Learning".

[ICCV 2021] Focal Frequency Loss for Image Reconstruction and Synthesis

Deep learning model for EEG artifact removal

🕹️ Official Implementation of Conditional Motion In-betweening (CMIB) 🏃

IAST: Instance Adaptive Self-training for Unsupervised Domain Adaptation (ECCV 2020)

Implementation of GGB color space

A universal framework for learning timestamp-level representations of time series

Do you like Quick, Draw? Well what if you could train/predict doodles drawn inside Streamlit? Also draws lines, circles and boxes over background images for annotation.

This repo contains source code and materials for the TEmporally COherent GAN SIGGRAPH project.

The first dataset on shadow generation for the foreground object in real-world scenes.

Full Transformer Framework for Robust Point Cloud Registration with Deep Information Interaction

A weakly-supervised scene graph generation codebase. The implementation of our CVPR2021 paper ``Linguistic Structures as Weak Supervision for Visual Scene Graph Generation''

Distributed Arcface Training in Pytorch

Image-to-Image Translation with Conditional Adversarial Networks (Pix2pix) implementation in keras

Official implementation of the ICCV 2021 paper: "The Power of Points for Modeling Humans in Clothing".