Self-paced Ensemble (SPE) is a general learning framework for massive highly imbalanced classification.
This is the implementation for the paper "Self-paced Ensemble for Highly Imbalanced Massive Data Classification" (ICDE 2020 Research Paper). If you find this repository useful, please cite our work:
@inproceedings{
liu2020self-paced-ensemble,
title={Self-paced Ensemble for Highly Imbalanced Massive Data Classification},
author={Liu, Zhining and Cao, Wei and Gao, Zhifeng and Bian, Jiang and Chen, Hechang and Chang, Yi and Liu, Tie-Yan},
booktitle={2020 IEEE 36th International Conference on Data Engineering (ICDE)},
year={2020},
organization={IEEE}
}
This repository contains:
- Implementation of Self-paced Ensemble
- Implementation of 5 ensemble-based imbalance learning baselines
SMOTEBoost[1]SMOTEBagging[2]RUSBoost[3]UnderBagging[4]BalanceCascade[5]
- Implementation of 15 resampling based imbalance learning baselines
- Additional experimental results
NOTE: The implementations of [1],[3] and resampling methods are based on imbalanced-algorithms and imbalanced-learn.
SPE performs strictly balanced under-sampling in each iteration and is therefore very computationally efficient. In addition, SPE does not rely on calculating the distance between samples to perform resampling. It can be easily applied to datasets that lack well-defined distance metrics (e.g. with categorical features / missing values) without any modification. Moreover, as a generic ensemble framework, our methods can be easily adapted to most of the existing learning methods (e.g., C4.5, SVM, GBDT, and Neural Network) to boost their performance on imbalanced data. Compared to existing imbalance learning methods, SPE works particularly well on datasets that are large-scale, noisy, and highly imbalanced (e.g. with imbalance ratio greater than 100:1). Such kind of data widely exists in real-world industrial applications. The figure below gives an overview of the SPE framework.
Our SPE implementation requires following dependencies:
- python (>=3.5)
- numpy (>=1.11)
- scipy (>=0.17)
- scikit-learn (>=0.21)
- imblearn (>=0.0) (optional, for canonical resampling)
Currently you can install SPE by clone this repository. We'll release SPE on the PyPI in the future.
git clone https://github.com/ZhiningLiu1998/self-paced-ensemble.git
Our SPE implementation can be used much in the same way as the ensemble classifiers in sklearn.ensemble.
| Parameters | Description |
|---|---|
base_estimator |
object, optional (default=sklearn.tree.DecisionTreeClassifier()) The base estimator to fit on self-paced under-sampled subsets of the dataset. NO need to support sample weighting. Built-in fit(), predict(), predict_proba() methods are required. |
hardness_func |
function, optional (default=lambda y_true, y_pred: np.absolute(y_true-y_pred)) User-specified classification hardness function. Input: y_true and y_pred Output: hardness (1-d array) |
n_estimator |
integer, optional (default=10) The number of base estimators in the ensemble. |
k_bins |
integer, optional (default=10) The number of hardness bins that were used to approximate hardness distribution. |
random_state |
integer / RandomState instance / None, optional (default=None) If integer, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by numpy.random. |
| Methods | Description |
|---|---|
fit(self, X, y, label_maj=0, label_min=1) |
Build a self-paced ensemble of estimators from the training set (X, y). label_maj/label_min specify the label of majority/minority class. By default, we let the minority class be positive class ( label_min=1). |
predict(self, X) |
Predict class for X. |
predict_proba(self, X) |
Predict class probabilities for X. |
predict_log_proba(self, X) |
Predict class log-probabilities for X. |
score(self, X, y) |
Returns the average precision score on the given test data and labels. |
| Attributes | Description |
|---|---|
base_estimator_ |
estimator The base estimator from which the ensemble is grown. |
estimators_ |
list of estimator The collection of fitted base estimators. |
A minimal example
X, y = <data_loader>.load_data()
spe = SelfPacedEnsemble().fit(X, y)A non-minimal working example (It demonstrates some of the features of SPE)
import numpy as np
from sklearn import datasets
from sklearn.tree import DecisionTreeClassifier
from self_paced_ensemble import SelfPacedEnsemble
from utils import (make_binary_classification_target, imbalance_train_test_split)
X, y = datasets.fetch_covtype(return_X_y=True)
y = make_binary_classification_target(y, pos_label=7, verbose=True)
X_train, X_test, y_train, y_test = imbalance_train_test_split(X, y, test_size=0.2)
def absolute_error(y_true, y_pred):
"""Self-defined classification hardness function"""
return np.absolute(y_true - y_pred)
spe = SelfPacedEnsemble(
base_estimator=DecisionTreeClassifier(),
hardness_func=absolute_error,
n_estimators=10,
).fit(X_train, y_train)
print('auc_prc_score: {}'.format(spe.score(X_test, y_test)))We also provide a simple framework (run_example.py) for conveniently comparing the performance of our method and other baselines. It is also a more complex example of how to use our implementation of ensemble methods to perform classification. To use it, simply run:
python run_example.py --method=SPEnsemble --n_estimators=10 --runs=10
You should expect output console log like this:
Running method: SPEnsemble - 10 estimators in 10 independent run(s) ...
100%|█████████████████████████████████████████| 10/10 [00:14<00:00, 1.42s/it]]
ave_run_time: 0.686s
------------------------------
Metrics:
AUCPRC mean:0.910 std:0.009
F1 mean:0.872 std:0.006
G-mean mean:0.873 std:0.007
MCC mean:0.868 std:0.007
| Arguments | Description |
|---|---|
--method |
string, optional (default='SPEnsemble') support: SPEnsemble, SMOTEBoost, SMOTEBagging, RUSBoost, UnderBagging, Cascade, all When all, the script will run all supported methods. |
--n_estimators |
integer, optional (default=10) The number of base estimators in the ensemble. |
--runs |
integer, optional (default=10) The number of independent runs for evaluating method performance. |
We introduce seven small datasets to validate our method SPE. Their properties vary widely from one another, with IR ranging from 9.1:1 to 111:1, dataset sizes ranging from 360 to 145,751, and feature numbers ranging from 6 to 617. See the table below for more information about these datasets.
SPE was compared with other 5 ensemble-based imbalance learning methods:
Over-sampling-based ensemble: SMOTEBoost, SMOTEBagging
Under-sampling-based ensemble: RUSBoost, UnderBagging, Cascade
We use Decision Tree as the base classifier for all ensemble methods as other classifiers such as KNN do not support Boosting-based methods. We implemented SPE with Absolute Error as the hardness function and set k=10. In each dataset, 80% samples were used for training. The rest 20% was used as the test set. All the experimental results were reported on the test set (mean and standard deviation of 50 independent runs with different random seeds for training base classifiers).
Dataset links: Credit Fraud, KDDCUP, Record Linkage, Payment Simulation.
Comparisons of SPE with traditional resampling/ensemble methods in terms of performance & computational efficiency.
[1] N. V. Chawla, A. Lazarevic, L. O. Hall, and K. W. Bowyer, “Smoteboost: Improving prediction of the minority class in boosting,” in European conference on principles of data mining and knowledge discovery. Springer, 2003, pp. 107–119
[2] S. Wang and X. Yao, “Diversity analysis on imbalanced data sets by using ensemble models,” in 2009 IEEE Symposium on Computational Intelligence and Data Mining. IEEE, 2009, pp. 324–331.
[3] C. Seiffert, T. M. Khoshgoftaar, J. Van Hulse, and A. Napolitano, “Rusboost: A hybrid approach to alleviating class imbalance,” IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 40, no. 1, pp. 185–197, 2010.
[4] R. Barandela, R. M. Valdovinos, and J. S. Sanchez, “New applications´ of ensembles of classifiers,” Pattern Analysis & Applications, vol. 6, no. 3, pp. 245–256, 2003.
[5] X.-Y. Liu, J. Wu, and Z.-H. Zhou, “Exploratory undersampling for class-imbalance learning,” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 39, no. 2, pp. 539–550, 2009.