Imbalanced data stream classification with the changing prior probabilities

Implementation

Setup - experiment real data

Comparison between HDWE method and selected state-of-the-art method with base classifiers SVC and HDDT.

Data

Number of data streams: 3

• Poker data set
• Cover type data set
• Insects data set

Methods

1. HDWE
2. SEA
3. AWE
4. Learn++.CDS
5. Learn++.NIE
6. OUSE
7. REA

Setup - experiment 3a

Comparison between HDWE method and selected state-of-the-art method with base classifiers SVC.

Data

In general, numbers of data streams: 84

1. Generated from stream-learn

   • Number of class: 2
   • Number of concept drifts: 5
   • Types of concept drifts: sudden, incremental
   • Stationary imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Dynamically imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Number of samples: 10000 (where number of chunks: 200 and chunk size: 500)
   • Number of features: 20 (where informative: 15 and redundant: 5)
   • Random state: 1111, 1234, 1567

Methods

1. HDWE
2. SEA
3. AWE
4. Learn++.CDS
5. Learn++.NIE
6. OUSE
7. REA

Setup - experiment 3b

Comparison between HDWE method and selected state-of-the-art method with base classifiers HDDT.

Data

In general, numbers of data streams: 84

1. Generated from stream-learn

   • Number of class: 2
   • Number of concept drifts: 5
   • Types of concept drifts: sudden, incremental
   • Stationary imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Dynamically imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Number of samples: 10000 (where number of chunks: 200 and chunk size: 500)
   • Number of features: 20 (where informative: 15 and redundant: 5)
   • Random state: 1111, 1234, 1567

Methods

1. HDWE
2. SEA
3. AWE
4. Learn++.CDS
5. Learn++.NIE
6. OUSE
7. REA

Setup - experiment 2

Different base classifier in the HDWE (Hellinger Distance Weighted Ensemble) method. Check influence of the type of base classifier on the ensemble.

Data

In general, numbers of data streams: 84

1. Generated from stream-learn

   • Number of class: 2
   • Number of concept drifts: 5
   • Types of concept drifts: sudden, incremental
   • Stationary imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Dynamically imbalance ratio: 1%, 3%, 5%, 10%, 15%, 20%, 25%
   • Number of samples: 10000 (where number of chunks: 200 and chunk size: 500)
   • Number of features: 20 (where informative: 15 and redundant: 5)
   • Random state: 1111, 1234, 1567

Methods

1. HDWE(GNB)
2. HDWE(MLP)
3. HDWE(CART)
4. HDWE(HDDT)
5. HDWE(KNN)
6. HDWE(SVC)

Evaluation

1. Test Then Train

Metrics

1. Specificity
2. Recall
3. Precision
4. F1-score
5. Balanced accuracy score
6. Geometric-mean

Setup - experiment 1

Compare 2 methods HDWE and AWE with different imbalance ratio.

Data

1. Generated from stream-learn

   • Number of class: 2
   • Number of streams: 10 (random state has changed in range 1000 - 1550)
   • Number of concept drifts: 5
   • Types of concept drifts: sudden, incremental
   • Stationary imbalance ratio: 10%, 20%, 30%, 40%, 50%
   • Dynamically imbalance ratio: 10%, 20%, 30%, 40%
   • Number of samples: 10000 (where number of chunks: 200 and chunk size: 500)
   • Number of features: 20 (where informative: 15 and redundant: 5)

Methods

1. HDWE (Hellinger Distance Weighted Ensemble) - own contribution
2. HDDT (Hellinger Distance Decision Tree) - own implementation
3. AWE (Accuracy-Weighted Ensemble)
4. Gaussian Naive Bayes - base classifier

Evaluation

1. Test Then Train

Metrics

1. Balanced accuracy score
2. F1 score
3. Recall
4. Specificity