You work for a major banking firm as a machine learning specialist.
As part of the bank's fraud detection team, you build a machine learning model to detect fraudulent transactions.
Using your training dataset, you have produced a Receiver Operating Characteristic (ROC) curve, and it shows 99.99% accuracy.Your transaction dataset is very large, but 99.99% of the observations in your dataset represent non-fraudulent transactions.
Therefore, the fraudulent observations are a minority class.
Your dataset is very imbalanced. You have the approval from your management team to produce the most accurate model possible, even if it means spending more time perfecting the model.
What is the most effective technique to address the imbalance in your dataset?
Click on the arrows to vote for the correct answer
A. B. C. D.Answer: C.
Option A is incorrect.
The SMOTE technique creates new observations of the underrepresented class, in this case, the fraudulent observations.
These synthetic observations are almost identical to the original fraudulent observations.
This technique is expeditious, but the types of synthetic observations it produces are not as useful as the unique observations created by other oversampling techniques.
Option B is incorrect.
Random oversampling uses copies of some of the minority class observations (randomly selected) to augment the minority class observation set.
These observations are exact replicas of existing minority class observations, making them less effective than observations created by other techniques that produce unique synthetic observations.
Option C is correct.
The Generative Adversarial Networks (GANs) technique generates unique observations that more closely resemble the real minority observations without being so similar that they are almost identical.
This results in more unique observations of your minority class that improve your model's accuracy by helping to correct the imbalance in your data.
Option D is incorrect.
Using an undersampling technique would remove potentially useful majority class observations.
Additionally, you would have to remove a huge number of your majority class observations to correct your imbalance that you would render your entire training dataset useless.
Reference:
Please see the Wikipedia article titled Oversampling and undersampling in data analysis, and the article titled Imbalanced data and credit card fraud.
The most effective technique to address the imbalance in the dataset in this scenario is to use Synthetic Minority Oversampling Technique (SMOTE) oversampling.
Explanation: In this scenario, the dataset is imbalanced, and fraudulent transactions are a minority class. Therefore, the standard machine learning algorithms tend to classify most of the observations as non-fraudulent transactions, which results in a high accuracy rate, but it is not effective in identifying fraudulent transactions.
To address this issue, one of the most effective techniques is to use oversampling, which creates synthetic examples of the minority class to balance the dataset. SMOTE is a popular oversampling technique that generates synthetic examples by interpolating between existing minority class observations. It creates new minority class examples by randomly choosing an observation and its k nearest neighbors and interpolating between them.
Random oversampling is another oversampling technique that duplicates the minority class observations to balance the dataset. However, it can result in overfitting and does not consider the distribution of the minority class observations.
Generative Adversarial Networks (GANs) oversampling is a newer technique that uses a neural network to generate synthetic examples of the minority class. However, it requires more computation and may not always result in better performance than other oversampling techniques.
Edited Nearest Neighbor undersampling is a technique that removes some observations from the majority class to balance the dataset. However, it may result in the loss of important information from the majority class observations.
Therefore, SMOTE oversampling is the most effective technique in this scenario as it balances the dataset by creating synthetic examples of the minority class without losing any information from the majority class observations.