Journal of Operational Risk

Risk.net

Imbalanced data issues in machine learning classifiers: a case study

Mingxing Gong

  • Machine learning classifiers are widely used in financial applications. Due to the nature of certain classification problems, special care should be taken when dealing with imbalanced data.
  • Resampling is a common technique in building machine learning models to address imbalanced data issue. This study shows that balanced data does not necessarily result in the best model performance.
  • The degree to which different algorithms are impacted by class imbalance is not equal. It is demonstrated that certain classification algorithms, especially tree-based algorithms perform very well in the presence of imbalance, while other algorithms such as neural network and logistic regression perform very poorly.

Machine learning classifiers are widely used in financial applications. Due to the nature of certain classification problems, special care should be taken when dealing with imbalanced data. In practice, many model developers and validators fail to take this into account in their model development and validation. In addition, resampling is a common technique to address imbalanced data issues when building traditional logistic regression models. However, there has been no specific discussion regarding the resampling ratio used to rebalance the data or how the issue of imbalance impacts different kinds of machine learning classifiers, especially the more advanced ones. This paper aims to outline the special characteristics of the classifiers, compare different methods in dealing with imbalanced data issues and provide best practice in model development, evaluation and validation to avoid common pitfalls. Although the methods discussed in this paper can apply to general machine learning classifiers in applications with imbalanced data issues, by using a case study in credit card fraud detection this paper calls practitioners’ attention to the imbalanced data problems therein, where class imbalance is often mistreated and lacks theoretical discussion.

Sorry, our subscription options are not loading right now

Please try again later. Get in touch with our customer services team if this issue persists.

New to Risk.net? View our subscription options

If you already have an account, please sign in here.

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here: