This article first appeared in Data Science Briefings, the DataMiningApps newsletter. Subscribe now for free if you want to be the first to receive our feature articles, or follow us @DataMiningApps. Do you also wish to contribute to Data Science Briefings? Shoot us an e-mail over at email@example.com and let’s get in touch!
In many real-world classification tasks such as churn prediction and fraud detection, we often encounter the class imbalance problem, which means one class is significantly outnumbered by the other class. The class imbalance problem brings great challenges to standard classification learning algorithms. Most of them tend to misclassify the minority instances more often than the majority instances on imbalanced data sets. For example, when a model is trained on a data set with 1% of instances from the minority class, a 99% accuracy rate can be achieved simply by classifying all instances as belonging to the majority class. Indeed, the problem of learning on imbalanced data sets is considered to be one of the ten challenging problems in data mining research.
In order to solve the problem of learning from imbalanced data sets, many solutions have been proposed in the past few years. Resampling approaches try to solve the problem by resampling the data and act as a preprocessing phase. Their usage is assumed to be independent of the classifier and can be applied to any learning algorithm. Hence, resampling solutions are very popular in practice. One important question when we use resampling is whether we actually need a perfectly balanced data set. Our research on churn prediction shows that a balanced sampling is unnecessary.
We used 11 real-world data sets from the telecommunication industry in our experiments. Seven sampling methods were considered, which include random over-sampling, random under-sampling SMOTE sampling and so on. We consider three different settings for the class ratios: 1:3, 2:3 and 1:1 (minority vs majority). Four benchmark classifiers are used in the experiments: logistic regression, C4.5 decision tree, support vector machine (SVM) and random forests (RF), which are widely used in churn prediction. The following table shows part of the results by using a 5 × 2 cross-validation experimental setup, where each entry represents the mean performance of each sampling rate across different classifiers and sampling methods. Beside the AUC measure, we also consider the maximum profit measure, which measures the profit produced by a retention campaign (Verbraken et al., 2013).
Table 1 Experimental results on imbalanced data set from telecom industry
As the table shows, the ratio of 1:3 is the best on two data sets and the ratio of 2:3 ranks the first on two data sets. The balanced class ratio never reaches the top position. The results definitely show there is no need to produce balanced data sets after sampling and the less balanced strategy (1:3) would be our recommendation due to its relatively good performance. The complete results and more discussions can be found in our recent paper “Benchmarking sampling techniques for imbalance learning in churn prediction” published in JORS.
- He, E. Garcia, Learning from imbalanced data, IEEE Transactions on Knowledge and Data Engineering, 2009, 21(9): 1263-1284.
- Verbeke, K. Dejaeger, D. Martens, J. Hur, B. Baesens, New insights into churn prediction in the telecommunication sector: A profit driven data mining approach, European Journal of Operational Research, 2012, 218(1): 211 -229.
- Verbraken, W. Verbeke, B. Baesens., A Novel Profit Maximizing Metric for Measuring Classification Performance of Customer Churn Prediction Models, IEEE Transactions on Knowledge and Data Engineering, Volume 25, Issue 5, pp. 961-973, 2013.
- Zhu, B. Baesens, A. Backiel, S. vanden Broucke. Benchmarking sampling techniques for imbalance learning in churn prediction, Journal of the Operational Society, 2017.