This article has Open Peer Review reports available.
Learning to improve medical decision making from imbalanced data without a priori cost
© Wan et al.; licensee BioMed Central Ltd. 2014
Received: 1 August 2014
Accepted: 11 November 2014
Published: 5 December 2014
In a medical data set, data are commonly composed of a minority (positive or abnormal) group and a majority (negative or normal) group and the cost of misclassifying a minority sample as a majority sample is highly expensive. This is the so-called imbalanced classification problem. The traditional classification functions can be seriously affected by the skewed class distribution in the data. To deal with this problem, people often use a priori cost to adjust the learning process in the pursuit of optimal classification function. However, this priori cost is often unknown and hard to estimate in medical decision making.
In this paper, we propose a new learning method, named RankCost, to classify imbalanced medical data without using a priori cost. Instead of focusing on improving the class-prediction accuracy, RankCost is to maximize the difference between the minority class and the majority class by using a scoring function, which translates the imbalanced classification problem into a partial ranking problem. The scoring function is learned via a non-parametric boosting algorithm.
We compare RankCost to several representative approaches on four medical data sets varying in size, imbalanced ratio, and dimension. The experimental results demonstrate that unlike the currently available methods that often perform unevenly with different priori costs, RankCost shows comparable performance in a consistent manner.
It is a challenging task to learn an effective classification model based on imbalanced data in medical data analysis. The traditional approaches often use a priori cost to adjust the learning of the classification function. This work presents a novel approach, namely RankCost, for learning from medical imbalanced data sets without using a priori cost. The experimental results indicate that RankCost performs very well in imbalanced data classification and can be a useful method in real-world applications of medical decision making.
One of the challenging issues in medical data analysis is caused by the highly skewed proportion of differen sample types . This often happens when one class of samples (positive or abnormal) is of limited size and sometimes difficult to collect while the other class (negative or normal) is much more abundant and much easier to find. Learning an effective classification model can be a difficult task if the data used to train the model are imbalanced. When samples of the majority class greatly outnumber samples of the minority class, the traditional classification models usually have a bias in favor of the majority class. This is because the goal of traditional classification modeling is to construct a function (or a classifier) based on the properties of training data so as to make as few errors as possible when being used to predict the class membership of new samples . A range of classification methods, such as decision tree, neural network, nearest neighbor, logistic regression, and support vector machine, have been well developed. These methods, when applied to imbalanced medical data, will often produce high predictive accuracy over the majority class, but poor predictive accuracy over the minority class. Besides the medical data analysis, there are many other real world applications involving learning from imbalanced data, such as text classification ,, the fraudulent telephone call detection ,, oil spill detection , potential buyer selection in direct marketing , and etc. Nevertheless, the impact of this issue is particularly tremendous in medical data analysis because the cost of misclassifying a minority sample as a majority sample, e.g., patients miss the chance to be cured if they fail to be identified and diagnosed due to the wrong classification, is highly expensive and sometimes unaffordable.
There are three major approaches to dealing with imbalanced data sets, which are sampling, cost-sensitive learning, and boosting. The sampling approach is applied to create a more balanced class distribution in the training data by either over-sampling the minority class or under-sampling the majority class ,-. Both over-sampling and under-sampling have their benefits and drawbacks. They can be easily implemented and applied to all application domains with imbalanced data. But the classification performance can be very sensitive to the class ratio of the training data. One major drawback associated with over-sampling is that learning on duplicated samples can lead to overfitting . On the other hand, under-sampling may result in the loss of information that comes with deleting samples .
While sampling approaches address the imbalanced learning problem at the data level, cost-sensitive learning methods target this problem at both the data level and the algorithm level . Instead of creating balanced data distributions through different sampling strategies, cost-sensitive learning uses a cost matrix that describes the costs for misclassifying data samples -. The cost matrix encodes the penalty of misclassifying samples from one class as another. Some research works have provided the theoretical foundations of cost-sensitive methods in imbalanced learning problems , and various empirical studies have shown that cost-sensitive methods are superior to sampling methods in many application domains ,. However, there is one major disadvantage of using cost-sensitive learning to handle the imbalanced medical data. It is that misclassification costs are often unknown and hard to estimate in medical decision making and the performance of cost-sensitive learning is very sensitive to different misclassification costs .
In contrast to sampling methods and cost-sensitive methods that are specially designed to address imbalanced learning problem, boosting is an off-the-shelf approach that is particularly effective in handling imbalanced data. The most common boosting algorithm is AdaBoost , which iteratively builds an ensemble of models with weighted samples. During each iteration, incorrectly classified samples are given high weights so that they will have high chance to be correctly classified in the next iteration. In the imbalanced classification, it is most likely that the minority class samples are misclassified at the beginning and naturally given higher weights in subsequent iterations. AdaBoost is particularly suitable for medical decision making since it does not require a priori cost. However, AdaBoost is still an accuracy-oriented algorithm and its learning process may still bias toward the majority class because samples in the majority class contribute more to the overall classification accuracy. As a result, the empirical study  shows that cost-sensitive methods outperform AdaBoost.
In this work, we present a novel boosting algorithm for the classification of imbalanced data. Instead of focusing on improving the class-prediction accuracy, our approach is to maximize the difference between the minority class and the majority class by using a scoring function. Intuitively, the basic idea is to translate the imbalanced classification problem into a partial ranking problem. In this partial ranking problem, we shall find a scoring (or ranking) function that can assign samples in the minority class higher scores than samples in the majority class or vice versa. Therefore, the target of our approach is to infer the pairwise relationship between samples in two classes. Compared to the cost-sensitive learning that explicitly uses cost matrix to learn a biased classifier toward the minority class, our method naturally embeds the importance of identifying minority samples in the new formulation and the relative importance between two classes is automatically learned from the data without using any priori knowledge.
where f(x) is a linear combination of multiple classifiers . The loss function measures the difference between estimated and true values for an instance of data. To minimize this loss function, AdaBoost iteratively builds an additive model with an ensemble of classifiers where subsequent classifiers are learned in favor of those instances misclassified by previous classifiers. Therefore, in AdaBoost, the samples in the minority class that are often misclassified at start will be given higher weights in subsequent classifiers and then have higher chance to be correctly classified. Nevertheless, the loss function in Eq. (1) is defined on the overall prediction accuracy. Thus AdaBoost may still favor the majority class as it has higher impact in the loss function. Some cost-sensitive learning methods, such as AdaC1, AdaC2, AdaC3 , and AdaCost , extend AdaBoost with the pre-specified cost matrix, which gives high penalization to the misclassification of the samples in the minority class. But as we mentioned above, the misclassification cost is often unknown.
where τ is a scalar that is used to avoid the trivial solutions (making F as a constant). We may choose the absolute function instead of the square function but the absolute function is not continuous at changing point, which complicates the optimization process. Our goal is to find a function F that minimizes .
Because m a x(0,.)≥0, the square of m a x(0,.) is non-decreasing. Because is an affine function of F and m a x(0,.) is pointwise supremum (maximum), is convex. Therefore, is convex.
We choose the regression tree as the base function to fit the negative gradient and with respect to x k and , respectively. If the learned regression tree closely matches the target value, adding it with a multiplier ρ to the additive model will decrease the loss. The whole gradient boosting procedure for learning the function F(x) is described as follows:
We name our method “RankCost” as the goal of this method is to find a partial ranking function F to replace the predefined cost matrix to solve imbalanced classification problem. To evaluate the performance of RankCost in medical decision making, we compare it with AdaBoost , AdaCost , and Cost-sensitive decision tree .
Data set summary
Breast cancer data
The breast cancer data was released by the University Medical Centre, Institute of Oncology, Ljubljana, Yugoslavia. Each instance is described by 9 attributes, 3 of which are linear and 6 are nominal. There are 286 instances in this data set, 9 instances with missing values. Class distributions are 29.7% of recurrence-events (positive class) and 70.3% of no-recurrence-events (negative class).
The second data is from a study of hepatitis, which includes only 155 instances in the whole data set. Each instance is described by 19 attributes with only one being continuously valued. The data set is composed of 32 positive instances (20.65%) in class “DIE” and 123 negative instances (79.35%) in class “LIVE”.
The third data set is from a study of diabetes in Pima Indian population. Each sample is described by 8 continuously valued attributes. 268 samples were identified as positive and the other 500 samples were identified as negative. The two classes are non-evenly distributed with 34.9% of positive instances and 65.1% of negative instances, respectively.
Sick Euthyroid data
The fourth data set is from a study of euthyroid sick. The data were collected with 25 attributes, 7 being continuous and 18 being Boolean values. The data set contains 3,163 instances, with 9.26% of the instances being euthyroid and the remaining 90.74% being negative. There are several instances with missing attribute values.
Confusion matrix table
Predicted as positive
Predicted as negative
True positive (TP)
False negative (FN)
False negative (FP)
True negative (TN)
where β corresponds to the relative importance of TPR versus PPV and it is typically set to 1. The F-measure incorporates TPR and PPV into a single number. It basically represents a harmonic mean between them. It follows that the F-measure is high when both TPR and PPV are high . This indicates that F-measure is able to evaluate the performance of a learning algorithm on the class of our interest.
To evaluate our proposed method RankCost, we specially select three well-known methods to compare, which are AdaBoost , AdaCost , and Cost-sensitive decision tree . AdaBoost is chosen for the reason that it also does not require a priori cost in handling imbalanced data classification. AdaCost is a cost-sensitive variant of AdaBoost, which requires a priori cost to adjust the weights of samples in different classes. Cost-sensitive decision tree is a popular cost-sensitive classifier for imbalanced classification problems. However, as we mentioned above, the misclassification costs are often unknown in medical decision making and the performances of cost-sensitive classifiers may vary significantly to different misclassification costs. Therefore, in our experiments, we first test AdaCost and Cost-sensitive decision tree on various cost settings and then choose the cost settings with which AdaCOST and Cost-sensitive decision tree can achieve the best performance.
All experiments are performed by following the standard practice of 10-fold cross validation. Each data set is split into ten disjoint subsets using random sampling. Nine of them are used to train the model and the remaining one is used to test the model. This procedure is repeated 10 times so that each partition is used as the test data once. All four methods use exactly the same ten testing and validation data sets, each of which is 10% of the entire data. The results for each method are the average of the 10-fold cross-validation. Regarding the cross validation in our experiments, not only is the coefficient (or weight) of each predictor cross-validated, but also the selection of the predictors is also cross-validated. The cost settings for AdaCost and Cost-sensitive decision tree is chosen from the set [ 1.0:0.1,1.0:0.2,1.0:0.3,1.0:0.4,1.0:0.5,1.0:0.6,1.0:0.7,1.0:0.8,1.0:0.9]. The cost of misclassifying a minority sample as a majority sample is always set 1.0. The cost of misclassifying a majority sample as a minority sample is set from 0.1 to 0.9.
Performance comparison among AdaBoost, Cost-sensitive decision tree, AdaCost, and our method RankCost with respect to F-measure (F), TPR (R), and PPV (P) values and and their 95 % confidence intervals
0.494 [0.376, 0.612]
0.388 [0.104, 0.672]
0.906 [0.603, 1.209]
0.494 [0.270, 0.718]
0.579 [0.152, 1.000]
0.337 [0.263, 0.411]
0.347 [0.211, 0.483]
0.494 [0.392, 0.596]
0.603 [0.261, 0.945]
0.469 [-.215, 1.153]
0.688 [0.212, 1.164]
0.561 [0.159, 0.963]
In our experiments, we observe that the results on hepatitis data show high variance. The main reason is due to the number of attributes. There are 19 attributes in the hepatitis data, which requires a much large data set in order to train a reliable and consistent model across the multiple runs of validation. However, we only have 155 samples in total. In such a situation, the literature suggested evaluation method is the leave-one-out cross-validation, in which the test data only contains one sample and all the others are used in the training. The number of runs (or folds) is equal to the number of samples. However, in the evaluation using hepatitis data, adding a few more samples in the training data is still far from enough to train a stable model. Furthermore, there are some critical issue in leave-one-out cross-validation. Besides the low efficiency. The major one is that each run is highly correlated with the others. That correlation may lead to the significant underestimation of the variance when the trained model is applied to new data because most of the trained models in leave-one-out evaluation will be nearly identical. Therefore, the trained model from the leave-one-out cross validation is very prone to over-fitting. Taking all these issues into consideration, we eventually choose the most popular one, which is 10 cross-validation.
Convergence of RankCost
Discussion and conclusions
In medical data analysis, it often happens that data are composed of a minority (positive or abnormal) group and a majority (negative or normal) group and the cost of misclassifying a minority sample as a majority sample is highly expensive. It is a challenging task to learn an effective classification model based on imbalanced data. The traditional approaches often use a priori cost to adjust the learning process in the pursuit of optimal classification function. However, this priori cost is often unknown and hard to estimate in medical decision making. This work presents a novel approach, namely RankCost, for learning from medical imbalanced data sets without using a priori cost. In RankCost, the traditional imbalanced classification problem is reformulated into a partial ranking problem. Instead of focusing on the class prediction accuracy, RankCost is to learn a non-parametric scoring function which can maximize the difference between the minority class and the majority class. The boosting technique is adopted in RankCost to learn the scoring function, and the relative importance of the minority class over the majority class is naturally reflected in the learning process. The performance of RankCost is illustrated by tests on four medical data sets varying in size, dimension, and imbalanced ratio. The experimental results obtained indicate that our approach achieves comparable performance against two cost-sensitive methods and outperforms the non-cost-sensitive method AdaBoost. Importantly, our approach does not require any priori knowledge, which makes our method more practical in medical decision making.
There are some limitations in our works. First, our approach does sacrifice the performance of the majority class for the minority class since it only aims to improve the prediction accuracy of the minority class. In medical decision making, misclassifying a majority sample as a minority sample is also a serious issue in some situations. Second, our approach can only handle two class classification at this moment. Multi-class imbalanced learning problems are also very popular and very difficult to solve in medical decision making. Our future research will address these issues by considering different types of scoring functions.
This work was supported by Hong Kong Baptist University Strategic Development Fund, Hong Kong Baptist University grant FRG1/12-13/065, Hong Kong Research grant HKBU202711, and Hong Kong Research grant HKBU12202114.
- Zhang W, Zeng F, Wu X, Zhang X, Jiang R: A comparative study of ensemble learning approaches in the classification of breast cancer metastasis. In Bioinformatics, Systems Biology and Intelligent Computing, 2009. IJCBS’09. International Joint Conference On: IEEE; 2009:242–245.Google Scholar
- Maloof MA: Learning when data sets are imbalanced and when costs are unequal and unknown. In ICML-2003 Workshop on Learning from Imbalanced Data Sets II, Volume 2; 2003.Google Scholar
- Cardie C, Howe N: Improving minority class prediction using case-specific feature weights. In ICML; 1997:57–65.Google Scholar
- Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP: Smote: synthetic minority over-sampling technique. arXiv preprint arXiv:1106.1813 2011.Google Scholar
- Fawcett T, Provost F:Adaptive fraud detection. Data Mining Knowledge Discov. 1997, 1 (3): 291-316. 10.1023/A:1009700419189.View ArticleGoogle Scholar
- Phua C, Alahakoon D, Lee V:Minority report in fraud detection: classification of skewed data. ACM SIGKDD Explorations Newsl. 2004, 6 (1): 50-59. 10.1145/1007730.1007738.View ArticleGoogle Scholar
- Kubat M, Holte RC, Matwin S:Machine learning for the detection of oil spills in satellite radar images. Mach Learn. 1998, 30 (2): 195-215. 10.1023/A:1007452223027.View ArticleGoogle Scholar
- Cui G, Wong ML, Lui H-K:Mach Learn for direct marketing response models: Bayesian networks with evolutionary programming. Manag Sci. 2006, 52 (4): 597-612. 10.1287/mnsc.1060.0514.View ArticleGoogle Scholar
- Abe N, Zadrozny B, Langford J: An iterative method for multi-class cost-sensitive learning. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: ACM; 2004:3–11.Google Scholar
- Chan P, Stolfo S: Toward scalable learning with non-uniform class and cost distributions: a case study in credit card fraud detection. In Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, Volume 164; 1998:168.Google Scholar
- Chen C, Liaw A, Breiman L: Using random forest to learn imbalanced data. 2004, University of California, BerkeleyGoogle Scholar
- Zhou Z-H, Liu X-Y:Training cost-sensitive neural networks with methods addressing the class imbalance problem. Knowl Data Eng IEEE Trans. 2006, 18 (1): 63-77. 10.1109/TKDE.2006.17.View ArticleGoogle Scholar
- Kubat M, Matwin S: Addressing the curse of imbalanced training sets: one-sided selection. In ICML, volume 97; 1997:179–186.Google Scholar
- Drummond C, Holte RC: C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In Workshop on Learning from Imbalanced Datasets II, Volume 11. Citeseer; 2003.Google Scholar
- Batista GE, Prati RC, Monard MC:A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explorations Newsl. 2004, 6 (1): 20-29. 10.1145/1007730.1007735.View ArticleGoogle Scholar
- Sun Y, Kamel MS, Wong AK, Wang Y:Cost-sensitive boosting for classification of imbalanced data. Pattern Recognit. 2007, 40 (12): 3358-3378. 10.1016/j.patcog.2007.04.009.View ArticleGoogle Scholar
- Elkan C: The foundations of cost-sensitive learning. In International Joint Conference on Artificial Intelligence, Volume 17. Citeseer; 2001:973–978.Google Scholar
- Ting KM:An instance-weighting method to induce cost-sensitive trees. Knowl Data Eng IEEE Trans. 2002, 14 (3): 659-665. 10.1109/TKDE.2002.1000348.View ArticleGoogle Scholar
- Domingos P: Metacost: a general method for making classifiers cost-sensitive. In Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: ACM; 1999:155–164.Google Scholar
- Zadrozny B, Langford J, Abe N: Cost-sensitive learning by cost-proportionate example weighting. In Data Mining, 2003. ICDM 2003. Third IEEE International Conference On: IEEE; 2003:435–442.Google Scholar
- Fan W, Stolfo SJ, Zhang J, Chan PK: Adacost: misclassification cost-sensitive boosting. In ICML. Citeseer; 1999:97–105.Google Scholar
- Zhou Z-H, Liu X-Y:On multi-class cost-sensitive learning. Comput Intell. 2010, 26 (3): 232-257. 10.1111/j.1467-8640.2010.00358.x.View ArticleGoogle Scholar
- Weiss GM:Mining with rarity: a unifying framework. ACM SIGKDD Explorations Newsl. 2004, 6 (1): 7-19. 10.1145/1007730.1007734.View ArticleGoogle Scholar
- Chawla NV, Japkowic N, Kotcz A:Editorial: special issue on learning from imbalanced data sets. ACM SIGKDD Explorations Newsl. 2004, 6 (1): 1-6. 10.1145/1007730.1007733.View ArticleGoogle Scholar
- Liu X-Y, Zhou Z-H: The influence of class imbalance on cost-sensitive learning: An empirical study. In Data Mining, 2006. ICDM’06. Sixth International Conference On: IEEE; 2006:970–974.Google Scholar
- McCarthy K, Zabar B, Weiss G: Does cost-sensitive learning beat sampling for classifying rare classes? In Proceedings of the 1st International Workshop on Utility-based Data Mining: ACM; 2005:69–77.Google Scholar
- Freund Y, Schapire RE: Experiments with a new boosting algorithm. In ICML: Morgan Kaufmann Publishers, Inc.; 1996:148–156.Google Scholar
- Hastie T, Tibshirani R, Friedman J, Franklin J:The elements of statistical learning: data mining, inference and prediction. Math Intelligencer. 2005, 27 (2): 83-85.Google Scholar
- Asuncion A, Newman D J: UCI machine learning repository2007.Google Scholar
- Lewis DD, Gale WA: A sequential algorithm for training text classifiers. In SIGIR’94: Springer; 1994:3–12.Google Scholar
- Joshi MV, Kumar V, Agarwal RC: Evaluating boosting algorithms to classify rare classes: Comparison and improvements. In Data Mining, 2001. ICDM 2001, Proceedings IEEE International Conference On: IEEE; 2001:257–264.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.