Data mining: practical machine learning tools and techniques with Java implementations
Data mining: practical machine learning tools and techniques with Java implementations
Estimation of the number of clusters and influences zones
Pattern Recognition Letters
Learning from imbalanced data sets with boosting and data generation: the DataBoost-IM approach
ACM SIGKDD Explorations Newsletter - Special issue on learning from imbalanced datasets
Learning on the border: active learning in imbalanced data classification
Proceedings of the sixteenth ACM conference on Conference on information and knowledge management
Expert Systems with Applications: An International Journal
| Hi-index | 0.00 |
Improving the learner performance over imbalanced and multidimensional datasets raises a challenging task for machine learning community. Although a salient characteristic in data modeling is the amount of data provided for the learner, the proportional distribution of that data in each class has also direct relationship with the classifier performance. In imbalanced datasets when data is distributed into different classes, various in size, understanding of data structure and characteristics plays an important role in improving the learner accuracy. In this paper we introduce a new approach that combines the information gained from traditional classification algorithms, confusion matrix parameters and density-based clustering to generate artificial data in order to increase the learner performance. First a classification algorithm is run on training data. Then the confusion matrix is studied and the True Positive (TP) rate of each class is measured. The class with the lowest TP rate is selected. Using density-based clustering we identify the centroid of the class and measure the samples distribution in multidimensional space in the next step. With the values gained from Probability Density Function estimations for clusters, extra samples are generated and added to the original dataset to rebalance the class proportion and the weight of different classes in the whole training set. Our method has been evaluated in terms of TP, F-Measure and also overall accuracy against a number of Demetra (toxicology) and UCI datasets. Our method provides an insight view of the data structure and characteristics in order to identify how much and where the data need to be added for increasing the classification accuracy of the learner.