Updated Ljubljana Breast Cancer Data Set: reduced and cleaned version
Description
This dataset contains information for Machine Learning algorithms to forecast recurrence events (RE) for patients with breast cancer stages I to III. The dataset contains 252 instances and six attributes, including a binary class indicating whether RE occurred. This dataset has been reduced and denoised from the original Ljubljana, which holds 286 instances with ten attributes each (LBCD, Zwitter M. and Soklic M. (1988). Breast Cancer. UCI Machine Learning Repository. https://archive.ics.uci.edu/dataset/14/breast+cancer). The ranking results by eight different Machine learning algorithms and statistical handling of the ranking 8-component vectors for attributes allow one to reduce ten features to six of the most relevant ones. The most pertinent features were the following five: {deg_malig, irradiat, node_caps, tumor_size, inv_nodes}. Less relevant found four attributes: {age, breast_quad, breast, menopause}. The CAIRAD: Co-appearance based Analysis for Incorrect Records and Attribute-values Detection ( Rahman MG, Islam MZ, Bossomaier T, Gao J. CAIRAD: A co-appearance based analysis for incorrect records and attribute-values detection. Proc Int Jt Conf Neural Networks. 2012;(June). https://doi.org/10.1109/IJCNN.2012.6252669) filter has been determined the noises in attributes and class features. Per the filtering results, 34 instances of LBCD had noises in half (or even more than half) of their features. Those were removed from the data. It is known that the noises in the class are riskier and teasing than those of attributes. Meantime, the class attribute had 35 (14%) missed values from 252 after COIRAD filtering. It was unacceptable, considering the comparable number (only 85 cases) of recurrence events in the class of initial LBCD. The imputation (reconstruction, "cure") of missed values was performed via the algorithm offered in: Bai BM, Mangathayaru N, Rani BP. An approach to find missing values in medical datasets. In: ACM International Conference Proceeding Series. Vol 24-26-Sept. ; 2015. https://doi.org/10.1145/2832987.2833083. The noises presented in the remaining attributes, ranging from 1% to 14%, were neglected. There are 252 instances in the dataset, of which 206 do not have RE, and the remaining 46 have RE. Six attributes, including its class, define each instance. This dataset is obtained from the initial version of the LBCD betterment, and it provides a significant advantage in the performance over the original LBCD for most classifying algorithms of Machine Learning. However, the dataset is slightly more imbalanced than the LBCD, which is a minus.
Files
Steps to reproduce
This is ARFF file handy to WEKA software. Such files are openable via any text' redactor, let us say Notepad.