Breast cancer diagnosis using multi-attributed lens recursive partitioning algorithm

Breast cancer diagnosis can assist in detecting the early stage of breast cancer patients which help alleviate one of the causes of women death in the US. Although, many cancer diagnoses have been done clinically by medical doctors, the help from classification systems can further reduce the misclassification rate based on historical characteristics of patients. Decision tree is one of the classifiers that have been popularly applied. During the construction of a decision tree by a recursive partitioning algorithm a single attribute is selected from candidate attributes to split a dataset using information measures such as the information gain. This paper proposes a new technique, multi-attributed lens, which weighs all numeric attributes simultaneously. A lens is generated using a core vector from a farthest pair of the same class instances. Consequently, data is partitioned into two regions, the outside and the inside lens. All instances in the outside lens are marked as opposite classes to the core vector. The rests are split by their projections on the core vector using the same information measure. UCI Breast Cancer Wisconsin (Original) dataset is used since the characteristics of the breast cancer patients are believed to lie within the lens. Our result shows that relative performances of this algorithm are better than C4.5 algorithm based on this dataset.