Graph-based Feature Extraction for Drug Classification Using Machine Learning

Drug classification is a key task in medical decision-making, and this process can be supported through appropriate selection of drugs that fit patient attributes and medical history. Currently, state-of-the-art solutions for this problem tend to make use of attribute-based representation and may overlook the relational structure that could exist in the data points. This paper seeks to examine a machine learning algorithm for multiclass drug classification, where the target variable is the prescribed type of drug, using a graph-based feature extraction method. The proposed method can be used thorough data preprocessing strategies, including handling imbalance and removal of outliers. Evaluation of the method requires a 10-fold cross-validation strategy to achieve a genuine and impartial evaluation. In each training set, a similarity graph is created and graph-based attributes such as degree centrality measures and clustering coefficient are extracted by using three different distance measures to consider graph relationships between samples. In the case of the testing data, graph attributes are computed by supplanting a weighted k-nearest neighbor strategy, wherein there is absolute prevention of information leakage. The graph attributes obtained along with other attributes are used to train different machine learning classifiers. The results show that the performance across all distance measures plays a great role for such classifications.