Cancer diagnosis data, for example microarray gene expression profiling data and proteomic profiling data, are often descried by thousands of features To computationally make a diagnosis for new samples, these data are usually input to a learning algorithm, the algorithm then induces a classifier, the classifier then predicts a class label for any test sample. As the data is so high-dimensional, most of the resulting classifiers are very complicated particularly those based on kernel-functions such as support vector machines - the interpretation of the decision results must need all the features to be involved. In this paper, we discuss built-in features and use them to concisely characterize the data and to easily interpret the decisions. Built-in features are features that are used only in the classifiers, and that are only a small subset of the original features, e.g., the features in a decision tree. So, the notion of built-in features is different from input features and also from original features. As there is a significant reduction from the huge size of original features to a small number of relevant features, the complexity of the interpretation can be much eased. The use of built-in features also provides much potential for elucidating the translation between raw data and clinically useful knowledge. In this paper, we also report that the performance of classifiers using built-in features tends to remain stable even input feature space changes, but other types of classifiers fluctuate their performance. So, once again, we promote the use of classifiers that use built-in features since the algorithms can avoid the existing hard problem of selecting best number of features for learning. |
(from crpit.com)
