PCA-Based Representations of Graphs for Prediction in QSAR Studies

In recent years, more and more attention has been paid on learning in structured domains, e.g. Chemistry. Both Neural Networks and Kernel Methods for structured data have been proposed. Here, we show that a recently developed technique for structured domains, i.e. PCA for structures, permits to generate representations of graphs (specif- ically, molecular graphs) which are quite effective when used for predic- tion tasks (QSAR studies). The advantage of these representations is that they can be generated automatically and exploited by any tradi- tional predictor (e.g., Support Vector Regression with linear kernel).