Modeling Structure Property Relationships with Kernel Recursive Least Squares

Motivation. Modeling structure property relationships accurately is a challenging task and newly developed kernel based methods may provide the accuracy for building these relationships. Method. Kernelized variant of traditional recursive least squares algorithm is used to model two QSPR datasets. Results. All the datasets showed a good correlation between actual and predicted values of boiling points with root mean squared errors (RMSEs) comparable to other conventional methods. For the datasets from Espinosa et al., KRLS showed good prediction statistics with R value in the range of 0.97–0.99 and S value in the range 5.5– 8 as compared to multiple linear regression (MLR) with R value in the range 0.85–0.88 and S value in the range 22–26. For the dataset from Trinajstiü et al., KRLS performed consistently well with R values lying in the range of 0.95–0.99 and S in the range of 5–10 as compared to MLR with R values in the range of 0.7–0.85 and S in the range of 25–30. Conclusions. The KRLS method works better when more number of variables from the dataset are included as against other methods such as support vector learning or lazy learning technique which works better for smaller number of reduced relevant variables from the dataset.