Igor Baskin - Inductive Transfer of Knowledge: Application of Multi-Task Learning and Feature Net Approaches to Model Tissue-Air Partition Coefficients

Document created by Igor Baskin on Aug 22, 2014
Version 1Show Document
  • View in full screen mode

  Publication Details (including relevant citation   information):

  Varnek, A.; Gaudin, C.; Marcou, G.; Baskin, I.; Pandey, A. K.;   Tetko, I. V. J. Chem. Inf. Mod. 2009,   49 (1), 133-144.


  Two inductive knowledge transfer approaches - multitask learning   (MTL) and Feature Net (FN) - have been used to build predictive   neural networks (ASNN) and PLS models for 11 types of tissue-air   partition coefficients (TAPC). Unlike conventional single-task   learning (STL) modeling focused only on a single target property   without any relations to other properties, in the framework of   inductive transfer approach, the individual models are viewed as   nodes in the network of interrelated models built in parallel   (MTL) or sequentially (FN). It has been demonstrated that MTL and   FN techniques are extremely useful in structureв€’property   modeling on small and structurally diverse data sets, when   conventional STL modeling is unable to produce any predictive   model. The predictive STL individual models were obtained for 4   out of 11 TAPC, whereas application of inductive knowledge   transfer techniques resulted in models for 9 TAPC. Differences in   prediction performances of the models as a function of the   machine-learning method, and of the number of properties   simultaneously involved in the learning, has been discussed.

  Address (URL): http://pubs.acs.org/doi/abs/10.1021/ci8002914