Review user | www.CADASTER.eu

try another color:

try another fontsize: 60% 70% 80% 90%

www.CADASTER.eu

Name of method:

PLS regression in feature subspace combined with linear regression

Descriptors used:
(e.g., sets provided on the challenge site)

E-state indices	Dragon descriptors
SimulationsPlus	QuantumChemistry
MOE	Only structures (e.g. in graph kernels)

Additional descriptors (indicate type, vendor or/and provide brief description and references):

Chemical classes:

Have you used information about chemical classes to build your model?
No
Yes
If yes, please, describe:

Mechanism and/or mode of action:

Have you used information about mechanisms and/or modes of action to build your model?
No
Yes
If yes, please, describe:

Descriptor selection:

Have you done selection of descriptors?
No
Yes
If yes, specify, number of descriptors selected for the final model:

If yes, describe the used procedure:

Genetic algorithm was used to search for the best subspace of each descriptor set, using repeated 2-fold crossvalidation.

Method (select one):

If any other, please, specify:

Have you also used the Known set to develop the final model?

No
Yes

Software suite/vendor (select one):

If any other, please, specify:

Error Prediction

Have you provided the predicted errors?
No
Yes
If yes, provide the description:

Detailed description of the method:

The description should be sufficient to reproduce results, please, also provide references to the respective articles with more details, if available:

I have used genetic algorithm to select the best descriptor subset from all 5 sets. Then I have joint them together in various combinations (once again evaluated using genetic algorithm) to finally obtain 30 almost equally good sets. Then PLS and linear regression was trained on those sets, resulting in ~60 models in total, validated using 5-fold CV this time. Finally about half of the models have been chosen (with only one PLS, the rest was linear regression) and their outputs have been combined. 

The test have shown about 0.126 10-fold CV mse on the training data, 0.075 10-fold CV mse on training+known test data and about 0.055 mse on known test data.

I do suspect however, that the final combination might overfit a bit due to inclusion of PLS regression (I have identified two peaks in the predictions for blind test data - negative on object 20 and positive on object 119), which are not there when PLS is not included. But the CV results train+known are worse without PLS.