reference - How to improve prediction on a different dataset using a trained model?

Question

I have two datasets. These datasets include chemical molecules and their solubility value (numeric values). Target variable (dependent variable) is solubility. Input to the machine learning model is the features extracted from molecular structures (which are also called FingerPrints).

Dataset A: 193 observations or rows (used for training & validation)

Dataset B: 151 observations or rows (used only for prediction)

Problem type: Regression

Model: Gaussian process regressor

Features or columns of both datasets are similar.

This is the procedure I have tried:

1) Train / validate model on dataset A by LOOCV (R2 = 0.75)
2) Fit model on the whole train dataset or dataset A (R2 > 0.9)
3) Use trained model to predict dataset B (R2 < 0.1)

Note 1: I am not allowed to merge or shuffle datasets.

Note 2: I don't use molecular structure as input feature to the model. I extract features from molecules (which are called FingerPrints) using RDKit python library. No. of features and data preprocessing steps for both datasets are similar.

I noticed that molecules in dataset A & B are different. It seems that the model should tackle extrapolation.

My question is that how I can improve prediction when datasets are different.

One idea might be using an observation (one row of data) as "Reference", and train the model on the difference (explained below):

1) Select one observation from dataset A as Reference
2) For observations in dataset A, calculate the difference of features and target variable with the Reference. 
3) Train model on these difference values  
4) For observations in dataset B, calculate the difference of features with the Reference. 
5) Use these differences as input to the trained model to predict target variable
6) Add the value of target variable for Reference to the predicted target values in step 5 for dataset B

I am not sure if using Reference is a valid idea. If you have any idea to improve prediction, please let me know.

question from:https://stackoverflow.com/questions/65881194/how-to-improve-prediction-on-a-different-dataset-using-a-trained-model

Answer 1

Waitting for answers