| Re: Assessing A Machine Learning Method's Predictivity [message #1578 is a reply to message #1523] |
Tue, 05 April 2022 14:56   |
thomas
Messages: 660
Registered: June 2014
|
Senior Member |
|
|
Hello Christophe,
I don't quite understand the question, but assume, there is a misinterpretation of the result. When you running the analysis, the data set is split into 10 fractions:
1: the oldest 10% of the data
2: the second oldest 10%
...
10: newest 10%
Then 9 models are generated using 10%, 20%, 30% ... 90% of the data (always the oldest)
Then every model is used to predict the Y value for the oldest fraction, which was not part of the model creation, e.g. for model 1 it is fraction 2, for model 5 it is fraction 6.
Then for all fractions with predicted data (2-10) a correlation graph is shown with predicted versus known Y-values. Graph 'fraction 8', for instance answers the question: If I had used built a model at the time, when I had 70% of the data and if I had used that model to predict Y-values for the next molecules to synthesize, how well would the prediction have been.
Does this explain it or did I misunderstand the question?
Thomas
|
|
|
|
Search
Help
Members
Register
Login
Home
