I got the following result after training a model:
which zoomed in looks like:
From this link, it seems like my model is just right. Does my model fit well? How do I evaluate this result?
$\begingroup$
$\endgroup$
2
-
1$\begingroup$ What are the units for error? $\endgroup$– Michael R. ChernickCommented May 5, 2017 at 20:36
-
$\begingroup$ The unit for error is the amount of the links utilization. Please correct me if something's wrong here. Thanks, Mike! $\endgroup$– Tung LeCommented May 5, 2017 at 21:11
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
I have an answer on how to read learning curve here:
How to know if a learning curve from SVM model suffers from bias or variance?
I would suggest you use more data, instead of 1 to 27 data points. The reason is in your plot, the numbers get very small after passing $n=3$ and the $y$ axis is in a larger scale.
In sum,
- Try to use more data
- Try to make $y$ axis scale smaller to clearly see two lines.
answered May 5, 2017 at 20:47
-
1$\begingroup$ Thanks, hxd1011. From what I've learned from ML, there is a high bias (underfit) when J_{train}{\theta} is high and J_{cross-validate}(\theta) ~ J_{train}{\theta}, so getting more data will not help much. There is a high variance (overfit) when J_{train}(\theta) is low and J_{cross-validate}(\theta) >> J_{train}{\theta}, so getting more data is likely to help. But in my case, J_{train}(\theta) is low, J_{cross-validate}(\theta)~J_{train}(\theta). That's why I am confused about this situation. As the data are pretty scarce in my case, are there alternatives to solve it without using more data? $\endgroup$– Tung LeCommented May 5, 2017 at 21:19