I am working on a data mining project and some of features varies between 1 to 1,000,000. Before training machines, I do the standardization for all the data features and do the rest of machine learning process (Regression method). The problem is to reduce some error of cource, but MAE, MSE, ... they are all metrics that bias the lines to bigger differences between train data and machine hyperplane (I hope used the right expression!). I decided to find the best fitted line to relative error ((x_estimation - x_test) / x_estimation). One of my friend said you can use some trick (maybe on data or model) that can help you to find your model faster.
I searched the web for hours, yet haven't found anything. He does not show me explicitly what that trick is, so I decided to ask you researchers if you got any useful hints, references, or python codes on github to clarify the problem for me, let me know.
I appreciate your helps.
Have you used an data transformation and attributes selection method?
If the true weight of an object is x and a scale measures it as y, then
Absolute error A = Absolute value of (x-y)
Relative error R = Absolute error / x
% error P = Relative error * 100
Example: If 1Kg is measured as 990 gm, A= 10 gm, R = 0.01, P= 1%
However if 100gm is measured as 90 gm, A= 10 gm, R= 0.1, P= 10%
Though absolute errors are same in both the cases, relative errors are Different by a factor of 10.
Therefore, Relative error is a better indicator of error than Absolute error
Assuming that minimizing absolute (or squared) error means minimizing the mean of absolute (or squared) error over a number of predictions: If the difference between actual value and predicted value is large, the squared difference would be larger. So, squared error approach penalizes large errors more as compared to absolute error approach. Therefore, if you want the model to penalize large errors more, minimizing squared error would be better.
Also, large errors do not get magnified in case of absolute error approach, some points may be allowed to have large errors while most have comparatively small errors. So, a model learnt by minimizing absolute error might have more variance in the error values for the training samples, whereas a model which minimizes squared error would tend to learn a model where errors do not vary a lot (or are closer to the mean of the error values). This also implies that if there are more outliers, or more noise in target values, minimizing absolute error would be better.
nice and thorough answer Mr Shankar.
I got better vision about model and data now
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining