I build collaborative filtering recommender system using surprise library in python. My dataset contains of three columns ( 'ReviewerID', 'ProductID', 'Rating') , the rating scale [-30,40] , I calculate the RMSE and it equals to 0.9 , Then I make a normalization process for the rating to change the scale to [-0.4, 0.4] when I calculate the RMSE it equals to 0.003 .. The difference in the RMSE is big and not reasonable, is it wrong to normalize the rating scale in CF?
Mohammed is correct. Anytime you impose a transformation (here, I'm presuming a linear transform), you will change the values of measures of location, dispersion, or both. Hence, so will aggregated discrepancy measures shift in value (like RMSE). So, now you no longer have an apples-to-apples comparison by having changed scales.
Good luck with your work.
Please be aware of how an error measure is calculated. The absolute difference between actual and predicted values is different for your two scales. The maximal difference is 70 for scale 1, but only 0.8 for scale 2. These error differences are furthermore squared before summed up. The sums of the squared differences thus are different not only in their absolute numbers, but also in relation to the scales. You hence need to compare RMSEs that are based on equally scaled data or you need to use error measures that are invariant to the scale transformations you applied.
Good luck with your project.
Thank you Mohammed Deeb , David Morse and Michael Scholz for your answers.
- Badges
- Science topic