Neural Net classifer vs boosted trees classifier: data types incompatible

User 2574 | 11/10/2015, 7:29:59 PM

I am looking to compare the performance of boosted trees classifier and a neural net classifier. I started with boosted trees and built the input data with certain data types including 'dict'. It works great with boosted trees. Now when I want to use the same input data into the neural network classifier, it complains of incompatible data types - I think it is referring to 'dict' type in several of the columns. If indeed the input data cannot be ported between various classifiers then it is a major drawback since a lot of time is spent on feature engineering and making sure it is compatible with a certain group of classifiers. If I have to re-engineer those features to make it compatible with neural nets then I am spending time in an unnecessary direction. Let me know if this is the case or not

Comments

User 1207 | 11/17/2015, 12:12:39 AM

Hello Javiar,

Thank you for your feedback. The neural net classifier is a special case here due to some under-the-hood engineering details in how the layers are constructed and interact with the data, but we are currently working to fix this issue and make everything take the same input types. All the other classifiers should handle all types similarly.

-- Hoyt


User 2574 | 11/22/2015, 6:37:34 PM

Thanks Hoyt. I look forward to the update whenever it becomes available.


User 3147 | 2/7/2016, 9:40:15 PM

I had a similar problem recently. It would be great help if you manage to resolve this incompatibility.


User 2568 | 2/8/2016, 9:00:05 PM

Same. I did not try neural nets at that point as I did not have time to figure out how they worked in GraphLab Create