You can use thresholds values to reduce classes.

Ali Kamali Mohammadzadeh

(1) What do you think about this approach? do you have another suggestion?

The approach is ok. You need to collapse several of the classes because 900 independent classes may be too much

if you dont have large sample set , assuming you have 900 classes of different objects (cap, cup, glass, etc) and you only have 2 or 3 samples for each, the model will in this case not be able to learn properly , because you simply dont have enough samples for each class, for 900 classes, you should be looking at 10k+ samples. Moreover deep neural net needs lots of samples to learn, else it will simply overfit

(2) How should I code the aforementioned approach in Python

since its regression, find a way to collapse the classes , meaning instead of having: 1,2,3,4,5,6,7,8,9,10 you may have: 1-3 (1) 4-6 (2) 7-`10 (3) so it becomes only 3 classes .

you can do a simple clustering to see which set share similar traits ,

then you will get your classes from that and then program your DNN for training and testing.

Apply kmeans with N clusters to generate reduced classes

then training M layer DNN using "relu" activation function .

Dear Naveen,

Thank you for your response. Actually, my data set is quite large (more than 250 thousand observations), so I have more than enough samples for each class.

The clustering approach that you suggested did not occur to me and it sounds like a great way to solve the problem. I will definitely try this approach.

Bests,

Ali Kamali Mohammadzadeh try this

You can use thresholds values to reduce classes.

some methods are discussed here

https://www.analyticsvidhya.com/blog/2015/11/easy-methods-deal-categorical-variables-predictive-modeling/

Ali,

I'll suggest some sort of sampling technique and also utilization of principal component analysis for dimensionality reduction.

Thanks Etu,

I will definitely consider PCA as well as other discussed methods.

Using a threshold techniques to reduce the number of classes.