The input dimension are (100,104,1) in shape and each value could be either 1 or 0. This is basically a multilabel multiclass problem where output needs to be mapped to a 104 bit vector. 104 bit representing 104 classes and input dimensions could mapped to multiple-class at a given time. For that One hot shot isn't an option. I had used sequential neural network with relu as the activation for the hidden layers and sigmoid in the output layer(104 sigmoid function in output layer). The NN Model is predicting same values for every inputs while testing but during training the loss went down to quite low values. Binary cross-entropy has been used as the loss function.
Dear Hasibul Jamil ,
Great discussion point. Following the updates!
What is a type of units in the NN?
Does the notation (100,104,1) mean that you have 100 one-dimensional vectors each of 104-bit length?
Dear Hasibul Jamil
Try to use a hard-limiter activation functions in the output layer for classifier. In hidden layers use the sigmoid functions.
Regards
Plz use categorical-correspondence instead of binary. If you want multiclass we use that instead of binary
Hasibul Jamil
I think the tree-based classifier is the best for your case especially the random forest classifier.
I agree with Dr. K. Seetharaman . Properly trained random forrest can bring good results. RF in classification problem deals - by its nature - with binary data (Y/N). As a bonus, you will get the information which type of input influence the class assignment most. (i.e. all inputs will be ranked).
Regards
Hubert
- Badges
- Science topic