Open ankursharma-iitd opened 5 years ago
We can use the second method mentioned in the paper. Compatibility score can be calculated directly by changing the dimensionality of L_i to that of G and then taking dot product of L_i and G.
We can use the second method mentioned in the paper. Compatibility score can be calculated directly by changing the dimensionality of L_i to that of G and then taking dot product of L_i and G.
I also did in the same way.
Actually, we have to use both the methods for replicating the tables 1 and 2. (dp and cp) Anyway, adding a single-neuron layer works in my case for the cp part.
The other compatibility measure used in the paper involves taking the dot product of a vector u with the addition of L_i and G. I'm assuming that the weights of this vector u have to be learned during the training process. How can we implement this in our project? Can we simply map the addition result (L_i + G) to a scalar by passing it from perceptron (without non-linearity)?