PNN is a native ranker neural network implemented for label ranking using Spearman ranking error function and Positive/Smooth Staircase activation function (SS) and (PSS) to enhance the prediction probability to produce almost discrete value without data freedom between layers, thus it use one middle layer for learning.
SGPNN is extended PNN to rank subgroup of labels using one learning model. the subgroups are combined from multiple domains to find a hidden relations between these groups.
These two networks use a new type of multi-values activation functions. Smooth Staircase (SS) employed for ranking, The following equations shows the positive output values of SS function.
and 2b=n-1 where n is the number of steps and b is the boundary value on x axis.
Symmetric Smooth Staircase (SSS) function where # steps = 5 and boundaries between -1 and 1 is:
Manipulate[Plot[(-(s/2)(Sum[( Tanh[(c(b-x-(w*i)))] ), {i, 0, n - 1}]-(1)) ), {x, -4, 4}], {n, 5},{s,1,1000}, {c, 100}, {b, 2},{w,1}]
Symmetric Smooth Staircase (SSS) function for regression value up to 2 decimal value where # steps = 5 and boundaries between -1 and 1 is:
Using python +3.7
pnn = PNN()
train_error = pnn.loadData(filename=path,featuresno= 4,labelno=3,ssteps=2,epochs=500,lrate=0.005,hn=100,Fold=10,useFold=False)
For feedback kindly communicate using my email [email protected]
Video Demo available at https://drive.google.com/drive/folders/1yxuqYoQ3Kiuch-2sLeVe2ocMj12QVsRM?usp=sharing
Please Cite using the following links:
A. Elgharabawy, M. Prasad and C. -T. Lin, "Preference Neural Network," in IEEE Transactions on Emerging Topics in Computational Intelligence, doi: 10.1109/TETCI.2023.3268707.
A. Elgharabawy, M. Prasad, and C.-T. Lin, “Subgroup Preference Neural Network,” Sensors, vol. 21, no. 18, p. 6104, Sep. 2021, doi: 10.3390/s21186104.