regression : what to do with small number values for categorical features to tune deep neural network

It is regression task *** regression from scikit learn works fine , I want to train NN to get better results *** python code to tune deep NN for tabular data - not NLP , not Images what to do with small number values for categorical features to tune deep neural network **** MAIN DIFFICULTY IS SMALL NUMBER OF OBSERVATIONS FOR EACH UNIQUE VALUE *** do you know how to train NN with the low observation numbers? data 1 million rows for example 20 features each feature has unique 570 categorical values pls provide prove you can do it , I see many people placed bid , but vast majority even do not understand what is it about .. target is continues , mean is 800 data 1 million rows for example 20 categorical features each feature has unique 570 categorical values then in average 140 observations for one unique value 140~ = 100000 / 570 you need generate data by yourself, do not forget include dependency between features and features and target simulate data by yourself with predefined patterns, them prove your code can find these patterns data 1 million rows for example 20 categorical features each feature has unique 570 categorical values then in average 140 observations for one unique value 140~ = 100000 / 570 for example [login to view URL] [login to view URL] or find data with known patterns, them prove your code can find these patterns any ideas are welcome for example transfer learning per [login to view URL] or embedding [login to view URL] [login to view URL] [login to view URL] [login to view URL] [login to view URL] it is not small data but still this can be helpfull [login to view URL] [login to view URL] for example categorical feature "ABC" has 8000 categorical values, but each value observed in train data only ~ 30 times you should be deep NN master and know very well how to tune a NN or may be One-shot-Learning: [login to view URL] first prove you can do it!!