calistic Goto Github PK

Hidden Layers: 2 1st Layer: 12 Neurons 2nd Layer: 12 Neurons Iterations: 10

I played around with many different numbers of layers and neurons. Of the top performers, this configuration was the most efficient (fewest neurons).

The best I could achieve was around 72.5% accuracy

Data Preperation

• Delete Columns: EIN, NAME, & SPECIAL_CONSIDERATIONS_N (redundant with SPECIAL_CONSIDERATIONS_Y)
• Scale the data
• Delete outlier data
• Bin data into various group sizes

Training:

• Experiment with activation functions: relu, sigmoid, tanh
• Try differnt amounts of hidden layers
• Increase training iterations
• Try random forests

The dataset is a little small. If the dataset is made larger (synthetically or organically) I would stick with Deep Learning and Random Forest.

The statistical models are good for categorizing data, but may not be as good at making predictions.

I'm not certain what is best. Perhaps we could statistically categorize the data, then have a human manually process the groups, or push the groups into a machine learning model.