The purpose of this analysis is to help create a binary classifier that is capable of predicting whether applicatns will be successful if funded by Alphabet Soup. With a dataset that contains over 34,000 organizations that have received funding from Alphabet Soup over the years, a neural network model is compiled, trained, and evaluated. After that, 3 attempts were made to see if the data or model could be optimized such that it can result in over 75% predictive accuracy.
Target variable: IS_SUCCESSFUL Features variable: Columns other than 'IS_SUCCESSFUL' Neither targets nor features variable (removed): 'EIN' and 'NAME'
Without optimization: Hidden layers: 2 Neurons: 12 for 1st, 8 for 2nd Activation function (hidden layer): Relu Activation function (output layer): Sigmoid
Result:
1st attempt: Modification: dropped 'ORGANIZATION' column
Hidden layers: 2 Neurons: 12 for 1st, 8 for 2nd Activation function (hidden layer): Relu Activation function (output layer): Sigmoid
Result:
2nd attempt: Modification: replaced 'ASK_AMT' values to 'Others' if counts are less than 5
Hidden layers: 2 Neurons: 12 for 1st, 8 for 2nd Activation function (hidden layer): Tanh Activation function (output layer): Sigmoid
Result:
3rd attempt: Modification: did not drop the 'NAME' column, replaced its values to 'Others' if counts equal to 1, replaced 'APPLICATION_TYPE' values to 'Others' if counts are less than 200, replaced 'CLASSIFICATION' values to 'Others' if counts are less than 300, replaced 'ASK_AMT' values to 'Others' if counts are less than 5, increased neurons (15 neurons per hidden layer), used MinMaxScaler, used relu activation function
Hidden layers: 2 Neurons: 30 for all both layers Activation function (hidden layer): Tanh Activation function (output layer): Sigmoid
Result:
There is no particular reasons for choosing numbers of layers and neurons. However some trials and errors were made (using up to 200 neurons per hidden layer for up to 15 layers, etc) to see if the model is underfitting or overfitting. In the end, 15 hidden neurons with 2 hidden layers were enough to achieve better performance.
After 3 attempts to optimize the model. The model was able to achieve 79.81% accuracy.
After 3 attempts to optimize the model. it was able to achieve over 75% accuracy, with 79.81%.
Since there are multiple features in the dataset, a Random Forest Classifier model may be able to perform better on this dataset since one of its strength is the ability to weighting importance on features.