GrantworthyAI

This repository is a project aimed at developing a deep learning neural network model that assists AlphabetSoup in identifying and choosing funding applicants who are most likely to succeed in their entrepreneurial endeavors.

The project has two main parts:

1. Preprocess the Data

• Within this dataset are a number of columns that capture metadata about each organization, such as:

  • EIN and NAME—Identification columns
  • APPLICATION_TYPE—Alphabet Soup application type
  • AFFILIATION—Affiliated sector of industry
  • CLASSIFICATION—Government organization classification
  • USE_CASE—Use case for funding
  • ORGANIZATION—Organization type
  • STATUS—Active status
  • INCOME_AMT—Income classification
  • SPECIAL_CONSIDERATIONS—Special considerations for application
  • ASK_AMT—Funding amount requested
  • IS_SUCCESSFUL—Was the money used effectively

• The target variable is STATUS

• The feature variables are:

  • APPLICATION_TYPE—Alphabet Soup application type
  • AFFILIATION—Affiliated sector of industry
  • CLASSIFICATION—Government organization classification
  • USE_CASE—Use case for funding
  • ORGANIZATION—Organization type
  • INCOME_AMT—Income classification
  • SPECIAL_CONSIDERATIONS—Special considerations for application
  • ASK_AMT—Funding amount requested
  • IS_SUCCESSFUL—Was the money used effectively

• The EIN and NAME—Identification columns were deleted because they didn't provide valuable data to our model.

2. Compiling, Training, and Evaluating the Model

• The model has 305 neurons, four layers, a reLu activation function on hidden layers, and sigmoid activation function on the output layer. Using the reLu function on hidden layers is a common practice because it is nonlinear yet fast and does not suffer from vanishing gradient problems, and the sigmoid function is a great function for returning probability. I used 128 neurons per hidden layer because it is more than multiple of the input layer's nods, and using a number that is power of two is efficient for modern hardware.
• The model is performing almost perfectly, with an accuracy of 99.98% on the test data.
• Choosing the right number of neurons, suitable activation functions, and efficient data processing all end up improving model performance.

Summary

The problem of finding the best application to fund is a classification, supervised ML problem that various models can be used for it. In some models, the reasoning is more understandable, but they are less accurate(decision trees, random forest, logistic regression), and others are better performers, but the reasoning is not clear with them(SVM, K nearest neighbor, and neural network). If model interpretability is not important, the neural network is the most accurate of them all, as it's been used in this project.