Alphabet Soup Charity - Deep Learning Model

Overview of the Analysis

The purpose of this analysis is to build a binary classifier using a deep learning neural network. This model predicts the success of applicants funded by Alphabet Soup, helping the organization select applicants with the highest likelihood of success in their ventures.

I worked on developing a Neural Network model

Data Preprocessing

• Target Variable:

  • IS_SUCCESSFUL

• Features:

  • APPLICATION_TYPE
  • AFFILIATION
  • CLASSIFICATION
  • USE_CASE
  • ORGANIZATION
  • STATUS
  • INCOME_AMT
  • SPECIAL_CONSIDERATIONS
  • ASK_AMT

• Removed (Dropped) Variables:

  • EIN: Identification column not relevant to predictions.
  • NAME: Identification column that may introduce noise without adding predictive value.

• Handling Categorical Data: Categorical variables with more than 10 unique values were grouped into an "Other" category. and One-hot encoding was applied to categorical features to convert them into numeric format.

• Scaling the Features: StandardScaler was used to normalize numerical data to ensure consistency and improve model convergence.

• Data Splitting: The dataset was split into training (80%) and testing (20%) sets using train_test_split.

Compiling, Training, and Evaluating the Model

• Model Structure:

  • Neurons: Initial configurations used between 30 and 200 neurons per layer.
  • Hidden Layers: Multiple attempts with 2-5 layers to balance complexity and training efficiency.
  • Activation Functions: relu for hidden layers and sigmoid for the output layer (binary classification).

• Performance:

  • Initial attempts achieved ~73% accuracy.
  • Various techniques were used to optimize performance, including:
    • Adjusting neurons and layers.
    • Using dropout and batch normalization.
    • Adjusting learning rate and adding callbacks.

Optimization Steps

To achieve a target accuracy of 75%, the following optimizations were attempted:

1. Data Adjustments:

   • Encoding categorical variables with pd.get_dummies().
   • Combining rare categorical values into "Other".
   • Scaling features using StandardScaler().

2. Model Adjustments:

   • Increased the number of neurons in hidden layers.
   • Added more hidden layers for greater learning capacity.
   • Experimented with different activation functions (relu, tanh).

3. Training Adjustments:

   • Increased the number of epochs.
   • Used EarlyStopping and ModelCheckpoint callbacks to improve training efficiency.

Final Model

• Accuracy: Despite extensive optimization, the highest accuracy achieved was ~73%. Then I had to try different approach that is focused on identifying and handling low-frequency categories in a dataset:

• Replacing with "Other": Consolidate rare categories into a single "Other" category.
• Dropping: Remove the data associated with these rare categories (only if justified).
• Encoding Separately: Treat them differently in downstream processing.

Questions

• Which features were used as inputs? Categorical features as application type, organization type, special considerations were utilized, transformed into numerical data using one-hot encoding.

• What was the target variable? The target variable was IS_SUCCESSFUL, which represents whether the charity was successful.

• What methods were used to preprocess the data? The data was prepared by removing non-essential columns, applying one-hot encoding for categorical variables, scaling numerical features with StandardScaler, and splitting the dataset into training and testing subsets.

• How was the model structured? The model was a neural network with first one hidden layer then increased by 5 comprising 30 and 200 neurons, using ReLU activation for hidden layers and a Sigmoid function for the output layer.

• What were the model’s performance metrics? The optimized model achieved an accuracy of 72%-73% and showed a reduced loss value, starting from 0.577 and improving during the optimization process.

• How was the model optimized? Optimization included:

• Modifying the architecture by experimenting with layers and neurons.
• Tuning hyperparameters like learning rate and increasing training epochs.
• Incorporating early stopping to mitigate overfitting.

Summary

• The deep learning model performed well but did not achieve the target accuracy of 75%, by adding hidden layers.
• Recommendation:
  • Perform feature importance analysis to focus on the most predictive variables.