Neural Network Algorithm

The Neural Network algorithm in SQL Server Analysis Services (SSAS) is a powerful tool for classification and regression tasks. It simulates the structure and function of biological neural networks to learn complex patterns in data.

Overview

Neural networks are composed of interconnected nodes (neurons) organized into layers: an input layer, one or more hidden layers, and an output layer. Each connection between neurons has a weight, which is adjusted during the training process. The algorithm learns by iteratively adjusting these weights to minimize the difference between its predictions and the actual outcomes in the training data.

A simplified representation of a neural network architecture.

Key Concepts

Input Layer: Receives the independent variables (features) from the data mining model.
Hidden Layers: Intermediate layers that process the input and extract complex features. The number and size of hidden layers can significantly impact model performance.
Output Layer: Produces the prediction (e.g., class label for classification, continuous value for regression).
Weights: Numerical values associated with connections between neurons, determining the strength of the signal passed.
Activation Function: A non-linear function applied to the output of each neuron, enabling the network to learn non-linear relationships.

Applications

The Neural Network algorithm is well-suited for scenarios where:

Complex, non-linear relationships exist between variables.
The dataset is large and contains many features.
Accurate prediction is critical.

Common applications include:

Customer churn prediction
Fraud detection
Image and speech recognition (though typically more specialized architectures are used here)
Medical diagnosis

Parameters

Key parameters for the Neural Network algorithm in SSAS include:

HIDDEN_LAYERS: Specifies the number and size of hidden layers. For example, 2-10-5 would indicate two hidden layers, the first with 10 neurons and the second with 5.
MAX_NEURONS_PER_LAYER: Sets an upper limit on the number of neurons in any hidden layer.
LEARNING_RATE: Controls the step size for weight updates during training.
MAX_ITERATIONS: The maximum number of training iterations to perform.
OPTIMIZATION_GOAL: Defines the objective function to minimize (e.g., SUM_SQUARED_ERROR for regression, CROSS_ENTROPY for classification).

Implementation Example (DMX)

To create a Neural Network mining model:


CREATE MINING MODEL [MyNeuralNetworkModel]
(
    [CustomerID] LONG KEY,
    [IsChurn]   DISCRETIZED (1, 0) SHARED STRUCTURE,
    [Demographics] STRUCTURE
    (
        [Age] LONG,
        [Gender] TEXT,
        [Income] DOUBLE
    ) PREDICTED `IsChurn`
)
USING NEWOWAN_CLASSIFICATION(
    HIDDEN_LAYERS = '2-10-5',
    MAX_ITERATIONS = 500,
    OPTIMIZATION_GOAL = 'CROSS_ENTROPY'
)

Usage and Prediction

Once trained, the model can be used to predict the likelihood of churn for existing or new customers:


SELECT
    [CustomerID],
    Predict (`IsChurn`) AS PredictedChurn,
    PredictProbability (`IsChurn`, 1) AS ChurnProbability,
    PredictProbability (`IsChurn`, 0) AS NoChurnProbability
FROM
    [MyNeuralNetworkModel]
NATURAL PREDICTION JOIN
    OPENQUERY(DataSourceView, 'SELECT CustomerID, Age, Gender, Income FROM Customers WHERE CustomerID = 12345')

Performance Considerations

Training neural networks can be computationally intensive. Consider data normalization, feature selection, and careful parameter tuning to optimize performance and prevent overfitting.

MSDN Documentation