Poisson Regression

Creates a regression model that assumes data has a Poisson distribution

Category: Machine Learning / Initialize Model / Regression

Module overview

This article describes how to use the Poisson Regression module in Azure Machine Learning Studio to create a Poisson regression model.

Poisson regression is intended for use in regression models that are used to predict numeric values, typically counts. Therefore, you should use this module to create your regression model only if the values you are trying to predict fit the following conditions:

  • The response variable has a Poisson distribution.

  • Counts cannot be negative. The method will fail outright if you attempt to use it with negative labels.

  • A Poisson distribution is a discrete distribution; therefore, it is not meaningful to use this method with non-whole numbers.

Tip

If your target isn’t a count, Poisson regression is probably not an appropriate method. Try one of the other modules in this category. For help picking a regression method, see the Azure machine Learning algorithm cheat sheet.

After you have set up the regression method, you must train the model using a dataset containing examples of the value you want to predict. The trained model can then be used to make predictions.

More about Poisson regression

Poisson regression is a special type of regression analysis that is typically used to model counts. For example, Poisson regression would be useful in these scenarios:

  • Modeling the number of colds associated with airplane flights

  • Estimating the number of emergency service calls during an event

  • Projecting the number of customer inquiries subsequent to a promotion

  • Creating contingency tables

Because the response variable has a Poisson distribution, the model makes different assumptions about the data and its probability distribution than, say, least-squares regression. Therefore, Poisson models should be interpreted differently from other regression models.

How to configure Poisson Regression

  1. Add the Poisson Regression module to your experiment in Studio.

    You can find this module under Machine Learning - Initialize, in the Regression category.

  2. Add a dataset that contains training data of the correct type.

    We recommend that you use Normalize Data to normalize the input dataset before using it to train the regressor.

  3. In the Properties pane of the Poisson Regression module, specify how you want the model to be trained, by setting the Create trainer mode option.

    • Single Parameter: If you know how you want to configure the model, provide a specific set of values as arguments.

    • Parameter Range. If you are not sure of the best parameters, do a parameter sweep using the Tune Model Hyperparameters module. The trainer iterates over multiple values you specify to find the optimal configuration.

  4. Optimization tolerance: Type a value that defines the tolerance interval during optimization. The lower the value, the slower and more accurate the fitting.

  5. L1 regularization weight and L2 regularization weight: Type values to use for L1 and L2 regularization. Regularization adds constraints to the algorithm regarding aspects of the model that are independent of the training data. Regularization is commonly used to avoid overfitting.

    • L1 regularization is useful if the goal is to have a model that is as sparse as possible.

      L1 regularization is done by subtracting the L1 weight of the weight vector from the loss expression that the learner is trying to minimize. The L1 norm is a good approximation to the L0 norm, which is the number of non-zero coordinates.

    • L2 regularization prevents any single coordinate in the weight vector from growing too much in magnitude. L2 regularization is useful if the goal is to have a model with small overall weights.

      In this module, you can apply a combination of L1 and L2 regularizations. By combining L1 and L2 regularization, you can impose a penalty on the magnitude of the parameter values. The learner tries to minimize the penalty, in a tradeoff with minimizing the loss.

      For a good discussion of L1 and L2 regularization, see L1 and L2 Regularization for Machine Learning.

  6. Memory size for L-BFGS: Specify the amount of memory to reserve for model fitting and optimization.

    L-BFGS is a specific method for optimization, based on the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm. The method uses a limited amount of memory (L) to compute the next step direction.

    By changing this parameter, you can affect the number of past positions and gradients that are stored for computation of the next step.

  7. Connect the training dataset and the untrained model to one of the training modules:

    • If you set Create trainer mode to Single Parameter, use the Train Model module.

    • If you set Create trainer mode to Parameter Range, use the Tune Model Hyperparameters module.

      Warning

      • If you pass a parameter range to Train Model, it uses only the first value in the parameter range list.

      • If you pass a single set of parameter values to the Tune Model Hyperparameters module, when it expects a range of settings for each parameter, it ignores the values and uses the default values for the learner.

      • If you select the Parameter Range option and enter a single value for any parameter, that single value you specified is used throughout the sweep, even if other parameters change across a range of values.

  8. Run the experiment to train the model.

Examples

For examples of how Poisson regression is used in machine learning, see the Azure AI Gallery.

Technical notes

Poisson regression is used to model count data, assuming that the label has a Poisson distribution. For example, you might use it to predict the number of calls to a customer support center on a particular day.

For this algorithm, it is assumed that an unknown function, denoted Y, has a Poisson distribution. The Poisson distribution is defined as follows:

Given the instance x = (x0, …, xd-1), for every k=0,1, …, the module computes the probability that the value of the instance is k.

Given the set of training examples, the algorithm tries to find the optimal values for θ0, …,θD-1 by trying to maximize the log likelihood of the parameters. The likelihood of the parameters θ0, …,θD-1 is the probability that the training data was sampled from a distribution with these parameters.

The log probability can be viewed as logp(y = yi)

The prediction function outputs the expected value of that parameterized Poisson distribution, specifically: fw,b(x) = E[Y|x] = ewTx+b.

For more information, see the entry for Poisson regression in Wikipedia.

Module parameters

Name Range Type Default Description
Optimization tolerance >=double.Epsilon Float 0.0000001 Specify a tolerance value for optimization convergence. The lower the value, the slower and more accurate the fitting.
L1 regularization weight >=0.0 Float 1.0 Specify the L1 regularization weight. Use a non-zero value to avoid overfitting the model.
L2 regularization weight >=0.0 Float 1.0 Specify the L2 regularization weight. Use a non-zero value to avoid overfitting the model.
Memory size for L-BFGS >=1 Integer 20 Indicate how much memory (in MB) to use for the L-BFGS optimizer. With less memory, training is faster but less accurate the training.
Random number seed any Integer Type a value to seed the random number generator used by the model. Leave blank for default.
Allow unknown categorical levels any Boolean true Indicate whether an additional level should be created for each categorical column. Any levels in the test dataset not available in the training dataset are mapped to this additional level.

Outputs

Name Type Description
Untrained model ILearner interface An untrained regression model

See also

Regression
A-Z Module List