Training

Basic training requires a sparse matrix of feature vectors as columns per data point X and a corresponding label vector y.

using FactorizationMachines

T = [
5           1 0     1 0 0 0    1 0 0        12.5;
5           1 0     0 1 0 0    1 0 0        20;
4           1 0     0 0 1 0    1 0 0        78;
1           0 1     1 0 0 0    0 0 1        12.5;
1           0 1     0 1 0 0    0 0 1        20;
]

X = sparse(T[:,2:end])'
y = T[:,1]

fm = train(X, y)

Training Components

The training procedure has several components that are specified as keyword arguments to the training procedure: method, evaluator, model_params, and task_params. All instances of a component are specified within their corresponding submodule: Methods, Evaluators, Models, and Tasks. Each component has a default. If no components have explicit instances specified, the base training procedure train(X, y) is equivalent to the following specification of component instances.

fm = train(X, y,
        method       = Methods.sgd(alpha = 0.01, num_epochs = UInt(100), reg0 = .0, regv = .0, regw = .0),
        evaluator    = Evaluators.rmse(),
        task_params  = Tasks.regression(),
        model_params = Models.gauss(k0 = true, k1 = true, num_factors = 8, mean = .0, stddev = .01)

Methods

The Methods module contains a single method for learning:

• Methods.sgd for stochastic gradient descent

fm = train(X, y, method = Methods.sgd(alpha = 0.01, num_epochs = 100)

Evaluators

The Evaluators module contains two evaluation types:

• Evaluators.rmse for using RMSE in evaluations
• Evaluators.heaviside for using the heaviside function for evaluations

Tasks

The Tasks module contains two types of tasks:

• Tasks.regression specifies a regression task
• Tasks.classification specifies a classification task

Models

The Models module contains a single instance for model initialization:

• Models.gauss initializes from a multivariate normal distribution