golinear

Overview ¶

Package golinear trains and applies linear classifiers.

The package is a binding against liblinear with a Go-ish interface. Trained models can be saved to and loaded from disk, to avoid the (potentially) costly training process.

A model is trained using a problem. A problem consists of training instances, where each training instance has a class label and a feature vector. The training procedure attempts to find one or more functions that separate the instances of two classes. This model can then predict the class of unseen instances.

Consider for instance that we would like to do sentiment analysis, using the following, humble, training corpus:

Positive: A beautiful album.
Negative: A crappy ugly album.

To represent this as a problem, we have to convert the classses (positive/negative) to an integral class labels and extract features. In this case, we can simply label the classes as positive: 0, negative: 1. We will use the words as our features (a: 1, beautiful: 2, album: 3, crappy: 4, ugly: 5) and use booleans as our feature values. In other words, the sentences will have the following feature vectors:

            1   2   3   4   5
          +---+---+---+---+---+
Positive: | 1 | 1 | 1 | 0 | 0 |
          +---+---+---+---+---+

          +---+---+---+---+---+
Negative: | 1 | 0 | 1 | 1 | 1 |
          +---+---+---+---+---+

We can now construct the problem using this representation:

problem := golinear.NewProblem()
problem.Add(golinear.TrainingInstance{0, golinear.FromDenseVector([]float64{1, 1, 1, 0, 0})})
problem.Add(golinear.TrainingInstance{1, golinear.FromDenseVector([]float64{1, 0, 1, 1, 1})})

The problem is used to train a linear classifier using a set of parameters to choose the type of solver, constraint violation cost, etc. We will use the default parameters, which train a L2-regularized L2-loss support vector classifier.

param := golinear.DefaultParameters()
model, err := golinear.TrainModel(param, problem)
if err != nil {
	log.Fatal(err)
}

Of course, now we would like to use this model to classify other sentences. For instance:

This is a beautiful book.

We map this sentence to the feature vector that we used during training, simply ignoring words that we did not encounter while training the model:

          +---+---+---+---+---+
????????: | 1 | 1 | 0 | 0 | 0 |
          +---+---+---+---+---+

The Predict method of the model is used to predict the label of this feature vector.

label := model.Predict(golinear.FromDenseVector([]float64{1, 1, 0, 0, 0}))

As expected, the model will predict the sentence to be positive (0).