Linear Model for Multi-label Classification.

This guide will take you through how LibMultiLabel can be used to train a linear classifier in python scripts.

For this guide, we will only need the linear module:

import libmultilabel.linear as linear

To start, we need to read and preprocess the input data:

datasets = linear.load_dataset("txt", "data/rcv1/train.txt", "data/rcv1/test.txt")
preprocessor = linear.Preprocessor()
preprocessor.fit(datasets)
datasets = preprocessor.transform(datasets)

# Alternatively, you can fit and transform the dataset at the same time:
# datasets = preprocessor.fit_transform(datasets)

Note

The fit method calculates the parameters (such as label mapping) that will be used for later transformation from the given dataset.

The transform method applies the learned parameters to the given dataset and returns a new transformed dataset.

The preprocessor handles many issues such as: mapping the labels into indices and transforming textual data to numerical data. The loaded dataset has the structure:

{
    'train': {
        'x': # training features
        'y': # training labels
    },
    'test': {
        'x': # test features
        'y': # test labels
    },
}

Next we train the model:

model = linear.train_1vsrest(datasets["train"]["y"], datasets["train"]["x"], "")

The third argument is the options string for LibLinear. We may leave it as the default for now.

Once we have the model, we may predict with it:

preds = linear.predict_values(model, datasets["test"]["x"])

preds holds the decision values, i.e. the raw values outputted by the model. To transform it into predictions, you can apply the API get_positive_labels to get predicted labels and their corresponding scores by using label_mapping in preprocessor and preds from the last step.

pred_labels, pred_scores = linear.get_positive_labels(preds, preprocessor.label_mapping)

We now have the labels (pred_labels) and scores (pred_scores). You can use the following code to save the prediction to a list.

prediction = []
for label, score in zip(pred_labels, pred_scores):
    prediction.append([f"{i}:{s:.4}" for i, s in zip(label, score)])

The first instance looks like:

>>> print(prediction[0])
...
    ['GCAT:1.345', 'GSPO:1.519']

To see how well we performed, we may want to check various metrics with the test set. Since the dataset we loaded are stored as scipy.sparse.csr_matrix, we will first transform the dataset to np.array.

target = datasets["test"]["y"].toarray()

Then we will compute the metrics with compute_metrics.

metrics = linear.compute_metrics(
    preds,
    target,
    monitor_metrics=["Macro-F1", "Micro-F1", "P@1", "P@3", "P@5"],
)

print(metrics)

The results will look similar to:

{'Macro-F1': 0.5171960144875225, 'Micro-F1': 0.8008124243391698, 'P@1': 0.9573153795447128, 'P@3': 0.799074151109632, 'P@5': 0.5579924865442584}