Creating Balanced Dataset Using Scikits

Question

I have a classic User-Item dataset where each row (i.e., (user, item)) indicates the action of a user clicking/selecting an item. Now, the dataset only provides positive samples and does not specifically indicate whether a user has disliked an item. In order to create a balanced dataset, I would like to create random negative samples (for instance randomly pick a set of items which the user has never clicked). Of course, I can achieve this by writing a program; however, I found this library in python Scikits called imbalanced-learn, which seems to provide various sampling techniques. Can someone provide a small code-snippet that can use this library (or some other python library) to achieve this?.

Answer 1

If I understood correctly, you want to invent new negative samples to have a balanced dataset, because your current dataset only have positive samples.

However, the library you mentioned only re-samples a dataset where all classes are present. The sample creation process is by definition data-specific, as there can be impossible attribute combinations, etc.

If you want to blindly create combination of attribute values, you can use the stuff in module random:

• To create integer attributes:

  def random_integer(min_val, max_val):
      return randint.randint(0, 9))
  

• To create real-valued attributes:

  def random_float(min_val, max_val):
      return random.uniform(min_val, max_val)
  

• To create discrete attributes:

  def random_discrete(value_list):
      return random.choice(value_list)
  

And then, with a loop, generate your negative samples in a loop:

    def generate_negative_sample():
        return {'attr1': random_discrete(['a', 'b', 'c']),
                'attr2': random_float(-10., 10.),
                'attr3': random_integer(0, 4)}

    random_samples = [generate_negative_sample() for _ in range(1000)]

Answer 2

Simple Version

# generate 2d classification dataset

X, y = make_blobs(n_samples=100, centers=3, n_features=2)

Moons Classification Problem (swirl like data)

# generate 2d classification dataset

X, y = make_moons(n_samples=100, noise=0.1)

The make_circles generates a binary classification problem with datasets that fall into concentric circles and few more are also there.

References

Scikit documentation

Imbalanced learn