|
| 1 | +"""ADASYN over sampling algorithm. |
| 2 | +
|
| 3 | +See more in imblearn.over_sampling.ADASYN |
| 4 | +original authors: |
| 5 | +# Guillaume Lemaitre <g.lemaitre58@gmail.com> |
| 6 | +# Fernando Nogueira |
| 7 | +# Christos Aridas |
| 8 | +# Dzianis Dudnik |
| 9 | +# License: MIT |
| 10 | +""" |
| 11 | + |
| 12 | +import numpy as np |
| 13 | +from sklearn.utils import check_random_state |
| 14 | + |
| 15 | +from aeon.transformations.collection.imbalance._smote import SMOTE |
| 16 | + |
| 17 | +__maintainer__ = ["TonyBagnall"] |
| 18 | +__all__ = ["ADASYN"] |
| 19 | + |
| 20 | + |
| 21 | +class ADASYN(SMOTE): |
| 22 | + """ |
| 23 | + Adaptive Synthetic Sampling (ADASYN) over-sampler. |
| 24 | +
|
| 25 | + Generates synthetic samples for the minority class based on local data |
| 26 | + distribution. ADASYN extends SMOTE by adapting the number of synthetic samples |
| 27 | + according to the density of the minority class: more samples are generated for |
| 28 | + minority samples that are harder to learn (i.e., surrounded by more majority |
| 29 | + samples). |
| 30 | +
|
| 31 | + This implementation is adapted from imbalanced-learn's |
| 32 | + `imblearn.over_sampling.ADASYN`. |
| 33 | +
|
| 34 | + Parameters |
| 35 | + ---------- |
| 36 | + random_state : int or None, optional (default=None) |
| 37 | + Random seed for reproducibility. |
| 38 | + k_neighbors : int, optional (default=5) |
| 39 | + Number of nearest neighbours used to construct synthetic samples. |
| 40 | +
|
| 41 | + References |
| 42 | + ---------- |
| 43 | + .. [1] He, H., Bai, Y., Garcia, E. A., & Li, S. (2008). |
| 44 | + ADASYN: Adaptive synthetic sampling approach for imbalanced learning. |
| 45 | + In IEEE International Joint Conference on Neural Networks, pp. 1322-1328. |
| 46 | + https://doi.org/10.1109/IJCNN.2008.4633969 |
| 47 | +
|
| 48 | + Examples |
| 49 | + -------- |
| 50 | + >>> from aeon.transformations.collection.imbalance import ADASYN |
| 51 | + >>> import numpy as np |
| 52 | + >>> X = np.random.random(size=(100,1,50)) |
| 53 | + >>> y = np.array([0] * 90 + [1] * 10) |
| 54 | + >>> sampler = ADASYN(random_state=49) |
| 55 | + >>> X_res, y_res = sampler.fit_transform(X, y) |
| 56 | + """ |
| 57 | + |
| 58 | + def __init__(self, random_state=None, k_neighbors=5): |
| 59 | + super().__init__(random_state=random_state, k_neighbors=k_neighbors) |
| 60 | + |
| 61 | + def _transform(self, X, y=None): |
| 62 | + X = np.squeeze(X, axis=1) |
| 63 | + random_state = check_random_state(self.random_state) |
| 64 | + X_resampled = [X.copy()] |
| 65 | + y_resampled = [y.copy()] |
| 66 | + |
| 67 | + # got the minority class label and the number needs to be generated |
| 68 | + for class_sample, n_samples in self.sampling_strategy_.items(): |
| 69 | + if n_samples == 0: |
| 70 | + continue |
| 71 | + target_class_indices = np.flatnonzero(y == class_sample) |
| 72 | + X_class = X[target_class_indices] |
| 73 | + |
| 74 | + self.nn_.fit(X) |
| 75 | + nns = self.nn_.kneighbors(X_class, return_distance=False)[:, 1:] |
| 76 | + # The ratio is computed using a one-vs-rest manner. Using majority |
| 77 | + # in multi-class would lead to slightly different results at the |
| 78 | + # cost of introducing a new parameter. |
| 79 | + n_neighbors = self.nn_.n_neighbors - 1 |
| 80 | + ratio_nn = np.sum(y[nns] != class_sample, axis=1) / n_neighbors |
| 81 | + if not np.sum(ratio_nn): |
| 82 | + raise RuntimeError( |
| 83 | + "Not any neighbours belong to the majority" |
| 84 | + " class. This case will induce a NaN case" |
| 85 | + " with a division by zero. ADASYN is not" |
| 86 | + " suited for this specific dataset." |
| 87 | + " Use SMOTE instead." |
| 88 | + ) |
| 89 | + ratio_nn /= np.sum(ratio_nn) |
| 90 | + n_samples_generate = np.rint(ratio_nn * n_samples).astype(int) |
| 91 | + # rounding may cause new amount for n_samples |
| 92 | + n_samples = np.sum(n_samples_generate) |
| 93 | + if not n_samples: |
| 94 | + raise ValueError( |
| 95 | + "No samples will be generated with the provided ratio settings." |
| 96 | + ) |
| 97 | + |
| 98 | + # the nearest neighbors need to be fitted only on the current class |
| 99 | + # to find the class NN to generate new samples |
| 100 | + self.nn_.fit(X_class) |
| 101 | + nns = self.nn_.kneighbors(X_class, return_distance=False)[:, 1:] |
| 102 | + |
| 103 | + enumerated_class_indices = np.arange(len(target_class_indices)) |
| 104 | + rows = np.repeat(enumerated_class_indices, n_samples_generate) |
| 105 | + cols = random_state.choice(n_neighbors, size=n_samples) |
| 106 | + diffs = X_class[nns[rows, cols]] - X_class[rows] |
| 107 | + steps = random_state.uniform(size=(n_samples, 1)) |
| 108 | + X_new = X_class[rows] + steps * diffs |
| 109 | + |
| 110 | + X_new = X_new.astype(X.dtype) |
| 111 | + y_new = np.full(n_samples, fill_value=class_sample, dtype=y.dtype) |
| 112 | + X_resampled.append(X_new) |
| 113 | + y_resampled.append(y_new) |
| 114 | + X_resampled = np.vstack(X_resampled) |
| 115 | + y_resampled = np.hstack(y_resampled) |
| 116 | + |
| 117 | + X_resampled = X_resampled[:, np.newaxis, :] |
| 118 | + return X_resampled, y_resampled |
0 commit comments