Question

我有一个包含连续值（从0到1020）的Python列表，我希望使用K-Means策略将这些值从0到5的序数取舍。

我已经使用新的类sklearn.preprocessing.KBinsDiscretizer来执行该操作：

def descritise_kmeans(python_arr, num_bins):
    X = np.array(python_arr).reshape(-1, 1) 
    est = KBinsDiscretizer(n_bins=num_bins, encode='ordinal', strategy='kmeans')
    est.fit(X)  
    Xt = est.transform(X)
    return Xt

运行此方法时，出现错误：

/usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/sklearn/preprocessing/_discretization.py in transform(self, X)
    262             atol = 1.e-8
    263             eps = atol + rtol * np.abs(Xt[:, jj])
--> 264             Xt[:, jj] = np.digitize(Xt[:, jj] + eps, bin_edges[jj][1:])
    265         np.clip(Xt, 0, self.n_bins_ - 1, out=Xt)
    266 

ValueError: bins must be monotonically increasing or decreasing

仔细观察这一点，似乎numpy.descritize方法是引发错误的方法。这似乎是Sklearn库的错误。

当容器n_bins的数量为6时，将引发错误。但是，当n_bins为5时，它将起作用。

Answer 1

我遇到了类似的问题，但是在设置垃圾箱的值时发现了我的错误。我的代码很简单

bins = np.array([0.0, .33, 66, 1])
data = [0.1, .2, .4, .5, .7, 8]
inds = np.digitize(data, bins, right=False)

在.66之前我错过了一个点，我的垃圾箱不是单调的。尽管它可能不是这个问题的根源，但我希望它能对某人有所帮助。

Answer 2

Makeshift解决方案：
使用以下转换功能编辑sklearns源代码：sklearn / preprocessing / _discretization.py
从版本“ 0.20.2”开始，它位于第237行

def transform(self, X):
    """Discretizes the data.

    Parameters
    ----------
    X : numeric array-like, shape (n_samples, n_features)
        Data to be discretized.

    Returns
    -------
    Xt : numeric array-like or sparse matrix
        Data in the binned space.
    """
    check_is_fitted(self, ["bin_edges_"])

    Xt = check_array(X, copy=True, dtype=FLOAT_DTYPES)
    n_features = self.n_bins_.shape[0]
    if Xt.shape[1] != n_features:
        raise ValueError("Incorrect number of features. Expecting {}, "
                         "received {}.".format(n_features, Xt.shape[1]))


    def ensure_monotic_increase(array):
        """ 
        add small noise to the bin_edges[i] 
        when bin_edges[i] !> bin_edges[i-1]
        """
        noise_overlay = np.zeros(array.shape)
        for i in range(1,len(array)):
            bigger = array[i]>array[i-1] 
            if bigger:
                pass
            else:
                noise_overlay[i] = abs(array[i-1] * 0.0001)
        return(array+noise_overlay)

    bin_edges = self.bin_edges_
    for jj in range(Xt.shape[1]):
        # Values which are close to a bin edge are susceptible to numeric
        # instability. Add eps to X so these values are binned correctly
        # with respect to their decimal truncation. See documentation of
        # numpy.isclose for an explanation of ``rtol`` and ``atol``.
        rtol = 1.e-5
        atol = 1.e-8
        eps = atol + rtol * np.abs(Xt[:, jj])

        old_bin_edges = bin_edges[jj][1:]

        try:
            Xt[:, jj] = np.digitize(Xt[:, jj] + eps, old_bin_edges)
        except ValueError:
            new_bin_edges = ensure_monotic_increase(old_bin_edges)
            #print(old_bin_edges)
            #print(new_bin_edges)
            try:
                Xt[:, jj] = np.digitize(Xt[:, jj] + eps, new_bin_edges)
            except:
                raise

    np.clip(Xt, 0, self.n_bins_ - 1, out=Xt)

    if self.encode == 'ordinal':
        return Xt

    return self._encoder.transform(Xt)

（我遇到的）问题
垃圾桶边缘彼此之间过于靠近。可能，由于某种浮点错误，先前的bin边缘最终大于下一个bin边缘。
当打印边缘时（取消注释上述功能中的打印语句），前两个bin边缘可观察到彼此相等。印刷的bin_edges是：

[-0.1025641  -0.1025641   0.82793522]             # ValueError 
[-0.1025641  -0.10255385  0.82793522]             # After fix
[0.2075     0.2075     0.88798077]                # ValueError 
[0.2075     0.20752075 0.88798077]                # After fix
[ 0.7899066   0.7899066  24.31967669]             # ValueError 
[ 0.7899066   0.78998559 24.31967669]             # After fix
[5.47545572e-18 5.47545572e-18 2.36842105e-01]    # ValueError 
[5.47545572e-18 5.47600326e-18 2.36842105e-01]    # After fix
[5.47545572e-18 5.47545572e-18 2.82894737e-01]    # ValueError 
[5.47545572e-18 5.47600326e-18 2.82894737e-01]    # After fix
[-0.46762302 -0.46762302 -0.00969465]             # ValueError 
[-0.46762302 -0.46757626 -0.00969465]             # After fix

sklearn：垃圾箱必须单调增加或减少

2 个答案: