Iasi Open stats & predictions

[0]: from __future__ import division [1]: from builtins import object [2]: import numpy as np [3]: import bottleneck as bn [4]: from scipy import special as spl [5]: from ..stats import multivariate_t [6]: class MinCovDet(object): [7]: """Minimum Covariance Determinant Estimator [8]: Parameters [9]: ---------- [10]: store_precision : bool (default=False) [11]: Precompute fast ellipsoid containment test using cholesky [12]: decomposition. This avoids having to recompute this [13]: decomposition repeatedly using :term:`S` and [14]: :term:`precision_`, which requires an additional factorization [15]: for each call. [16]: bias : bool (default=True) [17]: Correct bias in covariance matrix estimate. This results in a [18]: biased estimate of the MCD location parameter (influence function [19]: is not bounded), but in empirical studies was seen to produce [20]: smaller error rates. This option effectively implements the [21]: Fixed Point Algorithm as described in Rousseeuw (1985), [22]: Algorithm 2.1. [23]: support_fraction : float (default=None) [24]: If not None, then this is the proportion of points to be included [25]: in the raw MCD estimate. This is more robust to very small [26]: number of outliers than ``npcs``. If None (default), then it is set [27]: to be 0.5*n_samples by default. [28]: random_state : int (default=None) [29]: The generator used to initialize the centers. If int, random_state [30]: is the seed used by the random number generator; If RandomState [31]: instance, random_state is the random number generator; If None, [32]: the random number generator is the RandomState instance used by [33]: `np.random`. [34]: n_init : int, default=1 [35]: The number of time the algorithm will be run with different [36]: centroid seeds. The final results will be the best output of any [37]: run in terms of inertia. [38]: stacklevel : int, default=2 [39]: Specifies how deeply to interpret relatively-based traceback offsets. [40]: See traceback.print_stack() for information on stacklevel parameter. [41]: verbose : bool, default=False [42]: Allow verbose output. [43]: Attributes [44]: ---------- [45]: raw_data_ : array-like of shape (n_samples, n_features) [46]: Data used when calling `fit`, where n_samples in the number of [47]: samples and n_features is the number of features. [48]: precision_ : array, shape (n_features, n_features) or [49]: (n_features, n_features, mincov_variable) [50]: Generalized inverse matrix corresponding to covariance_ matrix. [51]: covariance_ : array-like, shape (n_features, n_features) or [52]: (n_features, n_features, mincov_variable) [53]: Estimated covariance matrix. [54]: support_ : array-like, shape (n_samples,) [55]: Indices of samples considered as inliers. [56]: center_ : array-like, shape (n_features,) or [57]: (mincov_variable, n_features) [58]: Center(s) of class. [59]: mahal_dist_ : array, shape (n_samples,) [60]: Squared Mahalanobis distances of samples to their closest center. [61]: scale_ : float [62]: Determinant of covariance matrix. [63]: X_train_ : array-containing-training-data [64]: Training data that is used for computing precision matrix (`raw_data_ is None) [65]: References [66]: ---------- [67]: .. [1] Hubert and Vandervieren, "Fast Computation of Minimum Covariance [68]: Determinant Estimates", Computational Statistics & Data Analysis, [69]: Volume 35, Issue 1, 1 July 2001, [70]: Pages 43-64, ISSN 0167-9473, [71]: http://dx.doi.org/10.1016/S0167-9473(01)00086-0. [72]: .. [2] J. P. Thibaux and C. K. Iwayama, [73]: "Elliptically symmetric distributions in statistics, [74]: pattern recognition, and signal processing", [75]: IEEE Trans. on Pattern Analysis and Machine Intelligence, [76]: vol. 15, no. 10, pp. 1035-1044, Oct. 1993. [77]: """ [78]: def __init__(self, [79]: store_precision=False, [80]: support_fraction=None, [81]: random_state=None, [82]: bias=True, [83]: n_init=1, [84]: verbose=False, [85]: stacklevel=2): [86]: self.store_precision = store_precision [87]: self.support_fraction = support_fraction [88]: self.bias = bias [89]: self.n_init = n_init [90]: self.random_state = random_state [91]: self.verbose = verbose [92]: self.stacklevel = stacklevel [93]: def _check_parameters(self): [94]: if not isinstance(self.store_precision, bool): [95]: raise ValueError("store_precision must be bool.") [96]: def _compute_indicators(self): [97]: if self.raw_data_.dtype == np.float32: [98]: _prec_is_finite = bn.in1d(np.isfinite(self.precision_), [True]) [99]: _cov_is_finite = bn.in1d(np.isfinite(self.covariance_), [True]) [100]: _prec_is_pos = bn.in1d(np.linalg.eigvalsh(self.precision_) > 0., [True]) [101]: else: [102]: _prec_is_finite = bn.in1d(np.isfinite(self.precision_), [True]).all() [103]: _cov_is_finite = bn.in1d(np.isfinite(self.covariance_), [True]).all() [104]: _prec_is_pos = bn.in1d(np.linalg.eigvalsh(self.precision_) > 0., [True]).all() [105]: # indicator of invertable precision matrix (i.e., non-singular covariance matrix) [106]: if self.covariance_.ndim == 2: [107]: self.compatible = _prec_is_finite and _cov_is_finite and _prec_is_pos [108]: else: [109]: self.compatible = (bn.in1d(_prec_is_finite, [110]: [True]).all() and [111]: bn.in1d(_cov_is_finite, [112]: [True]).all() and [113]: bn.in1d(_prec_is_pos, [114]: [True]).all()) [115]: def _compute_sub_result(self): [116]: if self.covariance_.ndim == 2: [117]: # get distance from center for each samples [118]: diff = self._centering(self.raw_data_) [119]: # calculate squared mahalanobis distance for each sample [120]: self.mahal_dist_ = np.einsum('ij,jk,ik->i', [121]: diff, [122]: self.precision_, [123]: diff) [124]: self.scale_ = np.sqrt(np.linalg.det(self.covariance_))