U Kv f%@sndZddlZddlmZddlZddlZddlm Z ddl m Z ddl m Z Gdd d e ZGd d d ZdS) z4Canonical correlation analysis author: Yichuan Liu N)svd)Model)summary2)multivariate_statscs4eZdZdZd fdd Zd ddZd d ZZS) CanCorra Canonical correlation analysis using singular value decomposition For matrices exog=x and endog=y, find projections x_cancoef and y_cancoef such that: x1 = x * x_cancoef, x1' * x1 is identity matrix y1 = y * y_cancoef, y1' * y1 is identity matrix and the correlation between x1 and y1 is maximized. Attributes ---------- endog : ndarray See Parameters. exog : ndarray See Parameters. cancorr : ndarray The canonical correlation values y_cancoef : ndarray The canonical coefficients for endog x_cancoef : ndarray The canonical coefficients for exog References ---------- .. [*] http://numerical.recipes/whp/notes/CanonCorrBySVD.pdf .. [*] http://www.csun.edu/~ata20315/psy524/docs/Psy524%20Lecture%208%20CC.pdf .. [*] http://www.mathematica-journal.com/2014/06/canonical-correlation-analysis/ :0yE>noneNc s.tt|j||f||d|||dS)N)missinghasconst)superr__init___fit)selfendogexog tolerancer r kwargs __class__D/tmp/pip-unpacked-wheel-2v6byqio/statsmodels/multivariate/cancorr.pyr 0s zCanCorr.__init__csv|jj\}}|jj\}}t||g}t|j}||d}t|j}||d}t|d\}} } | j} | |k} | t | krt d| dd| f| | <t|d\} }}|j}||k} | t | krt d|dd| f|| <t|j | d\}}tfddt t D|_| |ddd|f|_| |jddd|f|_dS)a=Fit the model A ValueError is raised if there are singular values smaller than the tolerance. The treatment of singular arrays might change in future. Parameters ---------- tolerance : float eigenvalue tolerance, values smaller than which is considered 0 rzexog is collinear.Nzendog is collinear.cs g|]}tdt|dqS)rr)maxmin).0isrr Zsz CanCorr._fit..)rshapernprarrayZmeanrTsumlen ValueErrordotrangecancorrZ x_cancoefZ y_cancoef)rrnobsk_yvark_xvarkxyZuxZsxZvxZvx_dsmaskZuyZsyZvyZvy_dsuvrrrr5s.    "z CanCorr._fitcCs|jj\}}|jj\}}t|jd}tjddddddgtt t |dd d d }d}t t |dd d D]B}|d||9}||}||} ||d|| dd} || dd } || } |d| dd d kr t || dd |d| dd } nd} | | d| }t|d| }d|||| }|j||j |df<||j |df<| |j |df<||j |df<||j |df<t jj|| |}||j |df<ql|jjddd }|j |ddf}t|||||d}t||S)aIApproximate F test Perform multivariate statistical tests of the hypothesis that there is no canonical correlation between endog and exog. For each canonical correlation, testing its significance based on Wilks' lambda. Returns ------- CanCorrTestResults instance zCanonical Correlationz Wilks' lambdazNum DFzDen DFzF ValuezPr > Fr)columnsindexrN)rrrr powerr(pdZ DataFramelistr'r$sqrtlocscipystatsfZsfr5valuesrCanCorrTestResults)rr)r*r+Z eigenvalsr>prodrpqrr0Zdf1tZdf2ZlmdFZpvalindstats_mvrrr corr_test_sR  ,  zCanCorr.corr_test)rr N)r)__name__ __module__ __qualname____doc__r rrJ __classcell__rrrrrs *rc@s(eZdZdZddZddZddZdS) rAz Canonical correlation results class Attributes ---------- stats : DataFrame Contain statistical tests results for each canonical correlation stats_mv : DataFrame Contain the multivariate statistical tests results cCs||_||_dSN)r>rI)rr>rIrrrr szCanCorrTestResults.__init__cCs |SrP)summary__str__)rrrrrRszCanCorrTestResults.__str__cCsJt}|d||j|ddi|ddi||j|S)NzCancorr resultsz,Multivariate Statistics and F Approximations)rZSummaryZ add_titleZadd_dfr>Zadd_dictrI)rZsummrrrrQs   zCanCorrTestResults.summaryN)rKrLrMrNr rRrQrrrrrAs rA)rNZnumpyr Z numpy.linalgrr=Zpandasr9Zstatsmodels.base.modelrZstatsmodels.iolibrZmultivariate_olsrrrArrrrs