一种趋势线绘制方式思路¶

• scipy polyBfgs 在一定范围内寻找局部最小值主要是注意参数范围n（优化方式也可以使用np.polynomial.Chebyshev先反出多项式，再optimize.fmin_bfgs速度回快点）
• 聚类多个局部最小值，通过elbow寻找最优k个全局最小（聚类效果如下所示）
• 每个类别中选出最小值，组成趋势

最后趋势线的确定需要通过LinearRegression联合Cluster共同确定¶

    elif mode == K_TREND_KMEAN:
    for xy in other:
        if len(xy) == 2:
            xPred = xy[0]
            yPred = xy[1]
            if xPred.shape[0] == 2:
                pass
            elif xPred.shape[0] == 3:
                d = np.array([np.ravel(xPred), np.ravel(yPred)]).T
                dPd = pd.DataFrame(d, columns=['X', 'Y'])
                dPd.sort(['X'])
                xPred = dPd.iloc[0:2, 0:1].values
                yPred = dPd.iloc[0:2, 1:2].values
            else:
                ZLog.info('KMeans effect')
                select_k = 2
                kmeans = KMeans(n_clusters=select_k, init='random')
                d = np.array([np.ravel(xPred), np.ravel(yPred)]).T

                kmeans.fit(d)
                y_kmeans = kmeans.predict(d)

                dPd = pd.DataFrame(d, columns=['X', 'Y'])
                dPd['Cluster'] = y_kmeans
                dPdMin = dPd.groupby(['Cluster', 'Y', 'X'])['X', 'Y'].min()
                xPred = [dPdMin.loc[x, :].values[0][0] for x in xrange(0, select_k)]
                yPred = [yOrg[x] for x in xPred]
                xPred = np.array(xPred).reshape(-1, 1)
                yPred = np.array(yPred).reshape(-1, 1)

            linreg = LinearRegression()
            linreg.fit(xPred, yPred)
            xPred = np.array(xOrg).reshape(-1, 1)
            yPred = linreg.predict(xPred)
            xPred = np.ravel(xPred)
            yPred = np.ravel(yPred)