some snippets

# Decile chart, LIFT, KS 
def getDeciles(models, X, y, n=10):
    probs = predict_probas(gbms,X)[:,1]
    cuts =  pd.qcut(probs,n)
    df = pd.DataFrame({'DECILE':cuts,"P":probs,'Y':y}).groupby('DECILE').Y.value_counts().unstack(fill_value=0)[::-1]
    df.index = [ f"{i+1} {df.index[i]}" for i in range(n) ]
    df.columns = ['ZEROS','ONES']
    df['POPULATION'] = df.ZEROS+df.ONES
    df['EVENT_RATE'] = df.ONES/df.POPULATION
    df['CUMULATIVE_EVENT_RATE'] = df.ONES.cumsum()/df.POPULATION.cumsum()
    df['EVENTS_CAPTURED'] = df.ONES.cumsum() / df.ONES.sum()
    df['NON_EVENT_CAPTURED'] = df.ZEROS.cumsum() / df.ZEROS.sum()
    df['LIFT'] = df.EVENTS_CAPTURED*df.POPULATION.sum() / df.POPULATION.cumsum()
    df['KS'] = df.EVENTS_CAPTURED - df.NON_EVENT_CAPTURED
    return df


def calibration_plot(yi, pi, bins:int=10, plot=True, show_trend=True, show_bin_size = False, figsize=(8,5) ):
    """
    yi : Actual Targets
    pi : predicted probabilities
    bins : Number of bins (Default 10)
    plot : Give a scatter plot
           else return fraction of positives, mean of probabilities, sizes of bin
    show_trend : show trend using lines (Default True)
    show_bin_size : Show overlayed figure with size of bins (Default False)
    figsize : size of figure (Default (8,5))
    """
    bins = np.linspace(0,1,bins+1)
    labels = [ f"{s:.2f}-{e:.2f}".replace("0.",".") for s,e in zip(bins[:-1],bins[1:]) ]
    data = pd.DataFrame( {"yi_frac":yi, "pi_mean":pi, "cnt":yi, "bi":pd.cut(pi,bins=bins,labels=labels )} )
    data = data.groupby('bi').agg({"pi_mean":np.mean, "yi_frac": lambda x: sum(x)/len(x), "cnt": len})
    if plot==True:
        plt.figure(figsize=figsize)
        fig = plt.scatter( data.index, data.yi_frac, c="r" )
        plt.xlabel("Probability Bin");plt.ylabel("Fraction of Positives");plt.title("Calibration Plot")
        if show_trend: plt.plot( list(data.index), data.yi_frac, alpha=0.7 )
        if show_bin_size:
            plt.bar( data.index, data.cnt/sum(data.cnt), alpha=0.3)
            for idx,cnt,frac in zip(data.index,data.cnt,data.cnt/sum(data.cnt)): plt.annotate(cnt,(idx,frac+0.02),rotation=90)
        plt.xticks( data.index, rotation=45)
        return fig 
    else: return data