如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

huangapple go评论68阅读模式
英文:

How to subsample a pandas df so that its variable distribution fits another distribution?

问题

我有两张天文数据表,df_jpasdf_gaia。它们是包含星系红移 z 等信息的目录。我可以绘制这两个目录中红移的分布,如下所示:

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

现在我想创建一个子采样的 df_jpas,使其在红移范围 0.8<z<2.3 内的分布尽可能接近于 df_gaia 的分布,即我想要:

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

我应该如何做?

英文:

I am having 2 astronomical data tables, df_jpas and df_gaia. They are catalogues of galaxies containing among others the red-shifts z of the galaxies. I can plot the distribution of the redshifts of the 2 catalogs and it looks like this:

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

What I want now is to create a subsampled df_jpas, so that its distribution of z is as close as possible to the distribution of df_gaia within the z-range 0.8<z<2.3, means I want:

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

How do I do this?

答案1

得分: 1

这是一个解决方案。

首先,让我们将数据框切割成所需的 z 范围:

left_z_edge, right_z_edge = 0.8, 2.3
stepsize=0.02

df_jpas = df_jpas[(df_jpas.z>left_z_edge)&(df_jpas.z<right_z_edge)]
df_gaia = df_gaia[(df_gaia.z>left_z_edge)&(df_gaia.z<right_z_edge)]

接下来,我们想要计算这些数据框的分布(或直方图):

jpas_hist, jpas_bin_edges = np.histogram(df_jpas.z, bins = np.arange(left_z_edge,right_z_edge + stepsize, step=stepsize))
jpas_bin_centers = (jpas_bin_edges + stepsize/2)[:-1] # instead of using the bin edges I create the bin centers and use them later

gaia_hist, gaia_bin_edges = np.histogram(df_gaia.z, bins = np.arange(left_z_edge,right_z_edge + stepsize, step=stepsize))
gaia_bin_centers = (gaia_bin_edges + stepsize/2)[:-1]

完成这一步后,代码的关键部分来了 - 将 gaia_hist 除以 jpas_hist 给出了在特定 z-bin 中存在银河系的概率,这个概率是我们将用于子采样的概率:

jpas_occup_prob = gaia_hist/jpas_hist

接下来,我们创建一个要应用于 df_jpas 数据框的函数,它创建了一个额外的列,其中包含一个标志,指示是否应该"激活"(保留或删除)该特定星系,以提供所需的分布:

def activate_QSO(z_val):
    idx = (np.abs(jpas_bin_centers - z_val)).argmin() # find the closest desscrite z-value to the z of the current QSO
    ocup_prob = jpas_occup_prob[idx] # assign to this entry the its probability of occupation
    activation_flag = int(np.random.random() < ocup_prob)# either activate (1) or not (0) this QSO depending on the probability from above
    return(activation_flag)

使用此标志,我们可以绘制在此列中包含 1 的所有星系,从而获得所需的分布:

plt.hist(df_jpas[df_jpas.activation_flag==1].z, bins=100, alpha=0.5, label='jpas mock, subsampled')
plt.hist(df_gaia.z, bins=100, alpha=0.5, label='GAIA QSO')
plt.ylabel('N(z)')
plt.xlabel('z')
plt.legend()
plt.show()

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

英文:

Here is a solution.

Let's first cut the dataframes into the desired z-range:

left_z_edge, right_z_edge = 0.8, 2.3
stepsize=0.02

df_jpas = df_jpas[(df_jpas.z&gt;left_z_edge)&amp;(df_jpas.z&lt;right_z_edge)]
df_gaia = df_gaia[(df_gaia.z&gt;left_z_edge)&amp;(df_gaia.z&lt;right_z_edge)]

Next, we want to calculate the distributions (or histograms) of these dataframes:

jpas_hist, jpas_bin_edges = np.histogram(df_jpas.z, bins = np.arange(left_z_edge,right_z_edge + stepsize, step=stepsize))
jpas_bin_centers = (jpas_bin_edges + stepsize/2)[:-1] # instead of using the bin edges I create the bin centers and use them later

gaia_hist, gaia_bin_edges = np.histogram(df_gaia.z, bins = np.arange(left_z_edge,right_z_edge + stepsize, step=stepsize))
gaia_bin_centers = (gaia_bin_edges + stepsize/2)[:-1]

After this is done comes the critical part of the code - dividing gaia_hist by jpas_hist gives us the probability of a galaxy existing in the particular z-bin and this probability is what we will use for subsampling:

jpas_occup_prob = gaia_hist/jpas_hist

Next, we create a function to be applied on the df_jpas dataframe, it creates an additional column that contains a flag if this particular galaxy should be "activated" (dropped or remained) to provide the desired distribution:

def activate_QSO(z_val):
    idx = (np.abs(jpas_bin_centers - z_val)).argmin() # find the closest desscrite z-value to the z of the current QSO
    ocup_prob = jpas_occup_prob[idx] # assign to this entry the its probability of occupation
    activation_flag = int(np.random.random() &lt; ocup_prob)# either activate (1) or not (0) this QSO depending on the probability from above
    return(activation_flag)

df_jpas[&#39;activation_flag&#39;] = df_jpas[&#39;z&#39;].apply(activate_QSO)

Using this flag, we can plot all galaxies containing 1 in this column which gives us the desired distribution:

plt.hist(df_jpas[df_jpas.activation_flag==1].z, bins=100, alpha=0.5, label=&#39;jpas mock, subsampled&#39;)
plt.hist(df_gaia.z, bins=100, alpha=0.5, label=&#39;GAIA QSO&#39;)
plt.ylabel(&#39;N(z)&#39;)
plt.xlabel(&#39;z&#39;)
plt.legend()
plt.show()

如何对 Pandas 数据框进行子采样,以使其变量分布适应另一个分布?

huangapple
  • 本文由 发表于 2023年5月21日 00:54:56
  • 转载请务必保留本文链接:https://go.coder-hub.com/76296353.html
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定