基于Pandas Groupby实现列作为X轴的时序数据可视化优化

你现在的嵌套循环确实可以优化，利用pandas的groupby功能，我们可以把相同cluster和cat的时间序列归为一组，一次性绘制整组的序列，避免逐行循环，代码会更简洁高效。

具体实现步骤： #

首先，先确认我们需要的核心数据：

• 时间序列的列：就是多级索引中('frac_norm',2,1)对应的所有列（也就是DataFrame中从第2列到倒数第2列）
• X轴的取值：这些时间序列列的第二个索引层级month_rel的值（比如-5, -4, ..., 5）

然后，我们可以按cluster和cat的组合来分组，每个组内的所有序列共享相同的颜色和线型，直接批量绘制：

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

# 先生成你的DataFrame（和你提供的代码一致）
vals = np.array([['-1b', 0.7606747496046389, 0.7898535589129476, 0.959409594095941, 0.8677546569280126, 0.7901020186672455, 1.0, 0.5887286145588728, 0.7190726452719073, 0.6955719557195572, 0.6476962793343348, 0.6563233814156323, 4], ['-1b', 0.5804363905325444, 0.5467611336032389, 0.7134300126103406, 0.7420329670329671, 0.8021978021978022, 0.389957264957265, 0.861451048951049, 0.6517857142857143, 0.798265460030166, 0.4763049450549451, 0.8960720130932898, 0], ['-1b', 0.9461875843454791, 0.49953095684803, 0.8348848603625673, 0.8257715338553662, 0.7540183696900115, 0.6782302664655606, 0.43069179143004643, 0.35398860398860393, 0.33376068376068374, 0.28475935828877, 0.260501012145749, 2], ['-1b', 0.8267008985879333, 0.8120300751879698, 0.7651821862348178, 0.680161943319838, 0.7634749524413443, 0.8026315789473684, 1.0, 0.7801857585139319, 0.8440191387559809, 0.8686980609418281, 0.7226720647773278, 4], ['-1b', 1.0, 0.955656108597285, 0.6776470588235294, 0.9115556882651537, 0.766129636568003, 0.7437589670014347, 0.6179800221975582, 0.6065359477124183, 0.715527950310559, 0.6149019607843138, 0.4822670674109059, 3], ['-1b', 0.8700564971751412, 0.3719806763285024, 0.6402116402116402, 0.8351851851851851, 0.6731078904991948, 0.5365853658536585, 1.0, 0.8502415458937197, 0.55119825708061, 0.873015873015873, 0.6355555555555555, 4], ['-1b', 0.8196997807387418, 0.879872907246731, 0.961178456344944, 1.0, 0.8554654738607772, 0.8278240873814314, 0.8271388025408839, 0.3040112596762843, 0.5749778172138421, 0.47399605003291634, 0.35893441346004046, 3], ['-1b', 1.0, 0.9958592132505176, 0.7384176764076977, 0.9912165129556433, 0.8543355440923606, 0.9365176566646254, 0.9103471520053926, 0.8832054560954816, 0.9877955758962623, 0.6994328922495274, 0.8150724637681159, 4], ['-1b', 0.9490474080638015, 0.9902376128200405, 0.8995240613432046, 1.0, 0.9610655737704917, 0.837029893924783, 0.8351136964569011, 0.759142496847415, 0.7497267759562841, 0.8869130313976105, 0.9369612979550449, 4], ['-1b', 0.7805383022774327, 0.7021791767554478, 0.5686274509803921, 1.0, 0.6013824884792627, 0.7891156462585033, 0.0, 0.07142857142857142, 0.0, 0.11196911196911197, 0.07397959183673469, 2]], dtype=object)
cols = pd.MultiIndex.from_tuples([( 'cat', ''), (('frac_norm', 2, 1), -5), (('frac_norm', 2, 1), -4), (('frac_norm', 2, 1), -3), (('frac_norm', 2, 1), -2), (('frac_norm', 2, 1), -1), (('frac_norm', 2, 1), 0), (('frac_norm', 2, 1), 1), (('frac_norm', 2, 1), 2), (('frac_norm', 2, 1), 3), (('frac_norm', 2, 1), 4), (('frac_norm', 2, 1), 5), ( 'cluster', '')], names=[None, 'month_rel'])
idx = pd.MultiIndex.from_tuples([(3414845, 4232621), (4369232, 3370279), ( 22771, 3795428), (2660226, 3126314), (4154510, 4348009), (2860801, 164553), ( 120577, 3480468), (6692132, 5095003), (2515737, 4263756), ( 707596, 2856619)], names=['user1', 'user2'])
df = pd.DataFrame(vals, columns=cols, index=idx)

# 准备颜色和线型字典（和你原来的一致）
colors = ['r','g','b','c','y','k']
lnst = ['-','--']
cats = np.sort(df['cat'].unique())
clusters = np.sort(df['cluster'].unique())
colordict = dict(zip(clusters, colors))
lnstdict = dict(zip(cats,lnst))

fig, ax = plt.subplots()

# 提取时间序列列和X轴值
ts_cols = df.columns[1:-1]  # 跳过第一个cat列和最后一个cluster列
x_vals = ts_cols.get_level_values(1)  # 获取month_rel作为X轴

# 按cluster和cat分组，批量绘制
for (clus_val, cat_val), group in df.groupby(['cluster', 'cat']):
    # 把组内的时间序列转置，这样每一列对应一个用户的序列，plot会自动绘制所有列
    ax.plot(x_vals, group[ts_cols].T, color=colordict[clus_val], linestyle=lnstdict[cat_val])

plt.show()

为什么这样更高效？ #

• 原来的代码是两层嵌套循环（先遍历cluster，再遍历每行），而现在只需要一层循环遍历分组
• group[ts_cols].T把每行的序列转成列，matplotlib的plot函数可以一次性绘制所有列，避免了逐行调用plot的开销
• 代码逻辑更清晰，直接按分组逻辑处理，和你的可视化需求（同cluster同颜色，同cat同线型）完全匹配

这样绘制出来的结果和你原来逐行绘制的效果完全一致，但代码更简洁，数据量大的时候性能提升会更明显。

内容的提问来源于stack exchange，提问作者irene