如何计算具有多列的多个 Pandas 数据框的准确度

Question

我有多个熊猫数据框如下：

data1 = {'1':[4], '2':[2], '3':[6]}
baseline = pd.DataFrame(data1)

 # baseline output  
   1  2  3
0  4  2  6

data2 = {'1':[3], '2':[5], '5':[5]}
forecast1 = pd.DataFrame(data2)

# forecast1 output
   1  2  5
0  3  5  5

data3 = {'1':[2], '3':[4], '5':[5], '6':[2]}
forecast2 = pd.DataFrame(data3)

# forecast2 output
   1  3  5  6
0  2  4  5  2

与基线数据框（即基线 vs 预测 1 和基线 vs 预测 2）相比，我如何计算预测 1和预测 2（分别）的准确性（或混淆矩阵）？

另请注意，与基线数据框相比，预测 1 和预测 2 可能有一些额外的列。因此，精度计算需要考虑可用列的数量并处理额外的列。有没有办法处理这种情况？

这些数据帧是我正在做的数据清理的结果，这就是为什么其中一些数据帧几乎没有基线数据帧中不可用的额外列。

我感谢您的帮助。

谢谢。

Answer 1

print(baseline.columns)
print(forecast1.columns)
print(forecast2.columns)

Index(['1', '2', '3'], dtype='object')
Index(['1', '2', '5'], dtype='object')
Index(['1', '3', '5', '6'], dtype='object')

您可以取列的交集来找出基线和预测之间哪些列是通用的，然后在这些列上应用 precision_score。

from sklearn.metrics import accuracy_score

common_columns = list(set(baseline.columns).intersection(forecast1.columns))

avg_acc = 0.0
for c in common_columns:
    c_acc = accuracy_score(baseline[c], forecast1[c])
    print(f'Column {c} acc: {c_acc}')
    avg_acc += c_acc/len(common_columns)

print(avg_acc)

编写一个函数来获取基线和预测以提供准确性。

from sklearn.metrics import accuracy_score

def calc_acc(baseline, forecast1):
    common_columns = list(set(baseline.columns).intersection(forecast1.columns))

    avg_acc = 0.0
    for c in common_columns:
        c_acc = accuracy_score(baseline[c], forecast1[c])
        print(f'Column {c} acc: {c_acc}')
        avg_acc += c_acc/len(common_columns)

    print(avg_acc)
    return avg_acc

from sklearn.metrics import accuracy_score

def calc_acc(baseline, forecast1):
    penalize = True
    common_columns = list(set(baseline.columns).intersection(forecast1.columns))

    avg_acc = 0.0
    for c in common_columns:
        c_acc = accuracy_score(baseline[c], forecast1[c])
        print(f'Column {c} acc: {c_acc}')
        if penalize:
            div = len(common_columns) + abs(len(forecast1.columns) - len(baseline.columns)) # it will penalize for both having more or less columns than baseline, you can change it based on your needs
            avg_acc += c_acc/div
        else:
            avg_acc += c_acc/len(common_columns)

    print(avg_acc)
    return avg_acc

对于回归尝试平均绝对误差，误差越低预测越好。

from sklearn.metrics import accuracy_score, mean_absolute_error

def calc_acc(baseline, forecast1):
    penalize = True
    common_columns = list(set(baseline.columns).intersection(forecast1.columns))

    avg_acc = 0.0
    for c in common_columns:
        c_acc = mean_absolute_error(baseline[c], forecast1[c])
        print(f'Column {c} mean absolute error: {c_acc}')
        if penalize:
            div = len(common_columns) + abs(len(forecast1.columns) - len(baseline.columns)) # it will penalize for both having more or less columns than baseline, you can change it based on your needs
            avg_acc += c_acc/div
        else:
            avg_acc += c_acc/len(common_columns)

    print(avg_acc)
    return avg_acc

通常，平均正确率约为 100% - 平均误差。因此，您可以从 100% 中减去误差。

def perc(a_list, b_list):
    ans = 0.0

    for i in range(len(a_list)):
        ans += (1. - abs(a_list[i]-b_list[i])/a_list[i])

    return ans

from sklearn.metrics import accuracy_score, mean_absolute_error

def calc_acc(baseline, forecast1):
    penalize = True
    common_columns = list(set(baseline.columns).intersection(forecast1.columns))

    avg_acc = 0.0
    for c in common_columns:
        c_acc = perc(baseline[c], forecast1[c])
        print(f'Column {c} mean percentange correct: {c_acc}')
        if penalize:
            div = len(common_columns) + abs(len(forecast1.columns) - len(baseline.columns)) # it will penalize for both having more or less columns than baseline, you can change it based on your needs
            avg_acc += c_acc/div
        else:
            avg_acc += c_acc/len(common_columns)

    print(avg_acc)
    return avg_acc