ClassificationScore
ClassificationScore
计算分类指标来评估预测数据相对于真实数据的质量。
与机器学习部分中的其他方法不同，这个指标适用于真实值的向量，而不是预测值的向量。
vector vector::ClassificationScore(
const matrix& pred_scores, // 包含每一类概率分布的矩阵
ENUM_CLASSIFICATION_METRIC metric // 指标类型
ENUM_AVERAGE_MODE mode // 平均模式
);
vector vector::ClassificationScore(
const matrix& pred_scores, // 包含每一类概率分布的矩阵
ENUM_CLASSIFICATION_METRIC metric // 指标类型
int param // 附加参数
);参数
- pred_scores
[in] 包含一组水平向量和每一类概率的矩阵。矩阵行数应与真值向量的大小相对应。
- metric
[in] ENUM_CLASSIFICATION_METRIC枚举的指标类型。使用CLASSIFICATION_TOP_K_ACCURACY， CLASSIFICATION_AVERAGE_PRECISION和CLASSIFICATION_ROC_AUC值。
- mode
[in] ENUM_AVERAGE_MODE枚举的平均模式。用于CLASSIFICATION_AVERAGE_PRECISION和CLASSIFICATION_ROC_AUC指标。
- param
[in] 对于CLASSIFICATION_TOP_K_ACCURACY指标，应指定整数K值来替代平均模式。
返回值
包含计算指标的向量。在AVERAGE_NONE平均模式的情况下，向量包含每个类别的指标值，而不进行平均。（例如，在二元分类的情况下，这将是分别用于’false’和’true’的两个指标）。
关于平均模式的说明
AVERAGE_BINARY只对二元分类有意义。
AVERAGE_MICRO — 通过将标签指标矩阵的每个元素视为一个标签来计算全局指标。标签指标矩阵是指一个矩阵，其中包含每个标签的一组概率。
AVERAGE_MACRO — 计算每个标签的指标并找到其未加权平均值。这没有考虑标签不平衡。
AVERAGE_WEIGHTED — 计算每个标签的指标并找到其按支持度加权的平均值（每个标签的真实实例数）。
注意
在二元分类的情况下，我们不仅可以输入一个n x 2矩阵，其中第一列包含负标签的概率，第二列包含正标签的概率，还可以输入一个由一列正概率组成的矩阵。这是因为二元分类模型可以返回正标签的两个概率或一个概率。
示例：
vector y_true={7,2,1,0,4,1,4,9,5,9,0,6,9,0,1,5,9,7,3,4,8,4,2,7,6,8,4,2,3,6};
//vector y_pred={7,2,1,0,4,1,4,9,5,9,0,6,9,0,1,5,9,7,3,4,2,9,4,9,5,9,2,7,7,0};
//--- label scores 0 1 2 3 4 5 6 7 8 9 true pred
matrix y_scores={{0.000109, 0.000186, 0.000449, 0.000052, 0.000002, 0.000022, 0.000005, 0.998059, 0.000010, 0.001104}, // 7 7
{0.000091, 0.081956, 0.916816, 0.001106, 0.000006, 0.000002, 0.000001, 0.000000, 0.000021, 0.000000}, // 2 2
{0.000108, 0.972863, 0.003600, 0.000021, 0.010479, 0.000015, 0.000131, 0.010385, 0.002339, 0.000060}, // 1 1
{0.925425, 0.000080, 0.002913, 0.000057, 0.000274, 0.000638, 0.063529, 0.000316, 0.000095, 0.006673}, // 0 0
{0.000060, 0.000126, 0.000006, 0.000000, 0.993513, 0.000000, 0.000003, 0.000222, 0.000001, 0.006069}, // 4 4
{0.000016, 0.982124, 0.000045, 0.000002, 0.008445, 0.000001, 0.000005, 0.009230, 0.000120, 0.000013}, // 1 1
{0.000000, 0.000040, 0.000001, 0.000000, 0.989395, 0.000167, 0.000004, 0.000070, 0.000177, 0.010146}, // 4 4
{0.000795, 0.002938, 0.023447, 0.007418, 0.021838, 0.002476, 0.000260, 0.047551, 0.000082, 0.893194}, // 9 9
{0.000091, 0.000226, 0.000038, 0.000007, 0.000048, 0.854910, 0.068644, 0.000080, 0.001097, 0.074860}, // 5 5
{0.000000, 0.000000, 0.000000, 0.000000, 0.003004, 0.000000, 0.000000, 0.000035, 0.000000, 0.996960}, // 9 9
{0.998856, 0.000009, 0.000976, 0.000002, 0.000000, 0.000013, 0.000131, 0.000006, 0.000000, 0.000007}, // 0 0
{0.000178, 0.000446, 0.000326, 0.000033, 0.000193, 0.000071, 0.998403, 0.000015, 0.000328, 0.000007}, // 6 6
{0.000005, 0.000016, 0.000153, 0.000045, 0.004110, 0.000012, 0.000015, 0.000031, 0.000076, 0.995537}, // 9 9
{0.994188, 0.000003, 0.002584, 0.000005, 0.000005, 0.000100, 0.000739, 0.001473, 0.000038, 0.000864}, // 0 0
{0.000173, 0.990569, 0.000792, 0.000040, 0.001798, 0.000035, 0.000114, 0.004750, 0.001716, 0.000013}, // 1 1
{0.000000, 0.000537, 0.000008, 0.005080, 0.000046, 0.992910, 0.000012, 0.000671, 0.000390, 0.000347}, // 5 5
{0.000127, 0.000003, 0.000003, 0.000000, 0.001583, 0.000000, 0.000002, 0.000555, 0.000016, 0.997712}, // 9 9
{0.000001, 0.000012, 0.000072, 0.000020, 0.000000, 0.000000, 0.000000, 0.999868, 0.000000, 0.000026}, // 7 7
{0.000020, 0.000105, 0.001139, 0.901343, 0.002132, 0.083873, 0.000124, 0.000097, 0.010981, 0.000186}, // 3 3
{0.000002, 0.000048, 0.000019, 0.000000, 0.999347, 0.000002, 0.000040, 0.000051, 0.000000, 0.000489}, // 4 4
{0.000059, 0.001344, 0.612502, 0.002749, 0.000229, 0.000678, 0.000038, 0.001844, 0.379727, 0.000831}, // 8 2
{0.000586, 0.000740, 0.001625, 0.000007, 0.269341, 0.000076, 0.016417, 0.000199, 0.000107, 0.710902}, // 4 9
{0.009547, 0.018055, 0.283795, 0.071079, 0.426074, 0.082335, 0.036379, 0.021188, 0.003924, 0.047623}, // 2 4
{0.002506, 0.002545, 0.001148, 0.005659, 0.020416, 0.000112, 0.006092, 0.272536, 0.003148, 0.685839}, // 7 9
{0.001263, 0.001769, 0.000293, 0.000011, 0.000302, 0.881768, 0.112019, 0.000125, 0.002327, 0.000123}, // 6 5
{0.002904, 0.002909, 0.013421, 0.001461, 0.007519, 0.001251, 0.000555, 0.106219, 0.107125, 0.756637}, // 8 9
{0.000055, 0.001080, 0.893158, 0.000000, 0.104492, 0.000159, 0.001042, 0.000013, 0.000000, 0.000000}, // 4 2
{0.000344, 0.002693, 0.071184, 0.000262, 0.000001, 0.000003, 0.000032, 0.924362, 0.000714, 0.000404}, // 2 7
{0.001404, 0.009375, 0.002638, 0.229189, 0.000064, 0.000896, 0.007516, 0.743557, 0.004462, 0.000897}, // 3 7
{0.491140, 0.000125, 0.000024, 0.000302, 0.000038, 0.034947, 0.473161, 0.000170, 0.000028, 0.000066}}; // 6 0
vector top_k=y_true.ClassificationScore(y_scores,CLASSIFICATION_TOP_K_ACCURACY,1);
Print("top 1 accuracy score = ",top_k);
top_k=y_true.ClassificationScore(y_scores,CLASSIFICATION_TOP_K_ACCURACY,2);
Print("top 2 accuracy score = ",top_k);
vector y_true2={0, 1, 2, 2};
matrix y_score2={{0.5, 0.2, 0.2}, // 0 is in top 2
{0.3, 0.4, 0.2}, // 1 is in top 2
{0.2, 0.4, 0.3}, // 2 is in top 2
{0.7, 0.2, 0.1}}; // 2 isn't in top 2
top_k=y_true2.ClassificationScore(y_score2,CLASSIFICATION_TOP_K_ACCURACY,2);
Print("top k = ",top_k);
Print("");
vector ap_micro=y_true.ClassificationScore(y_scores,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_MICRO);
Print("average precision score micro = ",ap_micro);
vector ap_macro=y_true.ClassificationScore(y_scores,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_MACRO);
Print("average precision score macro = ",ap_macro);
vector ap_weighted=y_true.ClassificationScore(y_scores,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_WEIGHTED);
Print("average precision score weighted = ",ap_weighted);
vector ap_none=y_true.ClassificationScore(y_scores,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_NONE);
Print("average precision score none = ",ap_none);
Print("");
vector area_micro=y_true.ClassificationScore(y_scores,CLASSIFICATION_ROC_AUC,AVERAGE_MICRO);
Print("roc auc score micro = ",area_micro);
vector area_macro=y_true.ClassificationScore(y_scores,CLASSIFICATION_ROC_AUC,AVERAGE_MACRO);
Print("roc auc score macro = ",area_macro);
vector area_weighted=y_true.ClassificationScore(y_scores,CLASSIFICATION_ROC_AUC,AVERAGE_WEIGHTED);
Print("roc auc score weighted = ",area_weighted);
vector area_none=y_true.ClassificationScore(y_scores,CLASSIFICATION_ROC_AUC,AVERAGE_NONE);
Print("roc auc score none = ",area_none);
Print("");
//--- 二元分类
vector y_pred_bin={0,1,0,1,1,0,0,0,1};
vector y_true_bin={1,0,0,0,1,0,1,1,1};
vector y_score_true={0.3,0.7,0.1,0.6,0.9,0.0,0.4,0.2,0.8};
matrix y_score1_bin(y_score_true.Size(),1);
y_score1_bin.Col(y_score_true,0);
matrix y_scores_bin={{0.7, 0.3},
{0.3, 0.7},
{0.9, 0.1},
{0.4, 0.6},
{0.1, 0.9},
{1.0, 0.0},
{0.6, 0.4},
{0.8, 0.2},
{0.2, 0.8}};
vector ap=y_true_bin.ClassificationScore(y_scores_bin,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_BINARY);
Print("average precision score binary = ",ap);
vector ap2=y_true_bin.ClassificationScore(y_score1_bin,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_BINARY);
Print("average precision score binary = ",ap2);
vector ap3=y_true_bin.ClassificationScore(y_scores_bin,CLASSIFICATION_AVERAGE_PRECISION,AVERAGE_NONE);
Print("average precision score none = ",ap3);
Print("");
vector area=y_true_bin.ClassificationScore(y_scores_bin,CLASSIFICATION_ROC_AUC,AVERAGE_BINARY);
Print("roc auc score binary = ",area);
vector area2=y_true_bin.ClassificationScore(y_score1_bin,CLASSIFICATION_ROC_AUC,AVERAGE_BINARY);
Print("roc auc score binary = ",area2);
vector area3=y_true_bin.ClassificationScore(y_scores_bin,CLASSIFICATION_ROC_AUC,AVERAGE_NONE);
Print("roc auc score none = ",area3);
/*
top 1 accuracy score = [0.6666666666666666]
top 2 accuracy score = [1]
top k = [0.75]
average precision score micro = [0.8513333333333333]
average precision score macro = [0.9326666666666666]
average precision score weighted = [0.9333333333333333]
average precision score none = [1,1,0.7,1,0.9266666666666666,0.8333333333333333,1,0.8666666666666667,1,1]
roc auc score micro = [0.9839506172839506]
roc auc score macro = [0.9892068783068803]
roc auc score weighted = [0.9887354497354497]
roc auc score none = [1,1,0.9506172839506173,1,0.984,0.9821428571428571,1,0.9753086419753086,1,1]
average precision score binary = [0.7961904761904761]
average precision score binary = [0.7961904761904761]
average precision score none = [0.7678571428571428,0.7961904761904761]
roc auc score binary = [0.7]
roc auc score binary = [0.7]
roc auc score none = [0.7,0.7]
*/