Share
Embed
Area under the receiver operating curve
A Jupyter notebook exploring the relationship between AUC and Cohen's effect size in multiple dimensions.
Allen Downey
April 17, 2019
Transcript
36 slides
-
Area under the receiver operating curve Area under the receiver
operating curve Copyright 2019 Allen Downey License: http://creativecommons.org/licenses/by/4.0/ (http://creativecommons.org/licenses/by/4.0/) -
Area under ROC Area under ROC As a way of
understanding AUC ROC, let's look at the relationship between AUC and Cohen's effect size. Cohen's effect size, d , expresses the difference between two groups as the number of standard deviations between the means. As d increases, we expect it to be easier to distinguish between groups, so we expect AUC to increase. I'll start in one dimension and then generalize to multiple dimensions.a -
Here are the means and standard deviations for two hypothetical
groups. In [2]: mu1 = 0 sigma = 1 d = 1 mu2 = mu1 + d; -
I'll generate two random samples with these parameters. In [3]:
n = 1000 sample1 = np.random.normal(mu1, sigma, n) sample2 = np.random.normal(mu2, sigma, n); -
If we put a threshold at the midpoint between the
means, we can compute the fraction of Group 0 that would be above the threshold. I'll call that the false positive rate. In [4]: thresh = (mu1 + mu2) / 2 np.mean(sample1 > thresh) Out[4]: 0.301 -
And here's the fraction of Group 1 that would be
below the threshold, which I'll call the false negative rate. In [5]: np.mean(sample2 < thresh) Out[5]: 0.325 -
Plotting misclassi cation Plotting misclassi cation To see what these
overlapping distributions look like, I'll plot a kernel density estimate (KDE). In [8]: def plot_misclassification(sample1, sample2, thresh): """Plot KDEs and shade the areas of misclassification. sample1: sequence sample2: sequence thresh: number """ kde1 = make_kde(sample1) clipped = kde1[kde1.index>=thresh] plot_kde(kde1, clipped, 'C0') kde2 = make_kde(sample2) clipped = kde2[kde2.index<=thresh] plot_kde(kde2, clipped, 'C1') -
Here's what it looks like with the threshold at 0.
There are many false positives, shown in blue, and few false negatives, in orange. In [9]: plot_misclassification(sample1, sample2, 0) -
With a higher threshold, we get fewer false positives, at
the cost of more false negatives. In [10]: plot_misclassification(sample1, sample2, 1) -
The receiver operating curve The receiver operating curve The receiver
operating curve (ROC) represents this tradeoff. To plot the ROC, we have to compute the false positive rate (which we saw in the gure above), and the true positive rate (not shown in the gure). The following function computes these metrics. In [11]: def fpr_tpr(sample1, sample2, thresh): """Compute false positive and true positive rates. sample1: sequence sample2: sequence thresh: number returns: tuple of (fpr, tpf) """ fpr = np.mean(sample1>thresh) tpr = np.mean(sample2>thresh) return fpr, tpr -
When the threshold is high, the false positive rate is
low, but so is the true positive rate. In [12]: fpr_tpr(sample1, sample2, 1) As we decrease the threshold, the true positive rate increases, but so does the false positive rate. In [13]: fpr_tpr(sample1, sample2, 0) Out[12]: (0.15, 0.501) Out[13]: (0.483, 0.846) -
The ROC shows this tradeoff over a range of thresholds.
I sweep thresholds from high to low so the ROC goes from left to right. In [14]: from scipy.integrate import trapz def plot_roc(sample1, sample2, label): """Plot the ROC curve and return the AUC. sample1: sequence sample2: sequence label: string returns: AUC """ threshes = np.linspace(5, -3) roc = [fpr_tpr(sample1, sample2, thresh) for thresh in threshes] fpr, tpr = np.transpose(roc) plt.plot(fpr, tpr, label=label) plt.xlabel('False positive rate') plt.ylabel('True positive rate') auc = trapz(tpr, fpr) return auc -
Here's the ROC for the samples. With d=1 , the
area under the curve is about 0.75. That might be a good number to remember. In [15]: auc = plot_roc(sample1, sample2, '') Out[15]: 0.761592 -
In [16]: mu1 = 0 sigma = 1 n =
1000 res = [] for mu2 in [3, 2, 1.5, 0.75, 0.25]: sample1 = np.random.normal(mu1, sigma, n) sample2 = np.random.normal(mu2, sigma, n) d = (mu2-mu1) / sigma label = 'd = %0.2g' % d auc = plot_roc(sample1, sample2, label) res.append((d, auc)) plt.legend(); -
This function computes AUC as a function of d .
In [17]: def plot_auc_vs_d(res, label): d, auc = np.transpose(res) plt.plot(d, auc, label=label, alpha=0.8) plt.xlabel('Cohen effect size') plt.ylabel('Area under ROC') -
The following gure shows AUC as a function of d
. In [18]: plot_auc_vs_d(res, '') Not suprisingly, AUC increases as d increases. -
Multivariate distributions Multivariate distributions Now let's see what happens if
we have more than one variable, with a difference in means along more than one dimension. First, I'll generate a 2-D sample with d=1 along both dimensions. In [19]: from scipy.stats import multivariate_normal d = 1 mu1 = [0, 0] mu2 = [d, d] rho = 0 sigma = [[1, rho], [rho, 1]] In [20]: sample1 = multivariate_normal(mu1, sigma).rvs(n) sample2 = multivariate_normal(mu2, sigma).rvs(n); Out[19]: [[1, 0], [0, 1]] -
The mean of sample1 should be near 0 for both
features. In [21]: np.mean(sample1, axis=0) And the mean of sample2 should be near 1. In [22]: np.mean(sample2, axis=0) Out[21]: array([ 0.01204411, -0.05193738]) Out[22]: array([0.97947675, 1.02358947]) -
The following scatterplot shows what this looks like in 2-D.
In [23]: x, y = sample1.transpose() plt.plot(x, y, '.', alpha=0.3) x, y = sample2.transpose() plt.plot(x, y, '.', alpha=0.3) plt.xlabel('X') plt.ylabel('Y') plt.title('Scatter plot for samples with d=1 in both dimensions'); -
Some points are clearly classi able, but there is substantial
overlap in the distributions. -
We can see the same thing if we estimate a
2-D density function and make a contour plot. In [25]: X, Y, Z = kde_scipy(sample1) plt.contour(X, Y, Z, cmap=plt.cm.Blues, alpha=0.7) X, Y, Z = kde_scipy(sample2) plt.contour(X, Y, Z, cmap=plt.cm.Oranges, alpha=0.7) plt.xlabel('X') plt.ylabel('Y') plt.title('KDE for samples with d=1 in both dimensions'); -
Classi cation with logistic regression Classi cation with logistic regression
To see how distinguishable the samples are, I'll use logistic regression. To get the data into the right shape, I'll make two DataFrames, label them, concatenate them, and then extract the labels and the features. -
In [26]: df1 = pd.DataFrame(sample1) df1['label'] = 1 df1.describe() In
[27]: df1[[0,1]].corr() Out[26]: 0 1 label count 1000.000000 1000.000000 1000.0 mean 0.012044 -0.051937 1.0 std 0.971861 0.976814 0.0 min -3.580857 -3.061129 1.0 25% -0.596927 -0.696824 1.0 50% 0.071937 -0.044057 1.0 75% 0.655457 0.615113 1.0 max 3.053507 3.292066 1.0 Out[27]: 0 1 0 1.000000 0.021376 1 0.021376 1.000000 -
In [28]: df2 = pd.DataFrame(sample2) df2['label'] = 2 df2.describe() In
[29]: df2[[0,1]].corr() Out[28]: 0 1 label count 1000.000000 1000.000000 1000.0 mean 0.979477 1.023589 2.0 std 0.983136 0.967058 0.0 min -2.231272 -2.027548 2.0 25% 0.291482 0.417082 2.0 50% 1.008545 1.008277 2.0 75% 1.670930 1.647037 2.0 max 3.869119 4.138071 2.0 Out[29]: 0 1 0 1.00000 -0.04433 1 -0.04433 1.00000 -
In [30]: df = pd.concat([df1, df2], ignore_index=True) df.label.value_counts() Out[30]: 2
1000 1 1000 Name: label, dtype: int64 -
X is the array of features; y is the vector
of labels. In [31]: X = df[[0, 1]] y = df.label; Now we can t the model. In [32]: from sklearn.linear_model import LogisticRegression model = LogisticRegression(solver='lbfgs').fit(X, y); And compute the AUC. In [33]: from sklearn.metrics import roc_auc_score y_pred_prob = model.predict_proba(X)[:,1] auc = roc_auc_score(y, y_pred_prob) With two features, we can do better than with just one. Out[33]: 0.853391 -
AUC as a function of rho AUC as a function
of rho The following function contains the code from the previous section, with rho as a parameter. -
In [34]: def multivariate_normal_auc(d, rho=0): """Generate multivariate normal samples and
classify them. d: Cohen's effect size along each dimension num_dims: number of dimensions returns: AUC """ mu1 = [0, 0] mu2 = [d, d] sigma = [[1, rho], [rho, 1]] # generate the samples sample1 = multivariate_normal(mu1, sigma).rvs(n) sample2 = multivariate_normal(mu2, sigma).rvs(n) # label the samples and extract the features and labels df1 = pd.DataFrame(sample1) df1['label'] = 1 df2 = pd.DataFrame(sample2) df2['label'] = 2 df = pd.concat([df1, df2], ignore_index=True) X = df.drop(columns='label') y = df.label # run the model model = LogisticRegression(solver='lbfgs').fit(X, y) y_pred_prob = model.predict_proba(X)[:,1] # compute AUC auc = roc_auc_score(y, y_pred_prob) return auc -
In [35]: res = [(rho, multivariate_normal_auc(d=1, rho=rho)) for rho in
np.linspace(-0.9, 0.9)] rhos, aucs = np.transpose(res) plt.plot(rhos, aucs) plt.xlabel('Correlation (rho)') plt.ylabel('Area under ROC') plt.title('AUC as a function of correlation'); -
AUC as a function of d AUC as a function
of d The following function contains the code from the previous section, generalized to handle more than 2 dimensions. -
In [36]: def multivariate_normal_auc(d, num_dims=2): """Generate multivariate normal samples and
classify them. d: Cohen's effect size along each dimension num_dims: number of dimensions returns: AUC """ # compute the mus mu1 = np.zeros(num_dims) mu2 = np.full(num_dims, d) # and sigma sigma = np.identity(num_dims) # generate the samples sample1 = multivariate_normal(mu1, sigma).rvs(n) sample2 = multivariate_normal(mu2, sigma).rvs(n) # label the samples and extract the features and labels df1 = pd.DataFrame(sample1) df1['label'] = 1 df2 = pd.DataFrame(sample2) df2['label'] = 2 df = pd.concat([df1, df2], ignore_index=True) X = df.drop(columns='label') y = df.label # run the model model = LogisticRegression(solver='lbfgs').fit(X, y) y_pred_prob = model.predict_proba(X)[:,1] # compute AUC -
Con rming what we have seen before: In [37]: multivariate_normal_auc(d=1,
num_dims=1) In [38]: multivariate_normal_auc(d=1, num_dims=2) Out[37]: 0.7680769999999999 Out[38]: 0.8322470000000001 -
Now we can sweep a range of effect sizes. In
[39]: def compute_auc_vs_d(num_dims): """Sweep a range of effect sizes and compute AUC. num_dims: number of dimensions returns: list of """ effect_sizes = np.linspace(0, 4) return [(d, multivariate_normal_auc(d, num_dims)) for d in effect_sizes] In [40]: res1 = compute_auc_vs_d(1) res2 = compute_auc_vs_d(2) res3 = compute_auc_vs_d(3) res4 = compute_auc_vs_d(4); -
And plot the results. In [41]: plot_auc_vs_d(res4, 'num_dim=4') plot_auc_vs_d(res3, 'num_dim=3')
plot_auc_vs_d(res2, 'num_dim=2') plot_auc_vs_d(res1, 'num_dim=1') plt.title('AUC vs d for different numbers of features') plt.legend(); With more features, the AUC gets better, assuming the features are independent.