Last week I met Joe Barr of Acronis who gave a talk at NAU about an application of machine learning to predicting security vulnerabilities based on source code analysis, and a database of exploits. The data they are using for training/testing their models is highly imbalanced so I suggested trying my new Area Under the Min(FP,FN) loss function (AUM).

After talking with them about the easiest way to try this new loss function on their data, I decided that it would be good to code it up in torch (for automatic differentiation). To do that we need the torch.argsort operation, which is also used for Area Under ROC Curve computation in _binary_clf_curve in roc in auroc in forward method of AUROC class, see pytorch_lightning.metrics.classification and metrics.function.classification source code.

Below I wrote an implementation of our AUM loss function,

import torch
def AUM(pred_tensor, label_tensor):
    """Area Under Min(FP,FN)

    Loss function for imbalanced binary classification
    problems. Minimizing AUM empirically results in maximizing Area
    Under the ROC Curve (AUC). Arguments: pred_tensor and label_tensor
    should both be 1d tensors (vectors of real-valued predictions and
    labels for each observation in the set/batch).

    """
    fn_diff = torch.where(label_tensor == 1, -1, 0)
    fp_diff = torch.where(label_tensor == 1, 0, 1)
    thresh_tensor = -pred_tensor.flatten()
    sorted_indices = torch.argsort(thresh_tensor)
    sorted_fp_cum = fp_diff[sorted_indices].cumsum(axis=0)
    sorted_fn_cum = -fn_diff[sorted_indices].flip(0).cumsum(axis=0).flip(0)
    sorted_thresh = thresh_tensor[sorted_indices]
    sorted_is_diff = sorted_thresh.diff() != 0
    sorted_fp_end = torch.cat([sorted_is_diff, torch.tensor([True])])
    sorted_fn_end = torch.cat([torch.tensor([True]), sorted_is_diff])
    uniq_thresh = sorted_thresh[sorted_fp_end]
    uniq_fp_after = sorted_fp_cum[sorted_fp_end]
    uniq_fn_before = sorted_fn_cum[sorted_fn_end]
    uniq_min = torch.minimum(uniq_fn_before[1:], uniq_fp_after[:-1])
    return torch.sum(uniq_min * uniq_thresh.diff())

The implementation above consists of vectorized torch operations, all of which should be differentiable almost everywhere. To check the automatically computed gradient from torch with the directional derivatives described in our paper, we can use the backward method,

def loss_grad(pred_vec, label_vec):
    pred_tensor = torch.tensor(pred_vec)
    pred_tensor.requires_grad = True
    label_tensor = torch.tensor(label_vec)
    loss = AUM(pred_tensor, label_tensor)
    loss.backward()
    return loss, pred_tensor.grad

Simple differentiable points

Let us consider an example from the aum R package,

bin.diffs <- aum::aum_diffs_binary(c(0,1))
aum::aum(bin.diffs, c(-10,10))

## $aum
## [1] 0
## 
## $derivative_mat
##      [,1] [,2]
## [1,]    0    0
## [2,]    0    0

The R code and output above shows that for one negative label with predicted value -10, and one positive label with predicted value 10, we have AUM=0 and directional derivatives are also zero. That is also seen in the python code below:

y_list = [0, 1]
loss_grad([-10.0, 10.0], y_list)

## (tensor(0., grad_fn=<SumBackward0>), tensor([-0., -0.]))

Another example is for the same labels but the opposite predicted values, for which the R code is below,

aum::aum(bin.diffs, c(10,-10))

## $aum
## [1] 20
## 
## $derivative_mat
##      [,1] [,2]
## [1,]    1    1
## [2,]   -1   -1

The R code and output above shows that we have AUM=20 and derivative 1 for the first/negative example, and derivative -1 for the second/positive example. Those derivatives indicate that the AUM can be decreased by decreasing the predicted value for the first/negative example and/or increasing the predicted vale for the second/positive example. Consistent results can be observed from the python code below,

loss_grad([10.0, -10.0], y_list)

## (tensor(20., grad_fn=<SumBackward0>), tensor([ 1., -1.]))

Non-differentiable points

As discussed in our paper, the AUM is not differentiable everywhere. So what happens when you use auto-grad on a non-differentiable loss? Apparently the backward method returns a subgradient, as explained in an issue discussing autograd of L1 loss function.

One example, again taken from the aum R package, is when both predicted values are zero,

aum::aum(bin.diffs, c(0,0))

## $aum
## [1] 0
## 
## $derivative_mat
##      [,1] [,2]
## [1,]    0    1
## [2,]   -1    0

The output above indicates AUM=0 with directional derivatives which are not equal on the left and right. The first row of the directional derivative matrix says that decreasing the first/negative predicted value will result in no change to AUM, whereas increasing will result in increasing AUM. The second row says that decreasing the second/positive predicted value results in increasing AUM, whereas increasing the predicted value results in no change to AUM. Is the python torch auto-grad code consistent?

loss_grad([0.0, 0.0], y_list)

## (tensor(0., grad_fn=<SumBackward0>), tensor([-0., -0.]))

The output above says that the gradient is zero, which is OK (these predicted values are a minimum) but it is missing some information about what happens nearby.

Non-convex points

The implementation above of AUM is only for binary classification, but in our paper we also discuss an application to changepoint detection problems. In the case of changepoint detection problems with non-monotonic error functions, the AUM may be non-convex, and there may be points at which there are no subgradients. Some are used as test cases in our R package. What does auto-grad return in that case? (exercise for the reader)