COCA - Anomaly Detection

[marciahon29]

Hello,

Please could you let me know how the Anomaly threshold is calculated? In your paper, you mention it is predefined. But how is it predefined?

Thanks

[ruiking04]

You need to find this within a range yourself. For example, to find the percentile of (1, 100), the code is as follows:

scores = np.random.rand(1000)
labels = np.random.randint(0, 1, 1000)

def find_best_threshold(scores, labels):
    best_f1 = 0
    best_threshold = 0
    for detect_nu in np.arange(1, 100) / 1e2:
        threshold = np.percentile(scores, 100-detect_nu)
        predicted_labels = (scores > threshold).astype(int)
        f1 = f1_score(labels, predicted_labels)

        if f1 > best_f1:
            best_f1 = f1
            best_threshold = threshold

    return best_threshold, best_f1

best_threshold, best_f1 = find_best_threshold(scores, labels)

[marciahon29]

I didn't do anything to find a threshold, I just used your code as is. Is this okay?

Also , which file has this code?

On Wed, Feb 14, 2024 at 7:37 PM RuiKing @.***> wrote:

You need to find this within a range yourself. For example, to find the percentile of (0, 100), the code is as follows:

scores = np.random.rand(1000) labels = np.random.randint(0, 1, 1000)

def find_best_threshold(scores, labels): best_f1 = 0 best_threshold = 0 for detect_nu in np.arange(1, 100) / 1e2: threshold = np.percentile(scores, 100-detect_nu) predicted_labels = (scores > threshold).astype(int) f1 = f1_score(labels, predicted_labels)

    if f1 > best_f1:
        best_f1 = f1
        best_threshold = threshold

return best_threshold, best_f1

best_threshold, best_f1 = find_best_threshold(scores, labels)

— Reply to this email directly, view it on GitHub https://github.com/ruiking04/COCA/issues/29#issuecomment-1945175797, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFOIJPTPTADQ3NJ23GSCZ7TYTVKENAVCNFSM6AAAAABDI7GATKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNBVGE3TKNZZG4 . You are receiving this because you authored the thread.Message ID: @.***>

[marciahon29]

Thank you. I have found the file that specifies this. It is in anomaly_predict.py with the function ad_predict.