Article Understanding and using sensitivity, specificity and predictive values

Hi Qi Teng , nice question. The problem you refering to is known as the one-class-classification. There is a nice entry: see http://rvlasveld.github.io/blog/2013/07/12/introduction-to-one-class-support-vector-machines/. Kind regards Rob

You can take a look-

https://machinelearningmastery.com/one-class-classification-algorithms/

It depends on the particular criteria. In my experience, the most stable and balanced results are obtained when using the inter-vector distance as a non-metric integral criterion (correct for non-normalized spaces) on the chi-square model or the amount of information according to Harkiewicz.

Preprint The structure of the description of scientific research usin...

Dear Qi Teng

I think this paper can help you

Article Advanced Meta-Heuristics, Convolutional Neural Networks, and...

POSITIVE PSEUDOPROGNOSIS.

Suppose the model gives such a forecast that everything will fall out: 1, 2, 3, 4, 5, and 6. It is clear that only one thing will fall out of all this. In this case, the model will not

predict what will not fall out, but it will certainly predict what will fall out. However, a lot of objects will be assigned to classes to which they do not belong.

Then the probability of true-positive solutions for the model will be 1/6, and the probability of false-positive solutions will be 5/6. It is clear that such a forecast is useless, which is why

I called it a pseudo-forecast.

NEGATIVE PSEUDOPROGNOSIS.

Imagine that we throw a cube with 6 faces, and the model predicts that nothing will fall out, i.e. neither 1, nor 2, nor 3, nor 4, nor 5, nor 6 will fall out, but something of

this, of course, will definitely fall out. Of course, the model did not predict what would fall out, but it predicted very well that it would not fall out. The probability of true-

negative solutions for the model will be 5/6, and the probability of false-negative solutions will be 1/6. Such a forecast is much more reliable than a positive pseudo-forecast, but it is also useless.

THE PERFECT FORECAST.

If, in the case of a cube, we predict that 1 will fall out, for example, and, accordingly, we predict that 2, 3, 4, 5, and 6 will not fall out, then this is an ideal forecast, having,

if it is carried out, 100% certainty of identification and non-identification. An ideal forecast that completely removes uncertainty about the future state of the object

forecasting, in practice it is extremely rare to get it, and usually we are dealing with a real forecast.

REAL FORECAST.

In practice, we most often encounter this type of forecast. A real forecast reduces uncertainty about the future state of the forecast object, but not

completely, as an ideal forecast, but leaves some uncertainty not removed. For example, for a dice, the following forecast is made: 1 or 2 will fall out, and,

accordingly, 3, 4, 5 or 6 will not fall out. It is clear that such a forecast cannot be fully realized in practice, because the variants of the dice are alternative, i.e.

both 1 and 2 cannot fall out at the same time. Therefore a real forecast will always have a certain identification error. Accordingly, if one or more fails

if there are several of the predicted options, then there will be a non-identification error, because this was not predicted by the model. Now imagine that you do not have 1 cube and

a forecast of its behavior, but thousands. Then you can calculate the weighted average characteristics of all these types of forecasts.