What we are trying to do in Machine Learning is to have an smart approach, a model that can be reproduced in different contexts (in our case, on different tunnels).įortunately for us, there is a performance analysis that allows us to give more importance to smart models than to naive models. That’s why a naive approach is not efficient.
Indeed, since our data is biased (there are many more cars than bikes), our performance is biased too. In fact, this type of prediction is not reliable. The problem is that this approach works when there are 80% of cars on the road, but if the context changes, this approach will not work anymore. This is what we call a naive prediction because there is no real reflection: we don’t predict, we arbitrarily decide that everything that comes out of the tunnel will be a car. If on this road 80% of the users are in cars and we predict each time that it will be a car that will come out of the tunnel… We will have at the end a success rate of 80%. We imagine that there are two possibilities here, either it is a car either a motorcycle. Let’s take the example of a choice with two possibilities : we place ourselves at the exit of a tunnel to predict what will come out. On the contrary, this approach can give us a distorted view of our model. To measure the performance of a Machine Learning model, we cannot simply look at the number of well-made predictions. To Know Why use True Positives / False Negatives ?
0 kommentar(er)
