In a machine learning application, the task is to estimate the market value of a property based on its features, e.g. property type, location, size, etc.

Let the model output for a property be and the trade price be , in the learning process, a criterion is employed to compare and y to assess the model:

Which of the following statements is appriopriate to describe the criterion?

Group of answer choices

is determined by the difference between the predicted price and the actual trading price , i.e. .

That is, the comparison must be depending on the absolute value of the difference between and and nothing else.

This means that in the following two cases: Case 1: , for a house. Case 2: for a car parking space. the criterion must yield the same value and measure the same "loss" for the model. Also, this means that over/under-estimation of the market value cannot be distinguished by the criterion.

is determined by the predicted price and the actual trading price .

If the trade price of the property is known, the loss reflects the realistic performance of the model on the task of market value estimation. This is regardless of all other factors, as the trade price is the golden standard for the market value of the property.

During the learning process, with each iteration to update the hypothesis, the value must reduce.

By "updating", it means to adjust the model, such as changing the parameters of the model, or changing the model structure, to better fit the data. Since in practical improvement of the algorithm, the step is designed to minimise the loss function evaluated on the training data.

It is believed that, no matter how the model update details are designed, the training loss is expected to reduce in each iteration. For example, if the model update is to change the parameters and the gradient descent is employed, the loss will reduce along the gradient direction regardless the other choices, such as the learning rate, the normalisation, etc.