what type of machine learning technique use in this case?

Question

I have found the following question:

I have made the following pre-assumptions and I would like if somebody could help me to see if they are right or wrong:

• I believe that there is the need to preprocess the data because the prices and the quantities can be in different scales, e.g. prices sold of each candy in thousand of dollars and quantities in tons or other metric unit.

• At first sight I can see that it is possible to use a neural network technique, a multilayer nn maybe with a backpropagation algorithm. The inputs could be the data obtained from the ten years regarding the price sold by us, the quantities we sold of each candy and the price sold by the competitors; three neurons for the output each one representing the quantities to be sell per each type of candy.

• I am not sure if I could use a linear regression technique, because maybe the data will not have linear characteristics and some features will invalid my prediction model.

• Logistic regression could be used, I suggest that the model of one vs all can be used for this case because we will have three outputs, each one per each type of candy. The inputs could be the prices sold per each type of candy by the competiros, the quantities we sold and the price we sold those (all over the amount of ten years).

Aside note: I can also use as an input the quantities of candies that were in the market for the last 2 years, but here I will have to rescale it because I am using a 10 years time period.

• Clustering? I am not sure how to use it here.

Any help?

Answer 1

Seems like a multiple choice question ("Which of the following...algorithms...?"), but you don't tell us explicitly what these alternatives are.

Really can't tell without more information, but here are some thoughts.

They are asking for a numerical outcome => this is a regression problem.

They are not asking for probabilities, or relative strengths: => not a logistical regression problem (which is used for classification/ categorical response variables anyway). Although with some assumptions + data processing you can treat it as a regression problem, I think.

Linear regression seems fine here, although some attributes (caketype) are nominal/categorical. The linear-model functions of most statistical software packages can deal with this.

fit = lm(q ~ price + price_of_competitors + year | caketype)

No clustering. no neural network.

Maybe a NN can be used, but it is more like a black box (applying logit function) to me. It does not seem to be the right tool for the job, as this is a problem with a simple linear relationship between price p and quantity.