设计神经网络进行销售预测

Question

I'm developing a distributedsystem in which a plan to use a neural network for sales prediction.

A small description of the system: the system combine the workflow of modules such as Pharmacy, Patient (electric health record) and a Doctor module.

the problem is: I have many information related to the sales of medicine and i'm a little bit confused in designing the neural network.

My current design : Inputs (according to the data that can be taken from the database):

1. Date: the date of selling the medicine to a customer (YYYY/MM/DD format).
2. Patient Age Category: a value that correspond to the age category of the patient ex: (1 -> 12: child, 13 -> 30: Young …..)
3. Patient Gender.
4. Medicine ID: a value that correspond to the medicine.
5. Disease ID: a value that correspond to the disease which was discovered by the doctor who generate the prescription.
6. Medicine Unit Cost: a value that correspond to the cost of purchasing a medicine.
7. Medicine Unit Sold: a value that corresponds to the cost of selling a medicine.
8. Pharmacy Address ID : a value that corresponds to the address of the pharmacy.
9. Season ID : a value that corresponds to the season of selling the medicine ex(summer, winter , … etc).

Outputs :

1. Quantity : A Value that correspond to the quantity of the medicine that will be sold.
2. Profit : A Value that correspond to the profit amount of selling the previous quantity.

the question is : I'm not sure if this design is meaningful, any better suggestion?

and what type of networks should I use to implement this design .. I'm planning to use multi-layer recurrent network ... It's a good choice or there is a better model?

note: I'm planning to implement the network using c# with "AForge.NET Framework".

hope this description is clear and simple, and sorry for my poor language.

Answer 1

IDs and categories make a poor choice of actual inputs, because they are not continuous and there is little meaning in their relative magnitudes. What IDs may be good for is to create separate networks for different categories, but, given the large number of IDs and categories you've defined here, this means you have a very large number of separate networks, requiring a large pool of training data, since it will be heavily diluted. Binary categories (like gender) might work, but anything with more than one category will probably not yield good results.

Be very careful with neural networks, as, given a big enough network, you can create something that appears to give predictions, but that is really meaningless outside the training set. Be sure you have a large validation set that does not take part in the training.

Answer 2

I question whether the date (in absolute terms) adds any value to the model. If you had many years of data, it might yield some meaning, but a measure like what quarter is the date in, is it a weekend or a weekday, or is it within n number of days from a major holiday (etc.) might be much more powerful.

I don't know your immediate field, so I don't have any opinions on the network topology. If you look up some papers, there may be review articles geared towards your area of study discussing what the trends are.

Answer 3

This appears to be missing some bits of info.

In order to predict the future you need a baseline to go from. For example, on a given prescription how likely is it that someone is going to refill it? Is that likelihood tied to any of your other metrics (age group, gender, date sold)?

Also season only has bearing given the location of the patient(s), and probably only for certain types of drugs (I'm thinking cold/allergy/flu). Further in order to provide any real seasonal accuracy you'd have to bring in data for that season such as allergen levels, etc.

Next, the cost of the medicine and profit expectations are really related only to the actual medicine and aren't necessary to predict sales likelihood.