How to build my own dataset and model for an LSTM neural network

I have a sort of mathematical problem and I'm not sure which model I should choose to make an LSTM neural network.

Currently in my country, there is a system in which certain groups of researchers upload information on products of scientific interest, such as research articles, books, patents, software, among others. Depending on the number of products, the system assigns a classification to each group, which can be A1, A, B and C, where A1 is the highest classification and C is the minimum.

The classification is done through a mathematical model whose entries are, the total number of each product, the total sum of all products, number of authors, among other indices that are calculated with the previous values.

Once the entries are obtained, these values ​​are processed by a set of formulas and the final result is a single number.

This number is located in a range provided by the mathematical model and this is how the group is classified.

What I want to do is given the current classification of a group, give suggestions of different values ​​to improve their classification.

For example, if there is a group with classification C, suggest how many products it should have, how many authors, what value should its indexes have, so that its category would be finally B.

I think the structure of my network should be: -1 input, which would be the classification you want to get. -Multiple output, one for each product and indexes.

But I do not understand how to make the network take into account the current classification of the group, in addition to the number of products and the value of the current indexes.

If you have further questions about the problem, please feel free to ask.

I appreciate your suggestions.

2 Answers

why do you want to use LSTM network? lstm is a variant of a recurrent neural network , recurrent neural networks are used for "sequential" tasks , i.e the dataset should have some sequential structure , like poems , songs etc.

Your model should be a simple classifier , once you fit a simple classifier , like a decision forest on the categories dataset products ,authors etc , you will have a model that will predict the class based on these attributes then you can say from the decision boundaries of the model what values you must have . if the relationship between the attributes is even more simpler you could try plotting distribution plots.

• I thought I could use it because I felt that my NN should have memory to make accurate recommendations. For example, if there is a group of lawyers the system should not recommend them to make medicine articles, but it could be a valid recommendation if they have products related to medicine in the group´s profile. Is it still a simple classifier if I have a model with one entry, which would be the classification and 100 outputs each representing a product and its respective quantity to achieve this classification?
– LP0
Jul 5, 2018 at 17:51

If you want to vary definitions of outcome, the problem is more about optimal segmentation/clustering and not a classification.

For clustering you could try latent class approaches.