I have a time series classification problem that uses a series of if-else statements to arrive at a particular label. I am attempting to use ML/DL to make the system simpler.

So far, I have tried using a tabular data approach where I take a snapshot of information up to a particular point. For example, this will be the rolling sum of certain columns and so on. I have also tried LSTM and CNN. All these approaches have failed to give me F1 scores significantly above 50 %.

Are there other ML/DL approaches that I should try before giving up? The models were built using AutoKeras and PyCaret.


1 Answer 1


Are you sure, that there is a need for ML? If there is a set of rules that allows to solve this problem without ML, it gives already 100% accuracy, whereas ML/DL will do it up to a certain accuracy, and this task may be even tough for supervised algorithm. Sorting problem is very difficult for neural networks.

Anyway, check whether your algorithm can express these if-else statements. For example such branch, given that the output is of the same size for both conditional branches, can be approximated by the following:

sigmoid(beta * condition) * output_1 + sigmoid(-beta * condition) * output_2

Here beta controls the slope of the sigmoid, condition is the conditional statement of the form $f(x) >(\geqslant) 0$, and output_1 and output_2 are the outcomes of both options.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .