First of all, I don’t know much about ML for sequential data. This is just an idea, I want you to check if it is valid to dig.
I’ve considered tabular sequential data. In this case, it is not well suited for DL because it’s tabular, but it is difficult to do recursion with a tree-based model. Couldn’t it be simpler when the data is categorical and not related to time? The idea is this:
When there is a row for each case, assign some feature with a value between 0 and 1, with each digit being 0 or 1. 1 if something happened, 0 otherwise. And more recent events come to the forefront.
For example, a value of 0.010001 is the value for the case that some event happened the first trial, then not happened three times, and happened and most recent trial.
If you assign a value like this
- All cases for each case can be separated.
- The biggest problem with sequential data, the form must be fixed, can be solved. For example, 0.101 is equivalent to 0.101000000…
- If a value with a sequence longer than the longest value in the training data comes into the input, it can be processed properly. Naturally more weighted on the most recent event.
Please check if this idea is already obsolete, typical or absurd! If it looks ok, I’ll dig it.