1
$\begingroup$

I need to answer the following question:

What is the probability that the event 1 will occur at some point in the time for a new sample? At which time point is this more likely to occur?

1: event occurred at a specific time point, 0: event did not occur, -1: the opposite of the requested event occurred (also not interesting).

For each sample, 6 time points are measured over time. Let us assume this is for example every week (1w, 2w, 3w, 4w, 5w, 6w).

sample 1 = (0, 0, 0, 0, 0, 1)
sample 2 = (0, 0, 1, 1, 1, 1)
sample 3 = (0, 0, 0, 0, 1, 1)
sample 4 = (0, 0, 0, 0, 0, 1)
sample 5 = (0, 0, 0, 1, 1, 1)
sample 6 = (0, 0, 0, 0, 0, 1)
sample 7 = (0, -1, -1, -1, -1, 0)
sample 8 = (0, 0, 0, 0, 0, 0)

What model would be appropriate for this?

Apparently, there are also "noisy" samples where we do not observe the event (samples 7, 8).

$\endgroup$

1 Answer 1

0
$\begingroup$

This can be as a typical time-series prediction and/or sequence modeling problem. The bad news is that 6 time-stamps is pretty short and you may find it more difficult to model compared to long sequences.

A simple statistical analysis (and not necessarily a ML model) is actually to calculate the probabilities. You have 6 time-stamps ($T$), 3 values ($X$) and $N$ samples. You can have an overall idea by calculating all $P(X_{i}=x|T=t, X_{i-1}=y)$ where $x,y\in X$ (i.e. $x$ and $y$ get values from -1, 0 and 1).

This is pretty Markovian as I took only the previous event responsible for predicting the current event. Sequence models can learn the history thus giving more holistically reliable results.

PS: Wait for better solutions. It is just a starter idea.

$\endgroup$
6
  • $\begingroup$ Thanks Kasra, you are right that samples are too few. But I could collect more data if there is a method for approaching the problem. Also this data is more like an example to explain what type of data I have. $\endgroup$
    – J. Doe
    Mar 16, 2022 at 10:04
  • $\begingroup$ Of course if you can collect "longer sequences" you can simply feed it to any sequence model, ranging from classic HMM to RNNs like LSTM and ESN, and get a pretty descent result. But still the idea above is not that stupid. I would suggest to at least try it in case you have a blocking point. This is very close to how all Bayesian Models work. Good luck and let me know if I you need more info and upvote if you got a little help from answer :))) $\endgroup$ Mar 16, 2022 at 10:10
  • 1
    $\begingroup$ So it seems that I need longer sequences to be able to perform HMM and probably more samples (sequences) to be able to compute the Markovian probabilities. But given the data I have I could still compute the probabilities based on Markovian but I should have in mind that the sequences are very few. Do I get it right? (seems I need a bit more reputation for the upvote :) ) $\endgroup$
    – J. Doe
    Mar 16, 2022 at 10:27
  • $\begingroup$ That is true! Sequence models "work" on even short sequences but it is about "how well they work". Longer sequences, better performance $\endgroup$ Mar 16, 2022 at 10:37
  • $\begingroup$ Thanks Kasra. Would a simple correlation metric also work here? Like if we would observe a pattern that the 0s increase from 0 to 1 across samples (independently of the time point that the transition from 0 to 1 occurs). $\endgroup$
    – J. Doe
    Mar 16, 2022 at 14:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.