1
$\begingroup$

I'm trying to solve a time series forecasting problem, where the main goal is to read data with various alarm logs and make a prediction about what may happen in the future.

Specifically, my data is a stream of alert data, where at each time stamp, information such as the alert monitoring system, the location of the problem etc. are stored in the alert. These fields are all categorical variables.

If someone already had an identical problem, it could help me find the best technologies and which ML models are most used in these cases and some practical examples of models.

I need to predict multiple occurrencies of future alarms. In this case the response variable is problem name.

enter image description here

**I know that in a pre-processing phase I have to encode the variables.

EDIT:

my goal is to have a multi-step output of the "problem name" label. This label is not numeric or percentual, its a string (categorical variable). i guess i will have to feed a model with data resembling something like this (after noramalization):

enter image description here

I guess that for this problem I will have to use LSTM models but my problem is that all the examples I see are trying to predicte continuous values, such as temperatures or stock values. Are any models (and special layers) targeting this kind of problems? Any good examples available online?

$\endgroup$
0
$\begingroup$

The first step is to pick a target variable. What are you trying to predict "problem name", "severity", "sub category", …?

Based on the target, it becomes either a conventional regression, ordinal regression, binary classification, multi-class classification, or multi-label classification problem.

Another issue might be how the date is encoded. "t1", "t2", … is not a conventional datestamp format.

$\endgroup$
1
  • $\begingroup$ i update my question, check pls $\endgroup$ – Valdemar sousa Oct 21 '20 at 9:15
0
$\begingroup$

There are two ways to setup prediction depending on your dataset:

  • If you have a row every minute/hour/day just find the rows with the right alarm and put 1 the observation 1/7/10 days ahead depending how much ahead of time you want to predict the alarm. Of course the less recent it is the more difficult to predict.
  • If you have a row every event it is harder because you don't have the right date of the next alarm but if in the next k events there is an alarm. The technique is pretty similar to the latter (find right alarm) but this time you just scale up by k rows and use a classifier to predict if the next kth event will be an alarm

If you want to predict multiple alarms just use multiple classification instead of 0/1

$\endgroup$
2
  • $\begingroup$ i update my question, check pls $\endgroup$ – Valdemar sousa Oct 21 '20 at 9:15
  • $\begingroup$ LSTM is just an hidden layer, if you want to predict multiple labels just apply after the last layer of your NN a Softmax activation function and use categorical crossentropy as loss function $\endgroup$ – Mikedev Oct 21 '20 at 10:48

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