I need to do a classification of binary labels. The dataset represents a loan procedure. So one row is a certain moment of a process for a certain case.

The dataset consists of several cases more than 9000 (attribute CaseID)


CaseID 1: consists of a process of 8 rows (each row represents a moment) and in the end the loan is accepted or refused.

CaseID 2: consists of 20 rows

etc etc

I was already able to combine these rows in one row per case and then predict. But I would also like to keep the order of these processes as maybe it might increase the performance of the models (random forest and XGboost).

My question is: How can I say to the model use the first 8 rows of this case to tell me for this case what its label is and use the first 20 rows for the next case and predict its label etc etc?

I hope the question is clear?

  • $\begingroup$ In general, multiple instance learning deals with multiple observations/instances that only together have a value of the dependent variable. For a start, read the Assumptions and Algorithms sections of the wiki article: en.wikipedia.org/wiki/Multiple_instance_learning. $\endgroup$ Feb 10, 2020 at 1:16
  • $\begingroup$ However, your introduction suggests there may be some additional structure to the input rows, esp. a time order to the interactions. In that case maybe something that deals with that more directly is in order. One key question is how you intend to apply the model: what will be known at the time of prediction? $\endgroup$ Feb 10, 2020 at 1:16
  • $\begingroup$ Yes indeed, so the given attributes are: starttime of each case (which is the same within the case for each row), the start time of each activity of each case process. As to the prediction my model should be able to say: After (random number here) 8 activities of a certain case will this case be accepted or refused. Do I need to explain something else about the dataset? Thanks for the time invested :) $\endgroup$ Feb 11, 2020 at 7:58
  • $\begingroup$ I'm not too familiar with this kind of problem, but perhaps an RNN can be of use, see e.g. datascience.stackexchange.com/q/41419/55122 and sciencedirect.com/science/article/pii/S1532046418300996 . Also consider feature engineering to flatten the data into a format better suited for a simpler model. $\endgroup$ Feb 11, 2020 at 14:50
  • 1
    $\begingroup$ I have already used LSTM's for this problem. And it performs well when the process is compressed in one row but I would like a comparison between the compressed and uncompressed.(It's for my thesis) But I have been looking into the multiple instance learning and it seems like it could help me. $\endgroup$ Feb 11, 2020 at 14:54


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.