I am attempting to use time-series classification algorithms for fraud detection applications. I have came across several works in the literature that propose novel techniques for multivariate time-series classification, however, most of these approaches treat each feature as an individual signal.

Now, my processing of my data transforms a transactions dataset into a tensor; 1 dimension where each observation is an account, 1 dimension where each element is a transaction and 1 dimension for the transaction attributes. The transactions dataset has a large number of features, many of which are one-hot encoded categorical variables. Therefore, I am not really sure that multivariate time-series classification algorithms such as CNN or LSTM will work in this case, since it will treat every one-hot encoded feature as a signal on its own.

What would be an alternative approach in this case? Would applying PCA on the data to capture the most significant features help instead of the ordinary features?

  • $\begingroup$ Why are you not using ordinal encoding? $\endgroup$ – Brian O'Donnell Jan 10 at 20:27
  • $\begingroup$ Why haven't you tried CNN and LSTM to find out? It doesn't take that much effort. $\endgroup$ – Brian O'Donnell Jan 10 at 20:28
  • $\begingroup$ My only issue with ordinal encoding is figuring out how many encodings per categorical feature I need to generate. I have some categorical features with 20 unique values, and a few other features have over 300. The reason I haven't attempted the methods yet is because I need to preprocess and generate more features, which will be a time consuming step. I want to know if deep learning in this case is feasible, given that there is a huge class imbalance in my dataset as well. $\endgroup$ – Pleastry Jan 11 at 10:06

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.