I have a dataset of parts/pieces of a whole, which I'll refer to as the end-item. The goal will be to take a list of these pieces and their attributes and then provide essentially the predict_proba, where we get the end-items that each piece could part of. There are hundreds of thousands of end-items, and therefore, an even higher number of parts/pieces. Example: we have nuts and bolts that would be part of item A, and are also on item D.

I have tried KNN, linear SVM, RBF SVM, decision trees, random forest, nns, adaboost, naive bayes, and QDA. None of these are working quite well.

Any modeling technique recommendations for this type of problem?

  • Are you performing a single multiclass classification or n - n being the number of parts in your end-item - classifications ? – AshOfFire Apr 19 at 15:50
  • I guess the more accurate term is multi-label classification. For each part, a multi-label classification is needed. Does that clear things up? – jma Apr 19 at 16:01
  • Yes, definitely. I would try problems transformations depending on the characteristics of your dataset : binary relevance if you assume that parts/pieces are not correlated, classifier chains if not, or any other method listed here – AshOfFire Apr 19 at 16:28
  • Thanks for the tip. For anyone else with a similar problem, I found this article which is pretty useful – jma Apr 19 at 16:39

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.