search

actionml / universal-recommender

Highly configurable recommender based on PredictionIO and Mahout's Correlated Cross-Occurrence algorithm
http://actionml.com/universal-recommender
Apache License 2.0
669 stars 172 forks source link

java.lang.NegativeArraySizeException error for secondary indicators #56

Open noeliaoses opened 6 years ago

noeliaoses commented 6 years ago

Hi,

The UR throws a java.lang.NegativeArraySizeException when there isn't at least one occurrence of the primary indicator. But it also throws the same error when no users in the secondary that are also in the primary (no cross-occurrences), as explained by Pat Ferrel in actionml's googlegroup. It would be great if the error messages were different for both cases and/or more informative.

Thank you very much for all the great work!!! noelia

pferrel commented 6 years ago

This will only happen when there is no data or no cross-occurrences. Therefore you should not be using the events in question, it just wastes cycles. I suggest some kind of serious warning here.

To collect data for maybe 10000 users and have no cross-occurrences is a serious data problem. It may mean you are collecting in the wrong way or have encoded user-ids differently in the primary vs the secondary events or the event has no meaning WRT cross-occurrences. Every time we have seen this is was a too small sample or a data encoding error. This should not happen in real-world situations or the recommender will have NO data to use. It means that not one of the users has meaningful data for the event.

As I understand it this happens for you with some test sample, so the error is annoying when doing a proof of concept. My worry is that a non-fatal error will allow data problems or malformed data to not be noticed.

pferrel commented 6 years ago

At very least the error should be more informative, giving the event name and better message.

wanxsb commented 6 years ago

Please decrease minEventsPerUser if you have make sure nothing wrong.