Refinements to macaronic user modeling
Things to consider regarding the testing of macaronic user model performance.
Randomly splitting test and train data. This feels wrong because there is an inherent assumption of stationarity. Users’ learning on the other hand is constantly changing as the perform more hits. Consider a test example of a user doing their first sentence, while we have trained the model on several examples of the user doing tasks after a lot more experience. User x Experience adaptation might help, but would increase the number of features needed to be learned.
Improving speed of inference. Implement top k using matrix masking. Top k would cause some issues. What if the messages from factors to a node do not overlap? The end result is a uniform message to the node. This seems bad, especially because the smaller k gets the more likely this will happen. If only there was some method where a node accepts all incoming messages (which have been top k truncated) and instead of just taking the product of the messages, it is aware of the truncation.
More to follow