Suggestions for future steps in temporal team formation

[VaghehDashti]

Hi @hosseinfani, Here are a few suggestions I have for future steps on temporal team formation:

• We can develop several temporal negative sampling heuristics. We can use the teams/experts (inverse) unigram distributions from past years to create negative samples for current years, e.g., use the distribution of t-1 when optimizing on training on time t. This can be done in several ways: we can use the data from 1 year ago, or 1 year ago to N years ago, or from beginning of data until 1 year ago.
• Using temporal graph embeddings to generate embeddings for input skills of neural model and use streaming learning on top of that.
• Since we have two forms of testing, shuffled test and future test, it can get confusing which folder is for which test set, so I think we should add a annotations like 'ftp' (future team prediciton) or 'ft' (future test) and 'st' (shuffled test).
• Right now, we use the data from last step aheads to create test set for future team prediction, however, the problem with this method is that some datasets have very few teams in the last time intervals like imdb which has less than 100 teams in the last two years. Another route is taking the last 15% of the sorted by time data for test set and use the rest for training and validation.

[hosseinfani]

@VaghehDashti Thank you very much for the suggestions. We'll take care of them in our future extensions.