search

boldchan / KGReasoning

0 stars 0 forks source link

ICEWS14_forecasting object index #6

Closed boldchan closed 4 years ago

boldchan commented 4 years ago

In ICEWS18_forecasting, entities in training set are indexed consecutively starting from 0. While when I preprocess ICEWS14 I happened to be ignored this property. Currently this change doesn't cause any error. But a NeighborFinder for training data only cannot be instantiated without modification.

https://github.com/boldchan/KGReasoning/blob/5add4963f1fbbf5e9a57d5ee14181a5f4f7451ba/tKGR/utils.py#L208-L226

boldchan commented 4 years ago

More explanation for the mistakes: I first loaded data from ICEWS14_completion, and indexed entity according to the appearance in training data, valid data and test data. But in completion dataset, events in valid data may happen later than events in training data. Therefore, in the generated training dataset, some entities are indexed with big number since they are in the valid or test set in ICEWS_completion.

And another mistake I didn't realise then: I split the dataset according to the amount of events. therefore I split events in same timestamp in training data and validation data, and also in valid data and test data.