Deduce synthetic carry events for event providers which don't annotate them

[lodevt]

Addressing Issue #379

I tried to follow a similar approach as in socceraction to add carry events for providers which don't annotate them: (https://github.com/ML-KULeuven/socceraction/blob/af297a490fc521d27622b947e93fa15c5e0092da/socceraction/spadl/base.py#L38).

The implementation and tests are not complete yet. I think that a great way to test this is to remove the carry events from a StatsBomb dataset and then compare the generated carries with the carries from the original dataset. It is impossible to have an accuracy of 100% but still the output should have high similarities.

I am happy to hear any feedback while further implementing this.

[lodevt]

To test this, I removed the carry events from a StatsBomb dataset and considered these events as the ground truth to compare the generated carries with.

True positives are generated carries where a matching StatsBomb carry from the same player within 5 seconds time is found.
False positives are generated carries where no matching StatsBomb carry is found.
False negatives are StatsBomb carries with a length greater than the min_carry_length_meters (3 meters) that have no match in the generated carries.

The current implementation reaches as accuracy of around 89% (466 TP, 38 FP, 18 FN).

[lodevt]

What do you think would be the best convention to follow to generate the event_ids of the synthetic carries?

The StatsBomb deserialiser follows the convention '<event_type>-<previous_event_id>' (see duels and ball out) and the Wyscout V3 deserializer similarly uses 'synthetic-<team_id>-<previous_event_id>' for synthetic formation change events (see formation change).

Based on this, I suggest using 'carry-<previous_event_id>' as the event_id for the synthetic carries.