Recsys Metric : MRR

[shivance]

Fixes #2631

Adding Mean Reciprocal Rank metric used in Recommendation Systems with tests that can directly be used
Reference

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[shivance]

cc : @vfdev-5 codeformatting, and unit tests are yet pending
This is the initial PR, as I am new to ignite ecosystem, looking for initial feedback !

[vfdev-5]

@shivance have you tried to run this locally ? I also wonder about expected result (ref implementation). The link on https://github.com/zuoxingdong/recsys_metrics/blob/master/recsys_metrics/mean_reciprocal_rank.py is not that satisfatory to me. Can you make a quick research on MRR and find another package computing it (something like sklearn)

Also please read carefully this : https://pytorch.org/ignite/metrics.html#ignite-metrics
We prefer to perform online computation (by batches) => precompute and accumulate things and finally compute the metric value instead of storing all predictions and targets and compute the metric on compute.
Take a look how Accuracy metric is implemented in ignite as example:

ignite/ignite/metrics/accuracy.py

Lines 221 to 222 in b76b825

	self._num_correct = torch.tensor(0, device=self._device)
	self._num_examples = 0

Main question is about if we can correctly accumulate self._relevance using batches vs if we were computing relevance over all data.

[vfdev-5]

This line is incorrect, but let's fix it later once we have a good understanding of how metric is computed