Research the detection of the primary name for an identified person

[zurk]

Now we have just a number that joins all person's identities together. But we should also set the name of the person.

See related discussion:
https://src-d.slack.com/archives/CJQ0DBAJV/p1564057922120100

one thing I've learned from the current demo and identity matching is that we need a simple way for the algorithm to determine a primary name

Plan:

• Get the data for proper validation. We may use Github API. We should also check data quality because what people set in their GitHub profile and use in commits can be different. We cannot predict something that does not appear in commits.
• Test simple heuristics like
  • take the most frequent name
  • take the longest name
  • ???
• If simple heuristics do not give good performance train something like GBDT.

[vmarkovtsev]

The primary user names are included in GHTorrent.
@EgorBu Can you please attach here the best identity matching results on GitHub?

[EgorBu]

I will launch the extraction of identities and dump it here

Update:
Old result found here: /user/egorbu/idmatching/cache/res_name_threshold_5_email_threshold_28_data_aggregated_deduplicated.pkl - it was launched on initial dataset (not full)
I will launch extraction on full dataset.

[vmarkovtsev]

No, the user names are no longer available. I can take all the logins from GHTorrent, but the names require some GitHub crawling. I will have to run it.

[vmarkovtsev]

I am crawling the full name for each github login on our 11 nodes. ETA 10 days.

[vmarkovtsev]

Fcking GitHub is throttling us with 429. I had to add download delays using the Scrapy's AutoThrottle extension, that will most probably increase ETA.

#38

Fetched ~900k in 10 hours.

[vmarkovtsev]

My experiments showed that the rate-limiting happens if you do more than or equal to 5 requests per second. I set the frequency to 4 and the process has finally stabilized.

[vmarkovtsev]

Progress: 4.6mm through the weekend

[vmarkovtsev]

I collected the dataset of full names: /user/internal-datasets/full_names

Random strategy

46% accuracy if we consider identities without a correct name. Otherwise, ~82%.

Biggest number of commits strategy

56% and 99.9%, respectively.

So 99.9% is a very good number and there is no need for digging deeper. We should count commits for each name and sort them, problem solved.