The following is an email conversation that I had with Mark Wang about ORES. I'm posting this so that it'll maybe gain some long-term usefulness. As you can see at the end of the thread, Mark agreed to me posting this publicly.
On Sat, Nov 17, 2018 at 8:50 PM Wang, Mark <> wrote:
Hi Scoring Platform Team!
I'm Mark, a CS student at Brown Univ. I'm working on experimenting with applying Snorkel (a framework for leveraging unlabeled data with noisy label proposals) to detect paid Wikipedia edits. I've got a few selfish requests / questions for you guys.
Snorkel code : https://github.com/HazyResearch/snorkel Snorkel paper : https://arxiv.org/pdf/1711.10160.pdf
Some selfish questions:
1) Is it possible for me to have access to edits and page-stats data that you work with? I can scrape them myself (with a reasonable crawl rate), but of course, it's less convenient and I'll end up working with less data.
2) How do you represent revisions? I'm thinking about using character embeddings here. What are some methods that worked well for you guys? And what should I probably not try?
3) What features seem to be strongly informative in your models for detecting low-quality edits?
4) Any additional recommendations /advice?
Thank you in advance for your time, Mark Wang
On Mon, Nov 19, 2018 at 4:49 PM Aaron Halfaker <ahalfaker@wikimedia.org> wrote: Hi Mark!
Thanks for reaching out! Have you seen our recent data release of known paid editors? https://figshare.com/articles/Known_Undisclosed_Paid_Editors_English_Wikipedia_/6176927
1) I'm not sure what page stats you are looking for, but you can see the features we use in making predictions by adding a "?feature" argument to an ORES query. For example, https://ores.wikimedia.org/v3/scores/enwiki/21312312/damaging?features shows the features extracted and a "is this edit damaging" prediction for https://en.wikipedia.org/wiki/Special:Diff/21312312
2) A revision is a vector that we feed into the prediction model. We do a lot of manual feature engineering, but we use vector embeddings for topic modeling. We're actually looking into just using our current word2vec strategies for implementing better damage detection too. See https://phabricator.wikimedia.org/T197007
3) Here's an output of our feature importance weights for the same model. This is estimated by sklearn's GradientBoosting model.
feature.log((temporal.revision.user.seconds_since_registration + 1)) 0.131 feature.revision.user.is_anon 0.036 feature.english.dictionary.revision.diff.dict_word_prop_delta_sum 0.033 feature.revision.parent.markups_per_token 0.029 feature.revision.parent.words_per_token 0.028 feature.revision.parent.chars_per_word 0.027 feature.log((wikitext.revision.parent.ref_tags + 1)) 0.026 feature.revision.diff.chars_change 0.026 feature.revision.user.is_patroller 0.026 feature.english.dictionary.revision.diff.dict_word_prop_delta_increase 0.025 feature.log((wikitext.revision.parent.chars + 1)) 0.023 feature.log((AggregatorsScalar(<datasource.tokenized(datasource.revision.parent.text)>) + 1)) 0.023 feature.log((AggregatorsScalar(<datasource.wikitext.revision.parent.words>) + 1)) 0.023 feature.revision.parent.uppercase_words_per_word 0.022 feature.log((wikitext.revision.parent.wikilinks + 1)) 0.021 feature.log((wikitext.revision.parent.external_links + 1)) 0.02 feature.log((wikitext.revision.parent.templates + 1)) 0.02 feature.wikitext.revision.diff.markup_prop_delta_sum 0.02 feature.english.dictionary.revision.diff.non_dict_word_prop_delta_sum 0.02 feature.log((AggregatorsScalar(<datasource.wikitext.revision.parent.uppercase_words>) + 1)) 0.018 feature.revision.diff.tokens_change 0.018 feature.log((wikitext.revision.parent.headings + 1)) 0.017 feature.wikitext.revision.diff.markup_delta_sum 0.015 feature.revision.diff.words_change 0.015 feature.english.dictionary.revision.diff.dict_word_delta_sum 0.015 feature.english.dictionary.revision.diff.dict_word_prop_delta_decrease 0.015 feature.english.dictionary.revision.diff.non_dict_word_prop_delta_increase 0.015 feature.revision.diff.markups_change 0.014 feature.english.dictionary.revision.diff.dict_word_delta_increase 0.014 feature.wikitext.revision.diff.markup_prop_delta_increase 0.013 feature.wikitext.revision.diff.markup_delta_increase 0.012 feature.wikitext.revision.diff.number_prop_delta_sum 0.011 feature.wikitext.revision.diff.number_prop_delta_increase 0.011 feature.english.dictionary.revision.diff.non_dict_word_delta_sum 0.011 feature.wikitext.revision.diff.number_delta_increase 0.01 feature.revision.diff.wikilinks_change 0.01 feature.revision.comment.has_link 0.01 feature.english.dictionary.revision.diff.dict_word_delta_decrease 0.01 feature.revision.page.is_mainspace 0.009 feature.wikitext.revision.diff.number_delta_sum 0.009 feature.wikitext.revision.diff.markup_prop_delta_decrease 0.008 feature.english.dictionary.revision.diff.non_dict_word_prop_delta_decrease 0.008 feature.revision.page.is_articleish 0.007 feature.revision.diff.external_links_change 0.007 feature.revision.diff.templates_change 0.007 feature.revision.diff.ref_tags_change 0.007 feature.english.informals.revision.diff.match_prop_delta_sum 0.007 feature.english.informals.revision.diff.match_prop_delta_increase 0.007 feature.wikitext.revision.diff.number_prop_delta_decrease 0.006 feature.revision.comment.suggests_section_edit 0.006 feature.english.dictionary.revision.diff.non_dict_word_delta_increase 0.006 feature.wikitext.revision.diff.markup_delta_decrease 0.005 feature.revision.user.is_bot 0.005 feature.revision.user.is_admin 0.005 feature.english.badwords.revision.diff.match_prop_delta_sum 0.005 feature.wikitext.revision.diff.number_delta_decrease 0.004 feature.wikitext.revision.diff.uppercase_word_prop_delta_sum 0.004 feature.revision.diff.headings_change 0.004 feature.revision.diff.longest_new_repeated_char 0.004 feature.english.badwords.revision.diff.match_prop_delta_increase 0.004 feature.english.informals.revision.diff.match_delta_increase 0.004 feature.english.dictionary.revision.diff.non_dict_word_delta_decrease 0.004 feature.wikitext.revision.diff.uppercase_word_delta_sum 0.003 feature.wikitext.revision.diff.uppercase_word_prop_delta_increase 0.003 feature.revision.diff.longest_new_token 0.003 feature.english.informals.revision.diff.match_delta_sum 0.003 feature.wikitext.revision.diff.uppercase_word_delta_increase 0.002 feature.wikitext.revision.diff.uppercase_word_prop_delta_decrease 0.002 feature.english.badwords.revision.diff.match_delta_sum 0.002 feature.english.badwords.revision.diff.match_delta_increase 0.002 feature.wikitext.revision.diff.uppercase_word_delta_decrease 0.001 feature.english.informals.revision.diff.match_prop_delta_decrease 0.001 feature.revision.page.is_draftspace 0.0 feature.revision.user.has_advanced_rights 0.0 feature.revision.user.is_trusted 0.0 feature.revision.user.is_curator 0.0 feature.english.badwords.revision.diff.match_delta_decrease 0.0 feature.english.badwords.revision.diff.match_prop_delta_decrease 0.0 feature.english.informals.revision.diff.match_delta_decrease 0.0
4) You'll note that time since registration and is_anon are strongly predictive. They don't overwhelm the predictions -- we can still differentiate good from bad among newcomers and anonymous editors. But the model generally doesn't predict that an edit by a very experienced editors is bad regardless of what's actually in the edit. The more we can move away from relying is_anon and seconds_since_registration, the more we'll be targeting the things that people do -- rather than targeting them for their status. See section 7.4 our systems paper for a more substantial discussion of this problem.
-Aaron
On Mon, Nov 19, 2018 at 6:47 PM Wang, Mark <> wrote:
Thanks a bunch for your help Aaron! This is all very informative.
One more question from me: May I borrow your features? And if so, is accessing them through the API the preferred method of access for an outsider?
Thanks again, Mark
On Tue, Nov 20, 2018 at 11:07 AM Aaron Halfaker <ahalfaker@wikimedia.org> wrote:
Say, I'd like to save this conversation publicly so that others might benefit from it. Would you be OK with me posting our discussion publicly on a wiki?
On Tue, Nov 20, 2018 at 10:06 AM Aaron Halfaker <ahalfaker@wikimedia.org> wrote: Yes. That is a good method for accessing the features. You'll notice that the features that the API reports are actually just the basic reagents for the features the model uses.
For example, we have features like this:
- words added
- words removed
- words add / words removed
- log(words added)
- log(words removed)
- etc.
In all of these features, the basic foundation is "words added" and "words removed" with some mathematical operators on top. So we only report those two via the API. To see the full set of features for our damage detection model, see https://github.com/wikimedia/editquality/blob/master/editquality/feature_lists/enwiki.py See also a quick overview I put together for feature engineering here: https://github.com/wikimedia/revscoring/blob/master/ipython/feature_engineering.ipynb
If I wanted to extract the raw feature values for the English Wikipedia "damaging" model, I'd install the "revscoring" library (pip install revscoring) and then run the following code from the base of the editquality repo:
$ python
Python 3.5.1+ (default, Mar 30 2016, 22:46:26)
[GCC 5.3.1 20160330] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from editquality.feature_lists.enwiki import damaging
/home/halfak/venv/3.5/lib/python3.5/site-packages/sklearn/cross_validation.py:41: DeprecationWarning: This module was deprecated in version 0.18 in favor of the model_selection module into which all the refactored classes and functions are moved. Also note that the interface of the new CV iterators are different from that of this module. This module will be removed in 0.20.
"This module will be removed in 0.20.", DeprecationWarning)
>>> from revscoring.extractors import api
>>> import mwapi
>>> extractor = api.Extractor(mwapi.Session("https://en.wikipedia.org"))
Sending requests with default User-Agent. Set 'user_agent' on mwapi.Session to quiet this message.
>>> list(extractor.extract(123456789, damaging))
[True, True, False, 10.06581896445358, 9.010913347279288, 8.079927770758275, 3.4965075614664802, 2.772588722239781, 5.402677381872279, 2.70805020110221, 1.791759469228055, 2.1972245773362196, 7.287484510532837, 0.3940910755707484, 0.009913258983890954, 0.06543767549749725, 0.0, 2.0, -2.0, 0.04273504273504275, 0.15384615384615385, -0.1111111111111111, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 2.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1, 1, False, False, False, False, True, False, False, 11.305126087390619, False, False, 0, 0, 0, 0.0, 0.0, 0.0, 0, 0, 0, 0.0, 0.0, 0.0, 0, 0, 0, 0.0, 0.0, 0.0, 0, 0, 0, 0.0, 0.0, 0.0]
This extracts the features for this edit: https://en.wikipedia.org/w/index.php?diff=123456789
-Aaron
Hi Aaron:
Thank you so much! This is all so helpful. And of course, feel free to publicize any of our conversations.
Mark