Jump to content

Research talk:Anonymous editor acquisition/Signup CTA experiment/Work log/2014-06-12

Add topic
From Meta, a Wikimedia project coordination wiki

Thursday, June 12th

[edit]

Time to gather some data for scoping out in R.

SELECT
    token_info.*,
    experimental_revisions
FROM token_info
LEFT JOIN (
    SELECT
        wiki,
        token,
        COUNT(rev_id) AS experimental_revisions
    FROM staging.token_revision
    WHERE timestamp BETWEEN "20140519180800" AND "20140526180800"
    GROUP BY wiki, token
) AS token_revision_count USING (wiki, token)
WHERE (first_user_id IS NULL OR first_user_registration > "20140519180800")
AND edit_link_clicks > 0

This is going to be a lot of data.

> SELECT
    ->     COUNT(*)
    -> FROM token_info
    -> WHERE (first_user_id IS NULL OR first_user_registration > "20140519180800")
    -> AND first_event BETWEEN "20140519180800" AND "20140526180800"
    -> AND edit_link_clicks > 0
    -> ;
+----------+
| COUNT(*) |
+----------+
|   690362 |
+----------+
1 row in set (1.34 sec)

But I should be able to handle it. --Halfak (WMF) (talk) 20:01, 12 June 2014 (UTC)Reply


The density of revisions per user (token) for eligible anons is plotted for each experimental condition.
Revision density per condition. The density of revisions per user (token) for eligible anons is plotted for each experimental condition.

As expected, this data is very highly skewed. Sadly, I think this means that a geometric mean is not the best approximation that we could use. I'll need to use Mann–Whitney U test to make sure that the differences are actually significant.


> token_activity.stats
      wiki    bucket revs.geo_mean revs.geo_se.upper revs.geo_se.lower      n
 1: dewiki   control    0.10285178        0.10481355        0.10089349  42441
 2: dewiki post-edit    0.09841916        0.10031747        0.09652413  38593
 3: dewiki  pre-edit    0.08392584        0.08576373        0.08209105  39331
 4: enwiki  pre-edit    0.11599710        0.11697464        0.11502041 198936
 5: enwiki   control    0.14224717        0.14329707        0.14119824 208205
 6: enwiki post-edit    0.14108991        0.14216908        0.14001176 189530
 7: frwiki  pre-edit    0.10927743        0.11171140        0.10684879  30115
 8: frwiki post-edit    0.13261317        0.13529830        0.12993438  28649
 9: frwiki   control    0.12999916        0.13250303        0.12750083  31714
10: itwiki  pre-edit    0.12519958        0.12854587        0.12186321  20310
11: itwiki   control    0.15190987        0.15557025        0.14826109  20297
12: itwiki post-edit    0.15375049        0.15764942        0.14986470  1835

Uhoh. It looks like I made a mistake -- and I think I know where it happened. The number of revisions performed by anons in the pre-edit condition is actually lower, not higher. I think that when I ran the SQL last time, I ended up generating stats on a column that had NULLs instead of zeros. I'm going to re-run an old query quick after converting NULLs to zeros and I expect to get a very different result.

SELECT
    wiki,
    bucket,
    ROUND(EXP(AVG(LOG(IFNULL(experimental_revisions, 0)+1)))-1, 3) AS geom_mean_revisions,
    SUM(experimental_revisions > 0) AS editing_clients,
    SUM(experimental_revisions > 0)/COUNT(*) AS editing_prop,
    COUNT(*) AS relevant_tokened_clients
FROM token_info
LEFT JOIN (
    SELECT
        wiki,
        token,
        COUNT(rev_id) AS experimental_revisions
    FROM staging.token_revision
    WHERE TIMESTAMP BETWEEN "20140519180800" AND "20140526180800"
    GROUP BY wiki, token
) AS token_revision_count USING (wiki, token)
WHERE (first_user_id IS NULL OR first_user_registration > "20140519180800")
AND link_clicks > 0
GROUP BY wiki, bucket;
+--------+-----------+---------------------+-----------------+--------------+--------------------------+
| wiki   | bucket    | geom_mean_revisions | editing_clients | editing_prop | relevant_tokened_clients |
+--------+-----------+---------------------+-----------------+--------------+--------------------------+
| dewiki | control   |               0.103 |            3972 |       0.0936 |                    42441 |
| dewiki | post-edit |               0.098 |            3663 |       0.0949 |                    38593 |
| dewiki | pre-edit  |               0.084 |            3005 |       0.0764 |                    39331 |
| enwiki | control   |               0.142 |           25751 |       0.1237 |                   208205 |
| enwiki | post-edit |               0.141 |           23602 |       0.1245 |                   189530 |
| enwiki | pre-edit  |               0.116 |           19673 |       0.0989 |                   198936 |
| frwiki | control   |               0.130 |            3750 |       0.1182 |                    31714 |
| frwiki | post-edit |               0.133 |            3448 |       0.1204 |                    28649 |
| frwiki | pre-edit  |               0.109 |            2881 |       0.0957 |                    30115 |
| itwiki | control   |               0.152 |            2562 |       0.1262 |                    20297 |
| itwiki | post-edit |               0.154 |            2325 |       0.1266 |                    18359 |
| itwiki | pre-edit  |               0.125 |            2063 |       0.1016 |                    20310 |
+--------+-----------+---------------------+-----------------+--------------+--------------------------+
12 rows in set (1 min 28.65 sec)

There we do. Well... This is *bad* news I guess. Better to be right then wrongfully ignorant. --Halfak (WMF) (talk) 21:35, 12 June 2014 (UTC)Reply


The geometric mean revisions saved per token is plotted for the experimental period with geometric standard errorbars.
Anon activity. The geometric mean revisions saved per token is plotted for the experimental period with geometric standard errorbars.

That shows the situation pretty clearly. The "productivity" as measured by edit count goes down for editors in the pre-edit condition. --Halfak (WMF) (talk) 22:25, 12 June 2014 (UTC)Reply


And now for a quick change of pace. Steven (WMF) asked in IRC how many impressions of the CTA users generally saw.

A histogram of the number of CTA impressions per user editor is plotted.
CTA impressions. A histogram of the number of CTA impressions per user editor is plotted.

Looks like we got a ton more impressions for the pre-edit click -- which makes perfect sense. Getting a second impression was pretty unlikely. --Halfak (WMF) (talk) 22:59, 12 June 2014 (UTC)Reply