Monthly Analysis of Automated Posts

It's been a month since I first looked at the rising tide of automated posts. Their escalation had been rapid, rising to 5,000 posts per day and approaching 20% of overall post volumes. How have things changed in the intervening month? Time to take another look.

Here's where we left the story. This chart shows the number of articles and rewards from users that posted over 1000 times (articles not comments) in a 30 day period from October 25 to November 24. Our light blue "articles per day" line is heading off the chart!

There were only 32 such accounts. Their headline statistics over the thirty day period were:

• 32 accounts, representing less than 0.1% of all accounts that posted.
• 107,956 posts, representing 17.5% of all articles posted.
• $3,341 in total rewards (author and curation added together) representing 0.3% of all rewards.

Here's the table of the top 20 accounts, showing the huge volumes in comparison to the small rewards:

Ranking	Posts	Total Rewards
20	2075	67.32
19	3093	92.65
18	3202	14.63
17	3310	65.48
16	3451	73.61
15	3716	155.63
14	3853	62.09
13	3924	101.01
12	3927	80.82
11	3931	92.13
10	4416	75.36
9	5229	73.02
8	5241	66.85
7	5770	72.2
6	6038	94.79
5	6164	118.79
4	6260	156.52
3	6382	78.99
2	6714	195.24
1	6757	1105.56

You can see the impact on these users on the right hand side of this distribution (number of posts per user over the 30 days along the x-axis, total posts generated on the y-axis).

Lots of posts being generated by very few users causing a bump in the tail of the distribution.

---

If we catch up with these same 32 users again one month later we see an interesting phenomenon: An almost vertical drop off of production, down by 80%. It looks like Steem's teams of cleaners moved into action (a quick check of the @spaminator account shows a number of these users have been listed in the "Spam, Comment & Post Farming Reports" and downvoted).

Problem solved? Well, let's have a look. If I run the same analysis again, this time from November 25 to December 24 I get the following distribution.

Our bump is still there!

---

Over the November - December window there were 49 accounts posting more than 1000 times. A rise of 17 users! Their headline statistics over the thirty day period were:

• 49 accounts, still representing less than 0.1% of all accounts that posted.
• 121,602 posts (a rise), representing 15.2% of all articles posted (a fall in percentage terms).
• $1,623 in total rewards (author and curation added together) representing less than 0.1% of all rewards (a significant fall).

Here's the table of the top 20 accounts for comparison to the above:

Ranking	Posts	Total Rewards
20	2343	0.81
19	2351	1.61
18	2419	10.25
17	2443	8.88
16	2444	16.87
15	2448	1.2
14	2476	4.57
13	2497	9.8
12	2514	15.29
11	2640	24.94
10	2758	21.06
9	2875	21.16
8	3006	49.52
7	4056	19.93
6	4138	635.15
5	4423	7.71
4	5645	17.13
3	6218	16.84
2	6380	57.44
1	7044	164.49

And here's the chart of article numbers and rewards from these 49 accounts:

So we have some decline from peak spam, but much less marked than in our previous chart. So what has happened?

---

Well we can break down the 49 accounts into three types:

• Those accounts unaffected and broadly unchanged over the period (approximately 10 accounts).
• Those accounts that stopped posting in large volume around the 11th December (approximately 20 accounts).
• Those accounts that started posting in large volume around the 12th December (approximately 20 accounts).

Conclusions

It's a slightly depressing chart in some ways but, with a little bit of licence to make assumptions, we can jump to two useful conclusions:

One: The actions of the steem cleaning teams do make a difference in stopping spam.

Two: Constant vigilance is required!

---

Questions

That's all for today.

If you have any questions or spot any errors please do not hesitate to leave a comment.

---

Methodology and Tools for Analysis

Tools

Raw data was obtained through sql queries of steemsql using Valentina Studio.
Data was analysed in LibreOffice and illustrated in Numbers (spreadsheet tools).
Data was obtained for various timescales, including a particular focus on October 25 to November 24 and comparison to November 25 to December 24.

SQL query

I used the following SQL query:

SELECT
    Comments.author,
    Count(Comments.author) AS [Posts],
    Count(distinct Comments.author) AS [DistinctCommentAuthor],
    count(Comments.parent_author) AS [ParentAuthor],
    count(distinct Comments.parent_author) AS [DistinctParentAuthor],
    sum(CONVERT(REAL,Comments.pending_payout_value)) AS [PendingPayoutValue],
    sum(CONVERT(REAL,Comments.curator_payout_value)) AS [CuratorPayoutValue],
    sum(CONVERT(REAL,Comments.total_payout_value)) AS [TotalPayoutValue]

FROM
    Comments (NOLOCK)

WHERE
    (YEAR(Comments.created) = 2017 AND MONTH(Comments.created) = 10 AND DAY(Comments.created) > 24 and depth = 0 ) OR
    (YEAR(Comments.created) = 2017 AND MONTH(Comments.created) = 11 AND DAY(Comments.created) < 25 and depth = 0) 

GROUP BY
    Comments.author

---

Posted on Utopian.io - Rewarding Open Source Contributors

---