If the Altmetric Attention Score for a research output decreases, it's usually due to one of two things:
Post Deletion
If a social media user (e.g Facebook page administrator, Google+ user or Twitter user) deletes a post or deletes their whole profile, we would automatically remove the post(s) from the relevant details pages, which would cause the score for the research output to go down.
We don't automatically remove blog posts or news stories if they are taken down from sites, but if the author of the blog or news story asked us to remove a blog or news story, we would delete the mention from the relevant details page.
Modifiers for Twitter
The social media modifiers mean that if we deem a Twitter account to have a high bias and a high promiscuity level, those tweets will contribute less to the score than tweets from unbiased, non-promiscuous accounts. Our algorithm classes a Twitter account as "promiscuous" if that person tweets about scholarly content a lot. This is not a problem in itself, and tweets from this account would not necessarily be scored lower than tweets from other accounts. It's more that the value of tweets from these accounts can degenerate over time, if our algorithm calculates that more than a certain number of these tweets are about the same domain. If we re-classify a tweeter as having a high bias and promiscuity level, their tweets would contribute 0.25 towards the Altmetric Attention Score, rather than 1. When the research output is re-scored, the score will likely decrease for this reason.
The modifiers mean that (for example), tweets from journal accounts are often scored lower than tweets from individual researchers or members of the general public, as our algorithm calculates that this account is tweeting lots of papers from the same domain, which is likely for promotional purposes, rather than out of general interest.
Our perspective is that tweets from accounts that share articles for promotional purposes shouldn't be scored the same as tweets that share articles out of pure interest, as the two types of attention are slightly different, and the implied "impact" of the research output is not the same.
Changes to the Scoring Algorithm
On very rare occasions, we do make small changes to the scoring algorithm and re-score all the articles in the database. This can cause the scores for some articles to decrease.