Online Retweet Recommendation with Item Count Limits
by Xiaoqi Zhao, Keishi Tajima
Abstract
Some Twitter accounts provide information to the followers not by
publishing their own tweets but by retweeting (i.e., forwarding)
useful information from their friends. These accounts need to select
an appropriate number of tweets that match the followers' interests.
If they retweet too many or too few tweets, it annoys the followers or
degrade the value of the accounts. They also need to retweet them in
a timely manner. If they retweet a tweet long after they receive it,
the informational value of the tweet may diminish before the followers
read it. There is, however, a trade-off between these two
requirements. If they select tweets after seeing all the candidates,
they can select the best given number of tweets, but in order to
provide timely information, they have to decide to (or not to) retweet
each tweet before seeing all the following candidates. In order to
help the management of such Twitter accounts, we developed a system
that reads a sequence of tweets from the friends one by one, and
select a given number of (or less) tweets in an online (or
near-online) fashion. In this paper, we propose four algorithms for
it. Two of them give priority to the timeliness, and make a decision
immediately after reading a new tweet by comparing its score with a
threshold. The other two give priority to the selection quality, and
make a decision after seeing some following tweets: after seeing
incoming tweets for a fixed length of time
or after seeing a fixed number of tweets. The former two
are truly online algorithms and the latter two are near-online
algorithms. Our experiment shows that the near-online algorithms
achieve high selection quality only with acceptable time delays.