The rise of the Web 2.0 has made content publishing easier than ever. Yesterday’s passive consumers are now active users who generate and contribute new data to the web at an immense rate. We consider evaluating data driven aggregation queries which arise in Web 2.0 applications. In this context, each user action is interpreted as an event in a corresponding stream e.g., a particular weblog feed, or a photo stream. The presented approach continuously tracks the most popular tags attached to the incoming items and based on this, constructs a dynamic top-k query. By continuous evaluation of this query on the incoming stream, we are able to retrieve the currently hottest items. To limit the query processing cost, we propose to pre-aggregate index lists for parts of the query which are later on used to construct the full query result. As it is prohibitively expensive to materialize lists for all possible combinations, we select those tag sets that are most beneficial for the expected performance gain, based on predictions leveraging traditional FM sketches. To demonstrate the suitability of our approach, we perform a performance evaluation using a real-world dataset obtained from a weblog crawl.