FOA Home | UP: Learning which documents to route

User drift and event tracking

One interesting feature of the training set generated by the routing task is the odd distribution of positive and negative examples it generates. Initially we can imagine that this filter is very inaccurate; i.e., we are likely to see many negative examples. Later, when we hope it is well-trained, the filter has nearly perfect performance and the system gets very few negative examples.

Further, no user's interests remain static. As discussed in the next chapter, one common purpose for the FOA activity is to become educated(cf. Section §8.3.4 , and this is an elusive, ever-changing goal. The world changes, and what they read changes a user's opinion of what needs to be done and what the new questions are. In brief, documents they used to find relevant aren't any longer. This has been called CONCEPT DRIFT [Klinkenberg98] . When the world changes, this corresponds to documents and the news they contain changing too. This side of the dynamic is called TOPIC TRACKING [Allan98] [Baker99] . Jaime Carbonell's (of CMU) approach is to first identify that a concept change has occurred, and then adjust a time window on the stream of incoming training data over which a new invariant is then identified (personal communication).

The distribution of RelFbk generated by the filtering task , where a standing query is allowed to adapt to a stream of RelFbk generated by users who receive and evaluate routed documents, (cf. Section §4.3.9 ) provides an especially interesting form of learning task, because of its TEMPORAL DIMENSION . Initially, the set of documents routed to users must depend on the same fundamental matching function shared by other search engine tasks. But as RelFbk in response to the first retrievals comes to affect the users' characterizations of interest, only a skewed sample (relative to the initial distribution) of potential documents is shown to the users, and only these can be the basis of subsequent \RelFbk. This tension between EXPLORATION of the universe of potentially relevant documents and EXPLOITATION of those that prior RelFbk makes it seem are most likely to be perceived as relevant by the users is familiar to other REINFORCEMENT LEARNING situations [Sutton98] .

Top of Page | UP: Learning which documents to route | ,FOA Home