FOA Home | UP: Background


Alternative Tasks for Learning

One obvious task we might ask of a learning algorithm is to learn the rank function. Following the Probability Ranking Principle §5.5.1 , we might attempt to learn the probability that a document is relevant, given a set of features it contains. Recall from Section §4.2.1 how RelFbk information can be used to modify a query vector so as to move it closer to those documents a user has marked as relevant. The same technique can be used to move documents towards the query! The radically different consequence of this change is that while query modification is only useful once to the user benefiting from it, document modification changes the document's representations for all subsequent users queries.

We will also be concerned with another task. Rather than assigning a single real number to each document ($for example), we shall attempt (using the same set of document features) to classify a document into one of a small number of potential categories. Most simply, we may be interested in BINARY classification of documents into one of two categories. An obvious use of binary classification would be to classify into Relevant and NonRelevant classes. More complex classifications into one of $C$ different classes are also possible. For example, imagine that you would like to have your Email client automatically sort your incoming Email into various mail folders you have used historically. Having decided how many classes to use, we must also determine whether our ASSIGNMENT to these classes is binary or whether we provide the probability that it belongs to a class. Classification techniques are discussed in §7.4 .


Top of Page | UP: Background | ,FOA Home


FOA © R. K. Belew - 00-09-21