In a recent tutorial on association and scalar clusters, http://www.miislita.com/information-retrieval-tutorial/association-scalar-clusters-tutorial-1.pdf, I introduced a back mapping technique wherein once features conforming clusters are extracted from objects, the clusters are mapped back to objects.

The technique works well with clusters of terms extracted from documents. The reverse case is also possible: given a cluster of documents extracted from terms, it is possible to map these back to terms.

What do we gain from such two-way manipulations? A lot. Consider the first scenario: Mapping term clusters back to documents; a tutorial on the second scenario will be available soon.

Back Mapping Term Clusters to Documents

A document is just a distribution over topics while topics are distributions over words. Thus, across a collection of documents there are topics hidden (latent) and waiting to be uncovered. Back mapping allows us to recover these, precisely.

Combinations of terms that do not amount to topics across the collection are discovered as well. Reasonably, one would expect these to be least relevant across other documents than those distributed across the collection. In addition, one would expect documents traced back to clusters to be the most relevant documents, from the collection and with respect to the topics.

The implications of this for search engine optimization and keyword bidding are quite obvious. Implementation is straightforward. To learn more about it, read Part 1 of the tutorial.

Advertisements