Selecting Relevance Thresholds to Improve a Recommender System in a Parliamentary Setting

摘要

In the context of building a recommendation/filtering system to deliver relevant documents to the Members of Parliament (MPs), we have tackled this problem by learning about their political interests by mining their parliamentary activity using supervised classification methods. The performance of the learned text classifiers, one for each MP, depends on a critical parameter, the relevance threshold. This is used by comparing it with the numerical score returned by each classifier and then deciding whether the document being considered should be sent to the corresponding MP. In this paper we study several methods which try to estimate the best relevance threshold for each MP, in the sense of maximizing the system performance. Our proposals are experimentally tested with data from the regional Andalusian Parliament at Spain, more precisely using the textual transcriptions of the speeches of the MPs in this parliament.