VAAs have proliferated in the last years in many European countries, and its effects have been thoroughly discussed. One of their main goals is to increase the political competence of the users by providing their closest political party according to ”their own preferences”. For doing that, VAAs implicitly place the users in a conceptual political space and compare their position within it to those of the political parties. In this paper, we define four normative criteria to evaluate the methods to recommend a party to the user: informativeness, respect of the user ”own preferences”, realism and accountability. We argue that in the current method of providing a recommendation the expert chooses deductively the user’s architecture of election in an unjustifiable way. Although the model is theoretically in-formative, the recommendation is not based on the users’ most important preferences and neither in a realistic combination of them. The only alter-native proposed so far has been a social VAA providing a recommended party based only on the other users’ preferences. This model has better empirical results, but we argue that it is unable to increase the political competence of the users. Using a new conceptual framework and machine learning techniques, we propose a new VAA which takes into account both a realistic aggregation of the voter’s preferences and the distances with the political parties to provide a recommended party. We argue that our method combine most of the advantages of the previous methods and use the machine learning tools to model the user’s architecture of election in a more empirical and realistic way. Finally, we explain that the main limitation of our proposed VAA is the lack of accountability.
