The 7th International Conference on Modeling Decisions for Artificial Intelligence Modelització de Decisions per a la Intel·ligència Artificial Perpinyà, Catalunya Nord, França, Octubre 27 - 29, 2010 http://www.mdai.cat/mdai2010

Termini de submissió (EXTENDED): 9 Abril, 2010

Abstract: In decision applications, especially multicriteria decision-making, numerical approaches are often questionable because it is hard to elicit numerical values quantifying preference, criteria importance or uncertainty. More often than not, multicriteria decision-making methods come down to number-crunching recipes with debatable foundations. One way out of this difficulty is to adopt a qualitative approach where only maximum and minimum are used. Such methods enjoy a property of scale invariance that insures their robustness. One of the most sophisticated aggregation operation making sense on qualitative scales is Sugeno integral. It is not purely ordinal as it assumes commensurability between preference intensity and criteria importance or similarly, utility and uncertainty. However, since absolute qualitative value scales must have few levels so as to remain cognitively plausible, there are as many classes of equivalent decisions as value levels. Hence this approach suffers from a lack of discrimination power. In particular, qualitative aggregations such as Sugeno integrals cannot be strictly increasing and violate the strict Pareto property. In this talk, we report results obtained when trying to increase the discrimination power of Sugeno integrals, generalizing such refinements of the minimum and maximum as leximin and leximax. The representation of leximin and leximax by sums of numbers of different orders of magnitude (forming a super-increasing sequence) can be generalized to weighted max and min (yielding a "big-stepped" weighted average) and Sugeno integral (yielding a "big-stepped" Choquet integral). This methodology also requires qualitative monotonic set-functions to be refined by numerical set-functions, and we show they can always be belief or plausibility functions in the sense of Shafer.

Abstract: Web search engines gather a lot of information on the preferences and interests of users. They actually gather enough information to create detailed user profiles which might enable re-identification of the individuals to which those profiles correspond, e.g. thanks to the so-called vanity queries or to linkage of several queries known to have been submitted by the same user. In this way, a broadly used search engine like Google becomes a ``big brother'' in the purest Orwellian style. In this talk, a survey will be offered of the solutions which have been proposed to preserve anonymity in web search and to fight profile creation. We will start with Private Information Retrieval (PIR) and we will highlight its lack of practicality. We will then look at some relaxations of PIR, based on standalone defense by the user or on a defense based on a peer-to-peer community in which one user submits queries by other users and viceversa. Finally, we will sketch a new theory, called co-privacy or co-operative privacy, whose goal is to find out under which conditions the best rational option for a peer-to-peer user is to help other peers in preserving their privacy.