14.00, Room 445 at PCRI

Abstract
Given a set of objects O and a boolean interest function q:2^O-> {true, false}, the border of 2^O is the set of extremal (minimal or maximal) subsets o of O s.t. q(o)=true. This concept is present in many contexts, e.g., maximal frequent
itemsets, functional dependencies and partial matrialization of data cubes. We present a parallel algorithm for computing borders when q is anti-monotonic and show its performance from theoretical and experimental point of views. Some extensions will be discussed s.t. distributed data and map reduce.

Resumé
Etant donnés un ensemble d’objets O et une fonction booléenne d’intérêt q: 2^O-> {vrai, faux}, la bordure de 2^O est l’ensemble des éléments o de 2^O extrémaux (minimaux ou maximaux) tels que q(o)=vrai. On retrouve le concept de bordures dans plusieurs applications, ex: les itemsets fréquents maximaux, les dépendances fonctionnelles approximatives et le stockage
partiel des cubes de données. Nous présentons un algorithme parallèle pour le calcul des bordures lorsque q est anti-monotone, en discutons ses performances théoriques et expérimentales. Quelques extensions seront abordées comme le cas des données distribuées et son utilisation sous le paradigme map reduce.

14:00, Room 445 at PCRI

Abstract
As both the volume of data and the diversity of users accessing them increase, user preferences offer a useful means towards improving the relevance of the query results to the information needs of the specific user posing the query. In this talk, we will focus on enhancing preferences with context. Context may express conditions on situations external to the database or related to the data stored in the database. We will outline models for expressing both types of preferences. Then, given a user query and its surrounding context, we will consider the problem of selecting related preferences to personalize the query.

15:45, Room 445 at PCRI

Abstract
Nowadays, many enterprises and organizations are faced with large volumes of data that have to be analyzed in a per-day basis. In particular, scientific datasets are growing at unprecedented rates and are likely to continue growing to the order of Exabytes. These current needs of data management require applications to run over a large number of computing nodes. However, databases management systems (DBMS) have proven inefficient to deal with very large datasets as well as to scale out to a large number of computing nodes. In this context, MapReduce and the Cloud computing are two alternative technologies that respond to this challenge. While MapReduce allows enterprises, organizations, and researchers to easily process very large volumes of data, the Cloud provides the required computing infrastructure to scale applications out to a large number of computing nodes. The beauty of these approaches are their ease-to-use and almost-free-admin cost properties. However, this simplicity comes at a price: the performance of MapReduce applications in the Cloud often do not match the one of a well-configured parallel DBMS. In this talk, we present some of the main features that allow DBMS to achieve orders of magnitude better performance than MapReduce applications. Then, we analyze how our Hadoop++ project allows MapReduce applications to match DBMS performance in the Cloud. We also discussed the design choices we made in the Hadoop++ project in order to preserve the ease-of-use and the almost-free-admin cost of MapReduce applications in the Cloud. Finally, we conclude this talk by discussing some of the challenges imposed by the Cloud to achieve data management efficiently.

14.00, Room 455 at PCRI

Abstract
The rapid increasing amount of data on the Web provides a huge source of information but makes efficient search more challenging. Distributing the search is appealing to improve both efficiency and scalability. In this talk, we first present, in the context of social tagging systems, two gossip-based approaches that personalize query processing in a peer-to-peer manner. The off-line approach relies on user’s past behavior to personalize the search, and the on-line approach relies on user’s past behavior and current query to further improve the result quality for queries depicting user’s emerging interests. We then present, for a multi-site search engine, two approaches that invalidate entries in result cache to guarantee the freshness of results served to users. The on-line invalidation approach invalidates an entry upon cache hit according to index changes in local search site. The threshold-based approach makes invalidations when index changes in remote search sites. Joint use of both approaches provides a promising solution to react to index updating that may arise in future multi-site search engines.