Les informations qui suivent étant extraites du rapport d’activité annuel de l’équipe, elles sont en anglais.
Valda’s focus is on both
foundational and systems aspects of complex data management, especially human-centric data. The data we are interested in is typically heterogeneous, massively distributed, rapidly evolving, intensional, and often subjective, possibly erroneous, imprecise, incomplete. In this setting, Valda is in particular concerned with the optimization of complex resources such as computer time and space, communication, monetary, and privacy budgets. The goal is to extract value from data, beyond simple query answering.
Data management 40, 49 is now an old, well-established field, for which many scientific results and techniques have been accumulated since the sixties. Originally, most works dealt with static, homogeneous, and precise data. Later, works were devoted to heterogeneous data 3841, and possibly distributed 72 but at a small scale.
However, these classical techniques are poorly adapted to handle the new challenges of data management. Consider human-centric data, which is either produced by humans, e.g., emails, chats, recommendations, or produced by systems when dealing with humans, e.g., geolocation, business transactions, results of data analysis. When dealing with such data, and to accomplish any task to extract value from such data, we rapidly encounter the following facets:
Heterogeneity: data may come in many different structures such as unstructured text, graphs, data streams, complex aggregates, etc., using many different schemas or ontologies.
Massive distribution: data may come from a large number of autonomous sources distributed over the web, with complex access patterns.
Rapid evolution: many sources may be producing data in real time, even if little of it is perhaps relevant to the specific application. Typically, recent data is of particular interest and changes have to be monitored.
Intensionality1: in a classical database, all the data is available. In modern applications, the data is more and more available only intensionally, possibly at some cost, with the difficulty to discover which source can contribute towards a particular goal, and this with some uncertainty.
Confidentiality and security: some personal data is critical and need to remain confidential. Applications manipulating personal data must take this into account and must be secure against linking.
Uncertainty: modern data, and in particular human-centric data, typically includes errors, contradictions, imprecision, incompleteness, which complicates reasoning. Furthermore, the subjective nature of the data, with opinions, sentiments, or biases, also makes reasoning harder since one has, for instance, to consider different agents with distinct, possibly contradicting knowledge.
Among all these aspects, intensionality is perhaps the one that has least been studied, so we pay particular attention to it. Consider a user’s query, taken in a very broad sense: it may be a classical database query, some information retrieval search, a clustering or classification task, or some more advanced knowledge extraction request. Because of intensionality of data, solving such a query is a typically dynamic task: each time new data is obtained, the partial knowledge a system has of the world is revised, and query plans need to be updated, as in adaptive query processing 55 or aggregated search 80. The system then needs to decide, based on this partial knowledge, of the best next access to perform. This is reminiscent of the central problem of reinforcement learning 78 (train an agent to accomplish a task in a partially known world based on rewards obtained) and of active learning 74 (decide which action to perform next in order to optimize a learning strategy) and we intend to explore this connection further.
Uncertainty of the data interacts with its intensionality: efforts are required to obtain more precise, more complete, sounder results, which yields a trade-off between
processing cost and data quality.
Other aspects, such as heterogeneity and massive distribution, are of major importance as well. A standard data management task, such as query answering, information retrieval, or clustering, may become much more challenging when taking into account the fact that data is not available in a central location, or in a common format. We aim to take these aspects into account, to be able to apply our research to real-world applications.
We intend to tackle hard technical issues such as query answering, data integration, data monitoring, verification of data-centric systems, truth finding, knowledge extraction, data analytics, that take a different flavor in this modern context. In particular, we are interested in designing strategies to
minimize data access cost towards a specific goal, possibly a massive data analysis task. That cost may be in terms of communication (accessing data in distributed systems, on the Web), of computational resources (when data is produced by complex tools such as information extraction, machine learning systems, or complex query processing), of monetary budget (paid-for application programming interfaces, crowdsourcing platforms), or of a privacy budget (as in the standard framework of differential privacy).
A number of data management tasks in Valda are inherently intractable. In addition to properly characterizing this intractability in terms of complexity theory, we intend to develop solutions for solving these tasks in practice, based on approximation strategies, randomized algorithms, enumeration algorithms with constant delay, or identification of restricted forms of data instances lowering the complexity of the task.