Section: Research Program

Research Directions

We now detail three main research axes within the research agenda of Valda. For each axis, we first mention the leading researcher, and other permanent members involved.

Foundations of data management (Luc Segoufin; Serge Abiteboul, Pierre Senellart).

Foundations of data management

The systems we are interested in, i.e., for manipulating heterogeneous and confidential data, rapidly changing and massively distributed, are inherently error-prone. The need for formal methods to verify data management systems is best illustrated by the long list of famous leakages of sensitive or personal data that made the front pages of newspapers recently. Moreover, because of the cost in accessing intensional data, it is important to optimize the resources needed for manipulating them.

This creates a need for solid and high-level foundations of DBMS in a manner that is easier to understand, while also facilitating optimization and verification of its critical properties.

In particular these foundations are necessary for various design and reasoning tasks. It allows for clean specifications of key properties of the system such as confidentiality, access control, robustness etc. Once clean specifications are available, it opens the door for formal and runtime verification of the specification. It also permits the design of appropriate query languages – with good expressive power, with limited usage of resources –, the design of good indexes – for optimized evaluation –, and so on. Note that access control policies currently used in database management systems are relatively crude – for example, PostgreSQL offers access control rules on tables, views, or tuples (row security policies), but provides no guarantee that these access methods do not contradict each other, or that a user may have access through a query to information that she is not supposed to have access to.

Valda involves leading researchers in the formal verification of data flow in a system manipulating data. Other notable teams involve the WAVE project (http://db.ucsd.edu/WAVE/default.html) at U. C. San Diego, and the Business Artifact (http://researcher.watson.ibm.com/researcher/view_group.php?id=2501) research program of IBM. One of Valda's objectives is to continue this line of research.

In the short run, we plan to contribute to the state of the art of foundations of systems manipulating data by identifying new scenarios, i.e., specification formalisms, query languages, index structures, query evaluation plans, etc., that allow for any of the tasks mentioned above: formal or runtime verification, optimization etc. Several such scenarios are already known and Valda researchers contributed significantly to their discovery [46], [62],[6], but this research is still in infancy and there is a clear need for more functionalities and more efficiency. This research direction has many facets.

One of the facet is to develop new logical frameworks and new automaton models, with good algorithmic properties (for instance efficient emptiness test, efficient inclusion test and so on), in order to develop a toolbox for reasoning task around systems manipulating data. This toolbox can then be used for higher level tasks such as optimization, verification [46], or query rewriting using views [6].

Another facet is to develop new index structures and new algorithms for efficient query evaluation. For example the enumeration of the output of a query requires the construction of index structures allowing for efficient compressed representation of the output with efficient streaming decompression algorithms as we aim for a constant delay between any two consecutive outputs [69]. We have contributed a lot to this fields by providing several such indexes [62] but there remains a lot to be investigated.

Our medium-term goal is to investigate the borders of feasibility of all the reasoning tasks above. For instance what are the assumptions on data that allow for computable verification problems? When is it not possible at all? When can we hope for efficient query answering, when is it hopeless? This is a problem of theoretical nature which is necessary for understanding the limit of the methods and driving research towards the scenarios where positive results may be obtainable.

A typical result would be to show that constant delay enumeration of queries is not possible unless the database verify property A and the query property B. Another typical result would be to show that having a robust access control policy verifying at the same time this and that property is not achievable.

Very few such results exist nowadays. If many problems are shown undecidable or decidable, charting the frontier of tractability (say linear time) remains a challenge.

Only when we will have understood the limitation of the method (medium-term goal) and have many examples where this is possible, we can hope to design a solid foundation that allowing for a good trade-off between what can be done (needs from the users) and what can be achieved (limitation from the system). This will be our long-term goal.

Uncertainty and provenance of data (Pierre Senellart; Luc Segoufin).

Uncertainty and provenance of data

This research axis deals with the modeling and efficient management of data that come with some uncertainty (probabilistic distributions, logical incompleteness, missing values, open-world assumption, etc.) and with provenance information (indicating where the data originates from), as well as with the extraction of uncertainty and provenance annotations from real-world data. Interestingly, the foundations and tools for uncertainty management often rely on provenance annotations. For example, a typical way to compute the probability of query results in probabilistic databases is first to generate the provenance of these query results (in some Boolean framework, e.g., that of Boolean functions or of provenance semirings), and then to compute the probability of the resulting provenance annotation. For this reason, we will deal with uncertainty and provenance in a unified manner.

Valda researchers have carried out seminal work on probabilistic databases [63], [36][12], provenance management [4], incomplete information [37], and uncertainty analysis and propagation in conflicting datasets [53], [34]. These research areas have reached a point where the foundations are well-understood, and where it becomes critical, while continuing developing the theory of uncertain and provenance data management, to move to concrete implementations and applications to real-world use cases.

In the short term, we will focus on implementing techniques from the database theory literature on provenance and uncertainty data management, in the direction of building a full-featured database management add-on that transparently manages provenance and probability annotations for a large class of querying tasks. This work has started recently with the creation of the ProvSQL extension to PostgreSQL, discussed in more details in the following section. To support this development work, we need to resolve the following research question: what representation systems and algorithms to use to support both semiring provenance frameworks [57], extensions to queries with negation [54], aggregation [39], or recursion [67]?

Next, we will study how to add support for incompleteness, probabilities, and provenance annotations in the scenarios identified in the first axis, and how to extract and derive such annotations from real-world datasets and tasks. We will also work on the efficiency of our uncertain data management system, and compare it to other uncertainty management solutions, in the perspective of making it a fully usable system, with little overhead compared to a classical database management system. This requires a careful choice of the provenance representation system used, which should be both compact and amenable to probability computations. We will study practical applications of uncertainty management. As an example, we intend to consider routing in public transport networks, given a probabilistic model on the reliability and schedule uncertainty of different transit routes. The system should be able to provide a user with itinerary to get to have a (probabilistic) guarantee to be at its destination within a given time frame, which may not be the shortest route in the classical sense.

One overall long-term goal is to reach a full understanding of the interactions between query evaluation or other broader data management tasks and uncertain and annotated data models. We would in particular want to go towards a full classification of tractable (typically polynomial-time) and intractable (typically NP-hard for decision problems, or #P-hard for probability evaluation) tasks, extending and connecting the query-based dichotomy [49] on probabilistic query evaluation with the instance-based one of [4] [38].

Another long-term goal is to consider more dynamic scenarios than what has been considered so far in the uncertain data management literature: when following a workflow, or when interacting with intensional data sources, how to properly represent and update uncertainty annotations that are associated with data. This is critical for many complex data management scenarios where one has to maintain a probabilistic current knowledge of the world, while obtaining new knowledge by posing queries and accessing data sources. Such intensional tasks requires minimizing jointly data uncertainty and cost to data access.

Personal information management (Serge Abiteboul; Pierre Senellart).

Personal information management

This is a more applied direction of research that will be the context to study issues of interest (see discussion in application domains further).

A typical person today usually has data on several devices and in a number of commercial systems that function as data traps where it is easy to check in information and difficult to remove it or sometimes to simply access it. It is also difficult, sometimes impossible, to control data access by other parties. This situation is unsatisfactory because it requires users to trade privacy against convenience but also, because it limits the value we, as individuals and as a society, can derive from the data. This leads to the concept of Personal Information Management System, in short, a Pims.

A Pims runs, on a user's server, the services selected by the user, storing and processing the user's data. The Pims centralizes the user's personal information. It is a digital home. The Pims is also able to exert control over information that resides in external services (for example, Facebook), and that only gets replicated inside the Pims. See, for instance, [1] for a discussion on the advantages of Pims, as well as issues they raise, e.g. security issues. It is argued there that the main reason for a user to move to Pims is these systems enable great new functionalities.

Valda will study in particular the integration of the user's data. Researchers in the team have already provided important contributions in the context of data integration, notably in the context of the Webdam ERC (2009–2013).

Based on such an integration, Pims can provide a functions, that goes beyond simple query answering:

• Global search over the person's data with a semantic layer using a personal ontology (for example, the data organization the person likes and the person's terminology for data) that helps give meaning to the data;

• Automatic synchronization of data on different devices/systems, and global task sequencing to facilitate interoperating different devices/services;

• Exchange of information and knowledge between "friends" in a truly social way, even if these use different social network platforms, or no platform at all;

• Centralized control point for connected objects, a hub for the Internet of Things; and

• Data analysis/mining over the person's information.

The focus on personal data and these various aspects raise interesting technical challenges that we intend to address.

In the short term, we intend to continue work on the ThymeFlow system to turn it into an easily extendable and deployable platform for the management of personal information – we will in particular encourage students from the M2 Web Data Management class taught by Serge and Pierre in the MPRI programme to use this platform in their course projects. The goal is to make it easy to add new functionalities (such as new source synchronizers to retrieve data and propagate updates to original data sources, and enrichers to add value to existing data) to considerably broaden the scope of the platform and consequently expand its value.

In the medium term, we will continue the work already started that focuses in turning information into knowledge and in knowledge integration. Issues related to intensionality or uncertainty will in particular be considered, relying on the works produced in the other two research axes. We stress, in particular, the importance of minimizing the cost to data access (or, in specific scenarios, the privacy cost associated with obtaining data items) in the context of personal information management: legacy data is often only available through costly APIs, interaction between several Pims may require sharing information within a strict privacy budget, etc. For these reasons, intensionality of data will be a strong focus of the research.

In the long term, we intend to use the knowledge acquired and machine learning techniques to predict the user's behavior and desires, and support new digital assistant functions, providing real value from data. We will also look into possibilities for deploying the ThymeFlow platform at a large scale, perhaps in collaboration with industry partners.