The incorporation of domain knowledge (DK) in AI has been studied for years and turns out to be critical in practice. Functions are also a basic notion for dealing with data science projects and are somehow related to DK. Consider the following scenario. Let \(D(y, z_1, \ldots , z_n)\) be a dataset, Alice a data scientist, Bob a domain expert and \(y = f(z_1, \ldots , z_n)\) a function known to Bob from his background knowledge. Alice is interested in the following simple yet crucial questions: How to define the satisfaction of f in D ? How to measure that satisfaction efficiently? How does this satisfaction relate to the supervised learning task of learning f from D ? It turns out that these problems are related to the study of counterexamples through the use of functional dependencies (FDs) and, in particular, FD measures used to quantify their satisfaction in a dataset such as the \(g_3\) indicator where the equality is replaced by more flexible predicates. In this article, we first examine the complexity of computing \(g_3\) . It is known that \(g_3\) can be computed in polynomial time when using equality, while it becomes NP -hard when using general predicates. Our goal is to refine this dichotomy by studying the impact of the following common properties: reflexivity, transitivity, symmetry, and antisymmetry. We show that symmetry and transitivity together are sufficient to guarantee that the \(g_3\) can be computed in polynomial time. However, removing one of them makes the problem NP -hard. Second, we study the computation of \(g_3\) in the polynomial and NP -hard cases identified previously. We propose different exact and approximate solutions for the computation of \(g_3\) in both cases. We compare these solutions in a detailed experimental study of time performance and approximation accuracy. All the algorithms are also made available via fastg3 , an open-source Python library with an underlying C++ implementation. Finally, we link counterexamples and \(g_3\) to supervised learning with a web application called adesit . adesit is intended to be part of an iterative data refinement process right after data selection and just before the machine learning process itself. It provides a way to evaluate the ability of a dataset to perform well for a given supervised learning problem through statistical and visual exploration. In a last section, we validate our approach by applying it to the industrial problem of air gap monitoring in compact hydro-generators.

Surrogate keys are now extensively utilized by database designers to implement keys in SQL tables. They are straightforward, easy to understand, enable efficient access, and are often considered a sufficient guarantee of data integrity despite lacking any real-world semantic meaning. In spite of all their benefits, one might wonder whether surrogate keys can negatively impact data quality. IT developers who rely exclusively on surrogate keys when designing database schemas may be tempted to not encode natural keys, as they are perceived as complex to manage at the application level. In such settings, surrogate keys allow the presence of so-called artificial unicity , a complex form of redundancy that can be propagated through foreign keys, and other underlying data-quality issues. In the presence of artificial unicity, most data cleaning techniques, especially unsupervised, are likely to fail, making data preparation and analytics very challenging. For relational databases implemented with surrogate keys but no natural keys, we developed RED2Hunt (RElational Databases REDundancy Hunting), a human-in-the-loop framework for identifying hidden redundancy and, if problems occur, clean the database. The framework was implemented on top of PostgreSQL within an eponym web-based platform to guide the expert through its application. In this paper, we present a demonstration of the RED2Hunt tool through three interactive scenarios on a polluted instance of the publicly available Perfect Pet database. During the demonstration, the visitor can take on one of two roles in the Perfect Pet database: a domain expert or a data scientist. As a domain expert, she will interact with RED2Hunt, for example to elicit natural keys, from simple yet very intuitive visualizations of tables' attributes. As a data scientist, she will explore two simple scenarios—executing SQL queries or applying learning models—on both the initial and cleaned databases to grasp the benefits of the approach.

In the context of process monitoring and predictive maintenance, an adapted visualisation of sensor data is essential in order to help the domain experts to make the right maintenance decision. The large volume and diversity of data leads us to aggregate the data to obtain semantically rich information useful to the domain expert. We study the case of industrial machinery equipped with several sensors producing time series, and we consider that this machinery has different operating states in its operation. We propose a method to identify an optimal representation of the data in 2 dimensions, understandable by the domain expert. This representation allows to easily identify the operating modes of the equipment and the possible deviation from a "normal" behavior. We use co-occurrence matrices to synthesise the time series data, and the features of interest and discretization are selected using two proposed criteria to measure the separation of working modes.

Taking domain knowledge into account is a long-standing issue in AI, especially nowadays where huge amounts of data are collected in the hope of delivering new in-sights and value. Let us consider the following scenario. Let D(y, x1, … ,xn) be a dataset, Alice a data scientist, Bob a domain expert and$y$=$f$(x1, … , xn) a function known by Bob from his background knowledge. We are interested in the following simple yet crucial questions for Alice: how to define the satisfaction of f in D and how difficult is it to measure that satisfaction? It turns out that those problems are related to functional dependencies (FDs) and especially FD measurements used to quantify their satisfaction in a dataset such as the g3 indicator. In this paper, we examine the computation of g3 with crisp FDs (aka. exact FDs) and a large class of non-crisp FDs replacing strict equality by more flexible predicates. Interestingly, it is known that the computation of g3 with crisp FDs is polynomial but turns out to be NP-Hard for non-crisp FDs. In this paper, we propose different exact and approximate solutions for the computation of g3 for both types. First, for crisp FDs with very large datasets, we propose solutions based on uniform and stratified random sampling. Second, for non-crisp FDs we present a detailed computation pipeline with various computation optimizations, including approximation algorithms and adaptations of recent developments in sublinear algorithms for NP-Hard problems. We also propose an in-depth experimental study of the algorithms presented in terms of time performances and approximation accuracy. All the algorithms are also made available through FASTG3, an open-source Python library designed to be intuitive and efficient thanks to an underlying C++ implementation.

Thanks to the numerous machine learning tools available to us nowadays, it is easier than ever to derive a model from a dataset in the frame of a supervised learning problem. However, when this model behaves poorly compared with an expected performance, the underlying question of the existence of such a model is often underlooked and one might just be tempted to try different parameters or just choose another model architecture. This is why the quality of the learning examples should be considered as early as possible as it acts as a go/no go signal for the following potentially costly learning process. With ADESIT, we provide a way to evaluate the ability of a dataset to perform well for a given supervised learning problem through statistics and visual exploration. Notably, we base our work on recent studies proposing the use of functional dependencies and specifically counterexample analysis to provide dataset cleanliness statistics but also a theoretical upper bound on the prediction accuracy directly linked to the problem settings (measurement uncertainty, expected generalization...). In brief, ADESIT is intended to be part of an iterative data refinement process right after data selection and right before the machine learning process itself. With further analysis for a given problem, the user can characterize, clean and export dynamically selected subsets, allowing to better understand what regions of the data could be refined and where the data precision must be improved by using, for example, new or more precise sensors.

This paper presents a virtual sensor for helicopter Main Gear Box (MGB) oil pressure. It is developed for anticipating failures of systems related to MGB lubrication. The virtual sensor is built using Machine Learning and massive in-service flight data collected from Airbus helicopters flying world-wide. The correlation between oil pressure values and other flight parameters is learnt during stable phases of flights in which the system is in its nominal state. At each flight, the values continuously estimated by the virtual sensor are compared to the measured ones, and an alert is raised when the difference becomes higher than a statistically predefined threshold. The virtual sensor was tested using normal and abnormal flights, and the results obtained so far in terms of anomaly detection performance are promising.

To help databases users who have just started learning SQL or are not familiar with their database, we propose ExplIQuE, an exploration interface with query extensions. Its purpose is to assist users to smoothly dive into data exploration, and to be able to express imprecise questions over their data. Indeed, such situations are more and more current with the increasing desire for users to get value out of their data. In this configuration, in addition to classic SQL querying possibilities, ExplIQuE offers the possibility to extend a given SQL query, by suggesting a set of possible selection predicates to add to the query, that aim at dividing the initial answer set to identify interesting exploration zones. In addition, ExplIQuE proposes some indicators to help the user in choosing its desire extension and in understanding her data, as well as interactive visualizations of the result set, in two dimensions revealed by PCA techniques. In this demonstration, we offer the audience the possibility to try the various functionalities of ExplIQuE by trying to express an imprecise question over a scientific database on bacterial colonies, through an iterative process. A video of the proposed demonstration is available at \urlhttps://youtu.be/oK8xWGCWj_A.

In addition to sensor heterogeneity, monitoring applications must handle different temporal data models (e.g time series, event sequences). In this paper, we address the problem of discovering directly actionable high level knowledge from such data. We model temporal information through interval-based streams describing environment states. We propose an approach to discover efficiently Complex Temporal Dependencies (CTD) between state streams, called CTD-Miner. A CTD is modeled similarly to a conjunctive normal form and describes temporal relations (time delays) between states. CTD-Miner is robust to temporal variability of data and uses a statistical independence test to determine the most appropriate time lags between states. This test is also used to perform pruning on sub-dependencies checking. Finally, we validate our approach via synthetic data and a case study in a real-world smart environment using outdoor cameras and real-time video processing.

International audience

Nowadays, every company could understand how its business evolves from the data (deluge) generated by its activities. Roughly speaking, two types of data co-exist: historical data and real-time data from which business analysts have to take their decisions in a timely fashion. In this context, the notions of time (application time and transaction time) and traceability turn out to play a crucial role to understand what happened in the company and what is currently happening. Tornado offers a full-fledged platform to deal with such data and is based on two key features: 1) a bi-temporal DB specifically designed for handling historical and real-time data, 2) a GUI that aims to facilitate query formulation for business analysts. In this demonstration, we provide the key resources to let the visitors play with the Tornado functionalities to interact with predefined data.

Temporal dependencies between multiple sensor data sources link two types of events if the occurrence of one is repeatedly followed by the appearance of the other in a certain time interval. TEDDY algorithm aims at discovering such dependencies, identifying the statically significant time intervals with a chi2 test. We present how these dependencies can be used within the GrizzLY project to tackle an environmental and technical issue: the deicing of the roads. This project aims to wisely organize the deicing operations of an urban area, based on several sensor network measures of local atmospheric phenomena. A spatial and temporal dependency-based model is built from these data to predict freezing alerts.

We demonstrate CarbonDB, a web application for Life Cycle Inventory data management. Life Cycle Assessment provides a well-accepted methodology for modelling environmental impacts of human activities. This methodology relies on the decomposition of a studied system into interdependent processes in a phase called Life Cycle Inventory. Several organisations provide processes databases containing thousands of processes with their interdependency links. The usual workflow to manage those databases is based on the manipulation of individual processes, which turns out to be a very harnessing work even if there are strong semantic similarities between the involved processes. In previous publications, we proposed a new workflow for LCA inventory databases maintenance based on the addition of semantic information to the processes they contained. This method considerably eases the modeling process and offers a synthetic view of the dependencies links. We created a web application based on this approach composed of a back-end for data management and a front-end for searching processes and visualize the dependencies links in a graph.

One of the leading challenges for pervasive computing is to ease the application development to smoothly handle the surrounding environment. We consider the case where the environment produces heterogeneous and continuous data, e. g. temperature readings, car positions... We have defined a scenario for containers transportation tracking in a medical context involving the transportation of fragile biological matter in sensor-enhanced containers. This scenario has been simulated as a testbed and offers a very nice setting to measure the agility of data-centric application development.

Stream processing systems compute continuous queries over increasingly large volumes of data, as monitoring applications emerge in a broad array of fields. These systems need to satisfy application-dependent constraints, one of the most important ones being accuracy demands and query response times. As system resources are limited, various query optimization techniques are proposed. To the best of our knowledge, none of the existing methods takes into account the size of the window, which is input to a query. We believe resource usage can be tackled with a novel approach, that attempts to compute an optimal window size for a given continuous query, thereby placing a minimal upper bound on the resource consumption for that query.

In a Mobile Ad-hoc NETwork (MANET), information sharing is usually performed by distributing advertisements and queries. In order not to load environments with unnecessary traffic, an advertisement policy should be designed according to users' information consumptions and provisions, which are limited by users' stay-times and their contexts. In this paper, we propose an algorithm to parameterize advertisement policies according to users' stay-times and contexts.

The progress in mobile devices and wireless communication technologies magnifies the importance of opportunistic networks like Mobile Ad-hoc NETworks (MANETs). In MANETs, information sharing is performed by distributing advertisements and queries. In order not to load the environment with unnecessary traffic, file advertisement and query resolution should be performed according to the interests of users. In this paper, we propose algorithms to identify and estimate users' interests. Experimentations are conducted to evaluate the performanace of the proposed algorithms on a mobile phone and a PC.

Decision tree methods generally suppose that the number of categories of the attribute to be predicted is fixed. Breiman et al., with their Twoing criterion in CART, considered gathering the categories of the predicted attribute into two supermodalities. In this article, we propose an extension of this method. We try to merge the categories in an optimal unspecified number of supermodalities. Our method, called Arbogodaï, allows during tree growing for grouping categories of the target variable as well as categories of the predictive attributes. It handles both categorical and quantitative attributes. At the end, the user can choose to generate either a set of single rules or a set of multiconclusion rules that provide interval-like predictions. © 2005 Wiley Periodicals, Inc. Int J Int Syst 20: 693–718, 2005.