Fig. 3. Average, minimum, and maximum execution times for the queries belonging to different **query** templates in the test dataset.
C. Prediction Models
To predict **query** execution time, we experiment with two regression models. We first experiment with Weka’s [17] implementation of k-nearest neighbors (k-NN) regression [18], [19]. The k-NN **algorithm** predicts **based** on the closest training data points. It uses a distance function to compute these closest data points. We use Euclidean distance as the distance function in our experiments. For predictions, we use the weighted average of the k nearest neighbors - weighted by the inverse of the distance from the querying data point. This ensures that the nearby neighbors contribute more to the prediction than the faraway neighbors. We use the k-dimensional tree (k-d tree) [20] data structure to compute the nearest neighbors. For N training samples, k-d tree can find the nearest neighbor of a data point with O (log N ) operations. We also experiment with the libsvm [21] implementation of Support Vector Machine (SVM) with the nu-SVR kernel for regression [22]. The approach in SVM regression is to map the features to a higher dimensional space and perform a regression in that space. The predictions in SVM are **based** on a subset of data points known as support vectors.

En savoir plus
database while SVM would have difficulties for managing databases with more than 10000 examples (and is intractable for more than 100000 examples).
6 Conclusion and Future Works
In this paper we have proposed a probabilistic classification method **based** on the maximization of a loss probabilistic function that takes into account the whole probability distribution and not only the probability of the class. We propose two algorithms that simultaneously learn a set of kernel functions that encodes a probability distribution over classes without any post-normalization process. Last, we show that the computation time of the approach is very little sensitive to the size of the dataset. Our method competes with the other kernel approaches on the used benchmark datasets. Experiments on the KDD Cup 2012 dataset confirm that the approach is efficient on very large datasets when kernel methods are not tractable. Moreover, the parameters of the **algorithm** can be tuned automatically as it has been done the whole experimentation.

En savoir plus
In information extraction, a named entity is a real object such as person, location, organization, etc that can be denoted with a proper name. Example of named entities could be authors in bibli- ographic databases. With vast and enormous amount of data being represented in such databases, which can be because of typographical errors and imperfect information, it is difficult to recog- nize named entities in a manner that each unique item refers to its citations only. This raises the problem of named entity disambiguation. Author name disambiguation is a type of named entity disambiguation that applies to scholarly articles and the aim is to find all the publication records of the same author and distinguish them from the others. In other words, it can be defined as par- titioning of a collection of data records into clusters where all records belong to a unique entity. Typical approaches of author name disambiguation rely on information about the authors such as email addresses, their affiliations, co-authors and citation graphs that helps identifying them among the all. Accordingly, this information can be used to implement a machine **learning** classifier that decides whether two entities refer to the same author or not.

En savoir plus
107 En savoir plus

Information Processing & Management pp 215-231, 39(2) 2003
Abstract
Recent studies suggest that significant improvement in information retrieval performance can be achieved by combining multiple representations of an information need. The paper presents a genetic approach that combines the results from multiple **query** evaluations. The genetic **algorithm** aims to optimise the overall relevance estimate by exploring different directions of the document space. We investigate ways to improve the effectiveness of the genetic exploration by combining appropriate techniques and heuristics known in genetic theory or in the IR field. Indeed, the approach uses a niching technique to solve the relevance multimodality problem, a relevance feedback technique to perform genetic transformations on **query** formulations and evolution heuristics in order to improve the convergence conditions of the genetic process.The effectiveness of the global approach is demonstrated by comparing the retrieval results obtained by both genetic multiple **query** evaluation and classical single **query** evaluation performed on a subset of TREC-4 using the Mercure IRS. Moreover, experimental results show the positive effect of the various techniques integrated to our genetic **algorithm** model.

En savoir plus
Although, the **learning** **algorithm** is Gold-style [15], i.e., it uses a given set of examples to infer the transducer, it could be used as core in an interactive learner in Angluin- style [1], similar to [5]. Further research is required to justify this claim. In the context of xml, ranked transducers seem particularly useful if input and output dtds are given. This allows to construct encodings into ranked trees which group similar items (with respect to the dtd) into an own subtree. A dtop can then delete, interchange, or copy the groups (here, “group” refers to a regular subexpression of the dtd). Are there other classes beyond dtop to which our results could be extended? An interesting and robust class are top- down tree transducers with regular look-ahead [9]. Such a transducer is allowed to first execute a bottom-up finite-state relabeling over the input tree, and then run the top-down translation on the relabeled tree. The XPath filters that are used in xslt programs naturally correspond to bottom-up look-ahead. How can we minimize dtop transformations with look-ahead (dtop R s)? Through changing the relabel-

En savoir plus
the study. As an entry gets replaced over time by new vassoc, the spatial similarity between the Vassoc and the original vref is more likely to decrease. For exa[r]

Search engine logs register the queries users run in the search engines to complete their search tasks. Mining those logs allows the identification of search tasks. Search session segmentation is the first step in multiple methods for search task identification[8, 10, 12]. In session segmentation, a sequential log of search queries is parti- tioned into smaller sequences of queries. The boundaries for the **query** log partitions lie in pairs of adjacent queries. To determine if a **query** pair is a boundary, one may use time spans between the queries [7, 10, 12, 18]. If the time span is larger than a certain thresh- old, the **query** pair is considered a session boundary, which means that each **query** belongs to a different session. More sophisticated approaches use heuristics-**based** models [5] or neural networks [3] to determine the boundaries in the sequential **query** log. However, the use of clicked URLs [5] or adjacent queries [3] in the search log - in both backward and forward directions - represents a limitation in practical setups, especially in user supporting applications that require modeling on the fly. In such applications, waiting for the clicked URL or future queries to populate the model input might be unfeasible.

En savoir plus
Abstract. The efficiency of a **query** execution plan depends on the accu- racy of the selectivity estimates given to the **query** optimiser by the cost model. The cost model makes simplifying assumptions in order to produce said estimates in a timely manner. These assumptions lead to selectivity estimation errors that have dramatic effects on the quality of the resulting **query** execution plans. A convenient assumption that is ubiquitous among current cost models is to assume that attributes are independent with each other. However, it ignores potential correla- tions which can have a huge negative impact on the accuracy of the cost model. In this paper we attempt to relax the attribute value indepen- dence assumption without unreasonably deteriorating the accuracy of the cost model. We propose a novel approach **based** on a particular type of Bayesian networks called Chow-Liu trees to approximate the distribu- tion of attribute values inside each relation of a database. Our results on the TPC-DS benchmark show that our method is an order of magnitude more precise than other approaches whilst remaining reasonably efficient in terms of time and space.

En savoir plus
5 Conclusion
In this paper, we have presented the corese implementation of the sparql **query** language and its pattern matching mechanism. We reformulated the prob- lem of answering sparql queries against rdf(s) data into a graph homomor- phism checking and the corese **algorithm** takes advantage of the structure of graphs translating rdf(s) and sparql data and constrains graph homomor- phism checking by sparql value constraints. Corese has proven its usability in a wide range of real world applications since 2000 [3]. Its implementation has widely evolved and it is now compliant with the core of sparql **query** language.

En savoir plus
In this section, we use INDIAN for a concrete DL problem. We train a DNN for image classification using the CIFAR-100 dataset [15]. Regarding the network, we use a slightly modified version (with P = 10 6 parameters to optimize) of Network in Network (NiN) [17]. We compare our **algorithm** to the classical stochastic gradient descent (SGD) **algorithm**, ADAGRAD [13] and ADAM [14]. The full methodology can be found in the supplementary materials. We present the results for three different choices of hyperparameters of INDIAN using the intuition given by (3), even though there are many other satisfying choices of parameters. INDIAN is available as an optimizer for Pytorch, Keras and Tensorflow: https://github.com/camcastera/Indian-for-DeepLearning/ [7]. Results. Our result are representative of what can be obtained with a moderately large network on CIFAR-100 with reasonable parameter tuning (though higher accuracy can be achieved by much larger networks). Fig. 2 (left) shows that ADAM and INDIAN (with adequate tuning) can outperform SGD and ADAGRAD for the training. Thus, INDIAN proves an efficient optimizer for this problem. Although ADAM’s and INDIAN’s performances are similar, the latter is more versatile since its hyperparameters are fully tunable compared to ADAM’s adaptive stepsizes. Fig. 2 (right) highlights a special aspect of INDIAN. Indeed, every versions of INDIAN present better testing performances compared to the ones trained with usual optimizers, besides, the evolution of the accuracy depends on the choice of α and β. This suggests that the tuning of (α, β) might be a new regularization strategy in addition to usual methods such as dropout [22] and weight decay [16].

En savoir plus
is different than those of the meta-parameters. A crucial point is to well adjust the two dynamics in order they work in concordance. Like the bandit methods used in previous works, the SEA uses an exploitation component (the global selection) and an exploration component (the state mutation). The way to manage the computational cost, i.e. the number of evaluations, between the operators (or evolutionary algorithms) is different. In the AOS methods, one operator (or optimization **algorithm**) is used at each generation, and the method controls the number of generation for each operator during the run. In the SEA, several evolutionary algorithms can be used at the same time, and the method controls the sub-population size of each **algorithm**. The SEA is more parallel and the AOS more sequential.

En savoir plus
Computation of the (sub)gradients and convergence proofs (in batch or mini-batch settings) typically rely on the sum-rule in smooth or convex settings, i.e., ∂( J 1 + J 2 ) = ∂ J 1 + ∂ J 2 . Unfortu-
nately this sum-rule does not hold in general in the nonconvex setting using the standard Clarke subdifferential. Yet, many DL studies ignore the failure of the sum rule: they use it in practice, but circumvent the theoretical problem by modeling their method through simple dynamics (e.g., smooth or convex). We tackle this difficulty as is, and show that such practice can create additional spurious stationary points that are not Clarke-critical. To address this question, we introduce the notion of D-criticality. It is less stringent than Clarke-criticality and it describes more accurately real-world implementation. We then show convergence of INDIAN to such D-critical points. Our theoretical results are general, simple and allow for aggressive step-sizes in o(1/ log k). We first provide adequate calculus rules and tame nonsmooth Sard’s-like results for the new steady states we introduced. We then combine these results with a Lyapunov analysis from Alvarez et al. (2002) and the differential inclusion approximation method (Benaïm et al., 2005) to characterize the asymptotics of our **algorithm** similarly to Davis et al. (2019), Adil (2018). This provides a strong theoretical ground to our study since we can prove that our method converges to a connected component of the set of steady states even for networks with ReLU or other nonsmooth activation functions. For the smooth deterministic dynamics, we also show that convergence in values is of the form O(1/t) where t is the running time. For doing so we provide a general result for the solutions of a family of differential inclusions having a certain type of favorable Lyapunov functions.

En savoir plus
In contrast, the user faces the so-called many answers (or information over- load) problem when she submits a ‘broad’ **query**, i.e. , she has a vague and poorly defined information need. Means to circumvent such problem include **query**-by-example [3] [4] and ranking [5] [6] **based** techniques. Such techniques first seek to approximate user’s requirement. For instance, systems proposed in [3] [4] use relevance feedback from the user (few records judged to be relevant by the user) while works in [5] [6] use past behavior of the user (derived from available workloads). Then, they compute a score (similarity measures) of ev- ery answer that represents the extent to which it is relevant to the estimated user’s requirement. Finally, the user is provided with a ranked list, in descend- ing order of relevance, of either all **query** results or only the top-k [7] subset. The effectiveness of the above approaches depends on their ability to accurately capture the user’s requirement, which is, unfortunately, a very difficult and time consuming task.

En savoir plus
Being model-free means that we do not assume that the decision maker’s reasoning follows some well known and explicitly described rules or logic system. The human judgment is probably too complex to be described by simple rules and may not be totally determinis- tic. In some contexts, in policy decision making for instance, such rules may be a good frame that helps the decision maker to build a coherent reasoning and some defendable argumentations about the accepted decisions. But in others it could be that none of the existing Multi Criteria Aggregation Procedure (MCAP) such as the ones cited on the previous paragraph can be rigorously defended. For instance when we classify instinctively companies into categories represent- ing how globally responsible they are, it is possible that our intuitive reasoning is not **based** on an additive utility or a majority or logi- cal rule but it is a mix of all of them with some additional noise. In our knowledge, there exist very few model-free methods for MCDA. ORCLASS ([Pinheiro et al., 2014]) is one of them. It is a model-free multi-criteria sorting method that includes an interaction procedure. This procedure is incremental and in each step, the aim is to find the best or the most appropriate object to add to the **learning** set in order to guarantee the monotonicity. Our method is inspired from ORCLASS but transforms it in a stochastic method where the inter- action is limited to one step where the decision maker provides the full **learning** set.

En savoir plus
132 En savoir plus

able for ALC extended with functional roles (ALCF). We adopt the standard notion of combined complexity, which is measured in terms of the size of the whole input (TBox, data vocabulary, and **query** or predicate symbol).
Because of the restricted data vocabulary ⌃ and the quantification over all ⌃-databases in their definition, **query** emptiness and predicate emptiness do not reduce to standard reasoning problems such as **query** evaluation and **query** containment. Formally, this is demonstrated by our undecid- ability result for ALCF, which should be contrasted with the decidability of **query** entailment and containment in this DL, cf. (Calvanese, De Giacomo, & Lenzerini, 1998). When emptiness is de- cidable, the complexity still often differs from that of **query** evaluation. To simplify the comparison, we display in Figure 1 known complexity results for **query** evaluation in the considered DLs; please consult (Baader, Brandt, & Lutz, 2005, 2008; Kr¨otzsch, Rudolph, & Hitzler, 2007; Eiter, Gottlob, Ortiz, & Simkus, 2008) for the results concerning EL and its Horn extensions, (Calvanese, De Gia- como, Lembo, Lenzerini, & Rosati, 2007; Artale, Calvanese, Kontchakov, & Zakharyaschev, 2009) for the results on DL-Lite, and (Tobies, 2001; Hustadt, Motik, & Sattler, 2004; Lutz, 2008; Ortiz, Simkus, & Eiter, 2008) for the results on DLs from the ALC family. By comparing the two sides of Figure 1, we observe that there is no clear relationship between the complexity of emptiness checking and the complexity of **query** evaluation. Indeed, while the problems are often of similar complexity, there are several cases in which emptiness checking is more difficult than the corre- sponding **query** evaluation problem. It can also be the other way around, and complexities can also be incomparable. Note that for the extension EL ? of EL with the bottom concept ? (used to ex-

En savoir plus
Users may not use the right concepts - from the viewpoint of the ontologist - when writing a **query**, and this mismatch may lead to missed answers. Some experiments of the Corese semantic search engine we have developed give us good examples of misunderstanding or misuse by the user of concepts stated by the ontologist: in the CoMMA project the Commerce concept has been used instead of the Business one, TechnicalReport instead of ResearchReport. Moreover, a user asking for a person working on a subject may appreciate, instead of a failure, the retrieval of a research group working on that subject, even if a research group is not exactly a person. Lastly, a user may search for some related resources without knowing how their possibly complex relation is stated in the annotations. For instance, a user may search for organizations related to human sciences while ignoring the diversity of relations used to express this relationship in the annotations. All these examples illustrate the prime interest of semantic approximations for efficiently searching the Semantic Web.

En savoir plus
In our approach, uncertainty is defined as the impact of undesirable events on project objectives (cost and duration). It must be considered while making decisions about the structure of the system and its asso- ciated project. The need to optimize each technical choice jointly with those related to project activities has been highlighted in previous works ( Pitiot et al., 2010 ) where a multi-criteria optimization method **based** on an evolutionary **algorithm** guided by knowledge has been proposed. The method principle was to optimize the selection of project scenarios, taking into account design choices and project activities associated with them. A scenario is a set of tasks, with precedence constraints, that must be planned. The aim was to obtain a set of Pareto-optimal scenarios in a two-dimensional objective space (the total cost and the overall duration of the project). However, uncertainty was not taken into account. Thus, in order to improve it, a third dimension can be integrated: the risk one. In previous works done in our research team ( Baroso et al., 2014 ), the integration of risk as a third objective to minimize was proposed using a multi-objective **algorithm** **based** on a standard ACO. The ACO meta-heuristic is selected for addressing the problem that this paper is dealing with because many works in the literature attest that the use of ACO is very promising in project management especially in providing near optimal solutions to handle issues that are too expensive computationally. In Fernandez et al. ( 2015 ), the authors have developed a hybrid approach **based** mainly on ACO meta-heuristic to handle many objectives in the case of portfolio problems. It provided high-quality portfolios compared with other powerful meta-heuristics that deal with Pareto-front solutions. Other works have proved the power of ACO algorithms for solving both deterministic and probabilistic networks such as CPM/PERT by providing good optimal and sub-optimal solutions ( Abdallah et al., 2009 ). In Chen and Zhang ( 2012 ), the authors used an Ant Colony System (ACS) approach and Monte Carlo simulation (MCS) to maximize under uncertainty the expected net present value (NPV) of cash flows in the case of scheduling multi-mode projects.

En savoir plus
Using supervised **learning** machine **algorithm** to identify future fallers **based** on gait patterns : a two-year longitudinal study
Sophie Gillain 1 , Mohamed Boutaayamou 2 , Cedric Schwartz 3 , Olivier Bruyère 4 , Olivier Brüls 3 ,
Jean-Louis Croisier 3,5 , Eric Salmon 6 , Jean-Yves Reginster 7 , Gaëtan Garraux 3,6 , Jean Petermans 1