Frédéric Pennerath

The ATLAS DAQ and event filter prototype “−1” project

Abstract A project has been approved by the ATLAS Collaboration for the design and implementation of a Data Acquisition and Event Filter prototype, based on the functional architecture described in the ATLAS Technical Proposal. The prototype consists of a full “vertical” slice of the ATLAS Data Acquisition and Event Filter architecture, including all the hardware and software elements of the data flow, its control and monitoring as well as all the elements of a complete on-line system. This paper outlines the project, its goals, structure, schedule and current status and describes details of the system architecture and its components.

The Model of Most Informative Patterns and Its Application to Knowledge Extraction from Graph Databases

This article introduces the class of Most Informative Patterns (MIPs) for characterizing a given dataset. MIPs form a reduced subset of non redundant closed patterns that are extracted from data thanks to a scoring function depending on domain knowledge. Accordingly, MIPs are designed for providing experts good insights on the content of datasets during data analysis. The article presents the model of MIPs and their formal properties wrt other kinds of patterns. Then, two algorithms for extracting MIPs are detailed: the first directly searches for MIPs in a dataset while the second screens MIPs from frequent patterns. The efficiencies of both algorithms are compared when applied to reference datasets. Finally the application of MIPs to labelled graphs, here molecular graphs, is discussed.

Graph-Mining Algorithm for the Evaluation of Bond Formability

The formability of a bond in a target molecule is a bond property related to the problem of finding a reaction that synthesizes the target by forming the bond: the easier this problem, the higher the formability. Bond formability provides an interesting piece of information that might be used for selecting strategic bonds during a retrosynthesic analysis or for assessing synthetic accessibility in virtual screening. The article describes a graph-mining algorithm called GemsBond that evaluates formability of bonds by mining structural environments contained in several thousand molecular graphs of reaction products. When tested on reaction databases, GemsBond recognizes most formed bonds in reaction products and provides explanations consistent with knowledge in organic synthesis.

Mining Intervals of Graphs to Extract Characteristic Reaction Patterns

The article introduces an original problem of knowledge discovery from chemical reaction databases that consists in identifying the subset of atoms and bonds that play an effective role in a given chemical reaction. The extraction of the resulting characteristic reaction patternis then reduced to a graph-mining problem: given lower and upper bound graphs g l and g u , the search of best patterns in an interval of graphsconsists in finding among connected graphs isomorphic to a subgraph of g u and containing a subgraph isomorphic to g l , best patterns that maximize a scoring function and whose score depends on the frequency of the pattern in a set of examples. A method called CrackReac is then proposed to extract best patterns from intervals of graphs. Accuracy and scalability of the method are then evaluated by testing the method on the extraction of characteristic patterns from reaction databases.

Fast extraction of locally optimal patterns based on consistent pattern function variations

This article introduces the problem of searching locally optimal patterns within a set of patterns constrained by some anti-monotonic predicate: given some pattern scoring function, a locally optimal pattern has a maximal (or minimal) score locally among neighboring patterns. Some instances of this problem have produced patterns of interest in the framework of knowledge discovery since locally optimal patterns extracted from datasets are very few, informative and nonredundant compared to other pattern families derived from frequent patterns. This article then introduces the concept of variation consistency to characterize pattern functions and uses this notion to propose GALLOP, an algorithm that outperforms existing algorithms to extract locally optimal itemsets. Finally it shows how GALLOP can generically be applied to two classes of scoring functions useful in binary classification or clustering pattern mining problems.