Rémi Lehn

A User-Driven Process for Mining Association Rules

This paper describes the components of a human-centered process for discovering association rules where the user is considered as a heuristic which drives the mining algorithms via a well-adapted interface. In this approach, inspired by experimental works on behaviors during a discovery stage, the rule extraction is dynamic : at each step, the user can focus on a subset of potentially interesting items and launch an algorithm for extracting the relevant associated rules according to statistical measures. The discovered rules are represented by a graph updated at each step, and the mining algorithm is an adaptation of the well-known A Priori algorithm where rules are computed locally. Experimental results on a real corpus built from marketing data illustrate the different steps of this process.

Minimizing crossings in hierarchical digraphs with a hybridized genetic algorithm

Producing clear and intelligible layouts of hierarchical digraphs knows a renewed interest in information visualization. Recent experimental results show that metaheuristics are well-adapted methods for this problem. In this paper, we develop a new Hybridized Genetic Algorithm for arc crossing minimization. It follows the basic scheme of a GA with two major differences: problem-based crossovers adapted from ordering GAs are combined with a local search strategy based on averaging heuristics. Computational testing was performed on a set of 180 random hierarchical digraphs of standard sizes with various structures. Results show that the Hybridized Genetic Algorithm significantly outperforms Tabu Search—which is one of the best known methods for this problem- and also a multi-start descent except for highly connected graphs.

Elements for the description of fitness landscapes associated with local operators for layered drawings of directed graphs

Minimizing arc crossings for drawing acyclic digraphs is a well-known NP-complete problem for which several local-search approaches based on local transformations (switching, median, ...) have been proposed. Their adaptations have been recently included in different metaheuristics. As an attempt to better understand the dynamics of the search processes, we study the fitness landscapes associated with these transformations. We first resort to a set of multi-start descents to sample the search space for three hundred medium-sized graphs. Then, we investigate complete fitness landscapes for a set of 1875 smaller graphs, this aims at showing some instance characteristics that influence search strategies. The underlying idea is to consider a fitness landscape as a graph whose vertices are drawings and arcs representing a transformation of a drawing into another. We confirm that the properties of basins of attraction closely depend on the instances. Also, we show that the probability of being stuck on a local optimum is linked to the specific shapes of the basins of attraction of global optima which may be very different from the regular image of the continuous case generally used as a reference.

A genetic algorithm for visualizing networks of association rules

In order to discover relevant information in a huge amount of data, a process commonly used in data mining consists in extracting logical association rules. As algorithms generally produce a large quantity of rules which hides the most interesting, it is essential to develop well-adapted rule mining tools which organize the rules and offer an intelligible representation of them.

An Ontology-Based Autonomic System for Improving Data Warehouse Performances

With the increase in the amount and complexity of information, data warehouse performance has become a constant issue, especially for decision support systems. As decisional experts are faced with the management of more complex data warehouses, a need for autonomic management capabilities is shown to help them in their work. Implementing autonomic managers over knowledge bases to manage them is a solution that we find more and more used in business intelligence environments. What we propose, as decisional system experts, is an autonomic system for analyzing and improving data warehouse cache memory allocations in a client environment. The system formalizes aspects of the knowledge involved in the process of decision making (from system hardware specifications to practices describing cache allocation) into the same knowledge base in the form of ontologies, analyzes the current performance level (such as query average response time values) and proposes new cache allocation values so that better performance is obtained.

Business Intelligence – CDMB – Implementing BI-CMDB to Lower Operation Cost Expenses and Satisfy Increasing User Expectations

Since the beginning of Decision Support Systems (DSS), benchmarks published by hardware and software vendors show that DSS technologies are more efficient by offering better performance for a stable cost and are aiming at lowering operations costs with resources mutualisation functionalities. Nevertheless, as a paradox, whereas data quality and company politics improves significantly, user dissatisfaction with poor performances increases constantly (Pendse, 2007). In reality, this surprising result points the difficulties in globally defining the efficiency of a DSS. As a matter of fact, we see that whereas a “performance/cost” ratio is good for a technician, from the user perspective the “real performance/ expected performance” ratio is bad. Therefore, efficient may mean all or nothing, due to the characteristic of user perspective: a DSS may be very efficient for a developers needs, and completely inefficient for production objectives. In order to improve the global DSS efficiency, companies can act upon (i) cost and/or (ii) expectations. They must be capable of well defining and exhaustively taking into account the costs and expectations relative to their DSS. Concerning the first aspect of cost, the Total Cost of Ownership (TCO) provides a good assessment. It contains two principal aspects: hardware/software costs and cost of operations. With hardware/software, companies are limited by market offers and technological evolutions. The cost of operations, in addition to the potential gain by resource mutualisation, strongly influences the organization costs in rapport with its pertinence. Over this problematic, companies have implemented rationalization based on best practice guidelines, such as the Information Technology Integration Library (ITIL). Nevertheless, an internal study we have conducted at SP2 Solutions in 2007 showed that these guidelines are little or not-known to the DSS experts within or outside company IT departments. Our experience shows that this hasn’t changed almost at all since then. Relating with the second aspect, we state that user expectations cannot be ignored when evaluating the efficiency of a DSS. We estimate that the expected results must be based on the Quality of Service (QoS) and not on the raw technical performances. This way we define the DSS efficiency as a ratio between the DSS QoS and the DSS TCO, which is the foundation of our propositions.

User-System Interaction for Redundancy-Free Knowledge Discovery in Data

A classical limit of association rule at the decider’s point of view is in the combinatorial nature of the association rules, resulting in numerous rules. As the overall quality of an association rule set can be considered as insight of the studied domain given to the decider by the interpretation of the rules, too many rules can make an harder interpretation then a worse quality of the overall process. To get more readable rules and thus improve this global quality criterion, we apply techniques initially designed for redundancy reduction in functional dependencies sets to association rules. Although the two kinds of relations have different properties, this method allow very concise representations that are easily understood by the decider and can be further exploited for automatic reasoning. In this paper, we present this method, compare it to other approaches and apply it to synthetic datasets. We end with a discussion about the information loss resulted of the simplification.