Catálogo de publicaciones - libros

Finding complex patterns in long temporal or spatial sequences from real world applications is gaining increasing interest in data mining. However, standard data mining techniques, taken in isolation, seem to be inadequate to cope with such a task. In fact, symbolic approaches show difficulty in dealing with noise, while non-symbolic approaches, such as neural networks and statistics show difficulty in dealing with very long subsequences where relevant episodes may be interleaved with large gaps. The way out we suggest is to integrate the logic approach with non-symbolic methods in a unified paradigm, as it has been already done in other Artificial Intelligence tasks. We propose a framework where a high level knowledge representation is used to incorporate domain specific knowledge, to focus the attention on relevant episodes during the mining process, and flexible matching algorithms developed in the pattern recognition area are used to deal with noisy data. The knowledge extraction process follows a machine learning paradigm combining inductive and deductive learning, where deduction steps can be interleaved with induction steps aimed at augmenting a weak domain theory with knowledge extracted from the data. Our framework is formally characterized and then is experimentally tested on an artificial dataset showing its ability at dealing with noise and with the presence of long gaps between the relevant episodes.

Frequent patterns are often used for discovery of several types of knowledge such as association rules, episode rules, sequential patterns, and clusters. Since the number of frequent itemsets is usually huge, several lossless representations have been proposed. Frequent closed itemsets and frequent generators are the most useful representations from application point of view. Discovery of closed itemsets requires prior discovery of generators. Generators however are usually discovered directly from the data set. In this paper we will prove experimentally that it is more beneficial to compute the generators representation in two phases: 1) by extracting the generalized disjunction-free generators representation from the database, and 2) by transforming this representation into the frequent generators representation. The respective algorithm of transitioning from one representation to the other is proposed.

Model trees are an extension of regression trees that associate leaves with multiple regression models. In this paper a method for the top-down induction of model trees is presented, namely the Stepwise Model Tree Induction (SMOTI) method. Its main characteristic is the induction of trees with two types of nodes: regression nodes, which perform only straight-line regression, and split nodes, which partition the sample space. The multiple linear model associated to each leaf is then obtained by combining straight-line regressions reported along the path from the root to the leaf. In this way, internal regression nodes contribute to the definition of multiple models and have a “global” effect, while straight-line regressions at leaves have only “local” effects. This peculiarity of SMOTI has been evaluated in an empirical study involving both real and artificial data.

Recently, many solutions and sites related to the intelligent agent are created in order to provide good services for customers. Moreover, some new proposals including the collaborative filtering are put forward in the field of electronic commerce (EC) solutions. However, these proposals are lack of the add-on characteristics. In fact, it seems that only a few intelligent systems could provide the recommendations to the customers for the items that they really want to purchase, by means of the collaborative filtering algorithm based on their previous evaluation data. In this paper, we propose the CLASG (Clustering And Similarity Grouplens) collaborative filtering agent algorithm. The CLASG algorithm is the one that uses both the GroupLens algorithm and the clustering method. We have evaluated its performance with enough experiments, and the results show that the proposed method provides more stable recommendations than GroupLens does. We developed the MindReader, which makes it possible to have the correct predictions and recommendations with less response time than ever, as an automated recommendation system that includes both of CLASG algorithm and WhoLiked agent. It can be readily integrated into the existing EC solutions since it has an add-on characteristic, which is lacked in the past solutions.

Sources of XML documents are proliferating on the Web and documents are more and more frequently exchanged among sources. At the same time, there is an increasing need of exploiting database tools to manage this kind of data. An important novelty of XML is that information on document structures is available on the Web together with the document contents. However, in such an heterogeneous environment as the Web, it is not reasonable to assume that XML documents that enter a source always conform to a predefined DTD in the source. In this paper we address the problem of document classification by proposing a metric for quantifying the structural similarity between an XML document and a DTD. Based on such notion, we propose an approach to match a document entering a source against the set of DTDs available in the source, determining whether a DTD exists similar enough to the document.

Data mining is a useful decision support technique that can be used to discover production rules in warehouses or corporate data. Data mining research has made much effort to apply various mining algorithms efficiently on large databases. However, a serious problem in their practical application is the long processing time of such algorithms. Nowadays, one of the key challenges is to integrate data mining methods within the framework of traditional database systems. Indeed, such implementations can take advantage of the efficiency provided by SQL engines.

In this paper, we propose an integrating approach for decision trees within a classical database system. In other words, we try to discover knowledge from relational databases, in the form of production rules, via a procedure embedding SQL queries. The obtained decision tree is defined by successive, related relational views. Each view corresponds to a given population in the underlying decision tree. We selected the classical Induction Decision Tree (ID3) algorithm to build the decision tree. To prove that our implementation of ID3 works properly, we successfully compared the output of our procedure with the output of an existing and validated data mining software, SIPINA. Furthermore, since our approach is tuneable, it can be generalized to any other similar decision tree-based method.

We present , an efficient method for mining frequent sequential association rules from multiple sequential data sets with a time lag between the occurrence of an antecedent sequence and the corresponding consequent sequence. This approach finds patterns in one or more sequences that precede the occurrence of patterns in other sequences, with respect to user-specified constraints. In addition to the traditional frequency and support constraints in sequential data mining, this approach uses separate antecedent and consequent inclusion constraints. Moreover, separate antecedent and consequent maximum window widths are used to specify the antecedent and consequent patterns that are separated by the maximum time lag.

We use multiple time series drought risk management data to show that our approach can be effectively employed in real-life problems. The experimental results validate the superior performance of our method for efficiently finding relationships between global climatic episodes and local drought conditions. We also compare our new approach to existing methods and show how they complement each other to discover associations in a drought risk management decision support system.

Problems of discovering association rules in data sets containing semantic information about preference orders on domains of attributes are considered. Such attributes are called criteria and they are typically present in data related to economic issues, like financial or marketing data. We introduce a specific form of association rules involving criteria. Discovering such rules requires new concepts: semantic correlation of criteria, inconsistency of objects with respect to the dominance, credibility index. Properties of these rules concerning their generality and interdependencies are studied. We also sketch the way of mining such rules.

Real-world data usually contain a certain percentage of unknown (missing) attribute values. Therefore efficient robust data mining algorithms should comprise some routines for processing these unknown values. The paper [] figures out that each dataset has more or less its own ’favourite’ routine for processing unknown attribute values. It evidently depends on the magnitude of noise and source of unknownness in each dataset. One possibility how to solve the above problem of selecting the right routine for processing unknown attribute values for a given database is exhibited in this paper. The covering machine learning algorithm CN4 processes a given database for six routines for unknown attribute values independently. Afterwards, a meta-learner (meta-combiner) is used to derive a meta-classifier that makes up the overall (final) decision about the class of input unseen objects.

The results of experiments with various percentages of unknown attribute values on real-world data are presented and performances of the meta-classifier and the six base classifiers are then compared.

In this paper, we present an intelligent buffer cache management algorithm in multimedia data retrieval called . The proposed ABM scheme automatically detects the reference pattern of each file and intelligently switches between different buffer cache management schemes on per-file basis. According to our simulation based experiment, the ABM scheme yields better buffer cache miss ratio than the legacy buffer cache schemes such as LRU or interval based scheme. The ABM scheme manifests itself when the workload exhibits not only sequential but also looping reference patterns.