Catálogo de publicaciones - libros

We show how a previously derived method of using reinforcement learning for supervised clustering of a data set can lead to a sub-optimal solution if the cluster prototypes are initialised to poor positions. We then develop three novel reward functions which show great promise in overcoming poor initialization. We illustrate the results on several data sets. We then use the clustering methods with an underlying latent space which enables us to create topology preserving mappings. We illustrate this method on both real and artificial data sets.

Proposing efficient techniques for discovery of useful information and valuable knowledge from very large databases and data warehouses has attracted the attention of many researchers in the field of data mining. The well-known Association Rule Mining (ARM) algorithm, Apriori, searches for frequent itemsets (i.e., set of items with an acceptable support) by scanning the whole database repeatedly to count the frequency of each candidate itemset. Most of the methods proposed to improve the efficiency of the Apriori algorithm attempt to count the frequency of each itemset without re-scanning the database. However, these methods rarely propose any solution to reduce the complexity of the inevitable enumerations that are inherited within the problem. In this paper, we propose a new algorithm for mining frequent itemsets and also association rules. The algorithm computes the frequency of itemsets in an efficient manner. Only a single scan of the database is required in this algorithm. The data is encoded into a compressed form and stored in main memory within a suitable data structure. The proposed algorithm works in an iterative manner, and in each iteration, the time required to measure the frequency of an itemset is reduced further (i.e., checking the frequency of n-dimensional candidate itemsets is much faster than those of n-1 dimensions). The efficiency of our algorithm is evaluated using artificial and real-life datasets. Experimental results indicate that our algorithm is more efficient than existing algorithms.

The Bio-Basis Function Neural Network (BBFNN) is a successful neural network architecture for peptide classification. However, the selection of a subset of peptides for a parsimonious network structure is always a difficult process. We present a Sparse Bayesian Bio-Kernel Network in which a minimal set of representative peptides can be selected automatically. We also introduce per-residue weighting to the Bio-Kernel to improve accuracy and identify patterns for biological activity. The new network is shown to outperform the original BBFNN on various datasets, covering different biological activities such as as enzymatic and post-translational-modification, and generates simple, interpretable models.

Support Vector Regression (SVR) is usually pursued using the –insensitive loss function while, alternatively, the initial regression problem can be reduced to a properly defined classification one. In either case, slack variables have to be introduced in practical interesting problems, the usual choice being the consideration of linear penalties for them. In this work we shall discuss the solution of an SVR problem recasting it first as a classification problem and working with square penalties. Besides a general theoretical discussion, we shall also derive some consequences for regression problems of the coefficient structure of the resulting SVMs and illustrate the procedure on some standard problems widely used as benchmarks and also over a wind energy forecasting problem.

A fuzzy rule-based classification system (FRBCS) is one of the most popular approaches used in pattern classification problems. One advantage of a fuzzy rule-based system is its interpretability. However, we’re faced with some challenges when generating the rule-base. In high dimensional problems, we can not generate every possible rule with respect to all antecedent combinations. In this paper, by making the use of some data mining concepts, we propose a method for rule generation, which can result in a rule-base containing rules of different lengths. As the next phase, we use rule-weight as a simple mechanism to tune the classifier and propose a new method of rule-weight specification for this purpose. Through computer simulations on some data sets from UCI repository, we show that the proposed scheme achieves better prediction accuracy compared with other fuzzy and non-fuzzy rule-based classification systems proposed in the past.

As data analysis tasks often have to face the analysis of huge and complex data sets there is a need for new algorithms that combine vector quantization and mapping methods to visualize the hidden data structure in a low-dimensional vector space. In this paper a new class of algorithms is defined. Topology representing networks are applied to quantify and disclose the data structure and different nonlinear mapping algorithms for the low-dimensional visualization are applied for the mapping of the quantized data. To evaluate the main properties of the resulted topology representing network based mapping methods a detailed analysis based on the wine benchmark example is given.

The appraisement of rules and rule sets is very important in data mining. The information content of rules is discussed in this paper and is categorized into inner mutual information and outer impartation information. We put forward the viewpoint that the outer impartation information content of rules and rule sets can be represented by relations from input universe to output universe. Then, the interaction of rules in a rule set can be represented by the union and intersection of binary relations expediently. Based on the entropy of relations, the outer impartation information content of rules and rule sets are well measured. Compared with the methods which appraise rule sets by their overall performance (accuracy, error rate) on the given test data sets, the outer impartation information content of rule sets is more objective and convenient because of the absence of test data sets.

Both the number and complexity of Data Mining projects has increased in late years. Unfortunately, nowadays there isn’t a formal process model for this kind of projects, or existing approaches are not right or complete enough. In some sense, present situation is comparable to that in software that led to ’software crisis’ in latest 60’s. Software Engineering matured based on process models and methodologies. Data Mining’s evolution is being parallel to that in Software Engineering. The research work described in this paper proposes a Process Model for Data Mining Projects based on the study of current Software Engineering Process Models (IEEE Std 1074 and ISO 12207) and the most used Data Mining Methodology CRISP-DM (considered as a “facto” standard) as basic references.

In this investigation we propose a novel approach for classifying polyphonic melodies. Our main idea comes from for automatic classification of polyphonic melodies by where the states correspond to well-tempered chords over the music and the observation sequences to some feature values called . The similarity among harmonies can be considered by means of the features and well-tempered chords. We show the effectiveness and the usefulness of the approach by some experimental results.

Frequently, organisations have to face complex situations where decision making is difficult. In these scenarios, several related decisions must be made at a time, which are also bounded by constraints (e.g. inventory/stock limitations, costs, limited resources, time schedules, etc). In this paper, we present a new method to make a good global decision when we have such a complex environment with several local interwoven data mining models. In these situations, the best local cutoff for each model is not usually the best cutoff in global terms. We use simulation with Petri nets to obtain better cutoffs for the data mining models. We apply our approach to a frequent problem in customer relationship management (CRM), more specifically, a direct-marketing campaign design where several alternative products have to be offered to the same house list of customers and with usual inventory limitations. We experimentally compare two different methods to obtain the cutoff for the models (one based on merging the prospective customer lists and using the local cutoffs, and the other based on simulation), illustrating that methods which use simulation to adjust model cutoff obtain better results than a more classical analytical method.