It is known that the data preparation phase is the most time consuming in the data mining process, using up to 50 % or up to 70 % of the total project time. Currently, data mining methodologies are of general purpose and one of their limitations is that they do not provide a guide about what particular task to develop in a specific domain. This paper shows a new data preparation methodology oriented to the epidemiological domain in which we have identified two sets of tasks: General Data Preparation and Specific Data Preparation. For both sets, the Cross-Industry Standard Process for Data Mining (CRISP-DM) is adopted as a guideline. The main contribution of our methodology is fourteen specialized tasks concerning such domain. To validate the proposed methodology, we developed a data mining system and the entire process was applied to real mortality databases. The results were encouraging because it was observed that the use of the methodology reduced some of the time consuming tasks and the data mining system showed findings of unknown and potentially useful patterns for the public health services in Mexico.

The object clustering problem, according to their similarity measures, can be formulated as a combinatorial optimization problem. The K-Means algorithm has been widely used for solving such problem; however, its computational cost is very high. In this work a new heuristics is proposed for reducing the computational complexity in the classification step of the algorithm based on a honeycomb structure, which the algorithm builds when clusters are visualized in a two-dimensional space. In particular it has been observed that an object can only change membership to neighboring clusters. The heuristics consists of performing distance calculations only with respect to centroids of neighboring clusters, which reduces the number of calculations. For assessing the performance of this heuristics, a set of experiments was carried out that involved 2500, 10000 and 40000 objects uniformly distributed in a two-dimensional space, as well as real-world instances of 3100 and 245 057 objects with 2 and 3 dimensions. The results were encouraging, since the calculation time was reduced 65% on average, with respect to the standard K-Means algorithm for the synthetic instance, and up to 62% on average for the real-world instances, while the quality was reduced on average by 0.05% and 2.5%, respectively.

With the increasing presence of Big Data there arises the need to group large instances. These instances present a number of objects with multidimensional features, which require to be grouped in hundreds or thousands of clusters. This article presents a new improvement to the K-means algorithm, which is oriented to the efficient solution of instances with a large number of clusters and dimensions. This heuristic is called Honeycomb (HC) and it is based on the relationship between the number of dimensions and the number of centroids that form a neighborhood. That is, the heuristic allows the reduction of the number of distance calculations for each object. The heuristic was validated by solving a set of synthetic instances obtaining reductions in execution time of up to 90% and a quality reduction of less than 1 %, with respect to standard K-means. For real instances of low and high dimensionality, HC obtained a reduction of execution time between 84.74% and 95.44% with a quality reduction between 1.07% and 1.62 %, respectively. The experimental results are encouraging because this heuristic would benefit those domains that require instances with a continuous increase of the number of objects, dimensions and clusters.