Out-of-hospital emergency departments receive multiple types of requests daily. Their management requires a balance to be found between available resources and the actual needs of the requesting party. Those regarding suicidal behaviour, which are resource heavy, are few in number in terms of the bulk of requests, and detecting them correctly is therefore important. Previous research, using machine learning algorithms to analyse suicide, has typically focused on discovering insights to be used by medical personnel. This proposal extends its use in two directions: knowledge that can be used by non-exclusively medical staff, such as telephone operators, and the models that have been incorporated into a software prototype to help in the decision-making of an emergency department. In addition, previous research has often included a range of information from different sources that are not available when processing an emergency call request, for example, data that is only obtained at the end of the intervention. A full-scale data mining process has been performed using data from the out-of-hospital emergency service in Malaga (Spain). Sensitivity has been the primary goal to avoid missing cases requiring special attention, but this objective has been pursued without overlooking a good trade-off with specificity. The best models can offer such a compromise between sensitivity and specificity, and show more than 80% in both metrics simultaneously. The experts validate that the modelling phase showed that the algorithms have automatically identified already known situations. This lays the groundwork for further iterations with a promising outlook.

This paper applies the Behavior Pattern Learning (BPL) approach to automatically discover the conditions previous to specific macroeconomic events from the historical evolution of macroeconomic variables, as well as historical events. Every time the target macroeconomic event (e.g. Consumer Price Index (CPI) inflation greater than 5.5%) takes place, BPL constructs several snapshots at specific previous times. Each snapshot, called a behavior summary, includes information about its present, past, and future. The set of behavior summaries is used by a regression tree approach to discover the conditions for the target event to take place. In this paper, we used BPL to automatically discover the conditions for the occurrence of specific events associated with the CPI inflation and Gross domestic product (GDP) deflactor using macroeconomic data, and war and political events. Comparisons with other prediction models on the same problem show that the BPL provides automatic forecasts that are similar in performance as those forecasts done by models assisted by experts.

Classification trees are a powerful tool for mining non-stationary data streams. In these situations, mas sive data are constantly generated at high speed and the underlying target function can change over time. The iadem family of algorithms is based on Hoeffding’s and Chernoff’s bounds and induces online deci sion trees from data streams, but is not able to handle concept drift. This study extends this family to deal with time-changing data streams. The new online algorithm, named iadem-3, performs two main actions in response to a concept drift. Firstly, it resets the variables affected by the change and main tains unbroken the structure of the tree, which allows for changes in which consecutive target functions are very similar. Secondly, it creates alternative models that replace parts of the main tree when they significantly improve the accuracy of the model, thereby rebuilding the main tree if needed. An online change detector and a non-parametric statistical test based on Hoeffding’s bounds are used to guaran tee this significance. A new pruning method is also incorporated in iadem-3, making sure that all split tests previously installed in decision nodes are useful. The learning model is also viewed as an ensem ble of classifiers, and predictions of the main and alternative models are combined to classify unlabeled examples. iadem-3 is empirically compared with various well-known decision tree induction algorithms for concept drift detection. We empirically show that our new algorithm often reaches higher levels of accuracy with smaller decision tree models, maintaining the processing time bounded, irrespective of the number of instances processed.

Incremental and online learning algorithms are more relevant in the data mining context because of the increasing necessity to process data streams. In this context, the target function may change overtime, an inherent problem of online learning (known as concept drift). In order to handle concept drift regardless of the learning model, we propose new methods to monitor the performance metrics measured during the learning process, to trigger drift signals when a significant variation has been detected. To monitor this performance, we apply some probability inequalities that assume only independent, univariate and bounded random variables to obtain theoretical guarantees for the detection of such distributional changes. Some common restrictions for the online change detection as well as relevant types of change (abrupt and gradual) are considered. Two main approaches are proposed, the first one involves moving averages and is more suitable to detect abrupt changes. The second one follows a widespread intuitive idea to deal with gradual changes using weighted moving averages. The simplicity of the proposed methods, together with the computational efficiency make them very advantageous. We use a Naive Bayes classifier and a Perceptron to evaluate the performance of the methods over synthetic and real data.

Multi-label learning is a very interesting field in Machine Learning. It allows to generalise standard methods and evaluation procedures, and tackle challenging real problems where one example can be tagged with more than one label. In this paper we study the performance of different multi-label methods in combination with standard single-label algorithms, using several specific multi-label metrics. What we want to show is how a good preprocessing phase can improve the performance of such methods and algorithms. As we will explain, its main advantage is a shorter time to induce the models, while keeping (even improving) other classification quality measures. We use the GNUsmail framework to do the preprocessing of an existing and extensively used dataset, to obtain a reduced feature space that conserves the relevant information and allows improvements on performance. Thanks to the capabilities of GNUsmail, the preprocessing step can be easily applied to different email datasets.

Discovering frequent factors from long strings is an important problem in many applications, such as biosequence mining. In classical approaches, the algorithms process a vast database of small strings. However, in this paper we analyze a small database of long strings. The main difference resides in the high number of patterns to analyze. To tackle the problem, we have developed a new algorithm for discovering frequent factors in long strings. This algorithm uses a new data structure to arrange nodes in a trie. A positioning matrix is defined as a new positioning strategy. By using positioning matrices, we can apply advanced prune heuristics in a trie with a minimal computational cost. The positioning matrices let us process strings including Short Tandem Repeats and calculate different interestingness measures efficiently. The algorithm has been successfully used in natural language and biological sequence contexts.

Exact calculation of the TF-IDF weighting function in massive streams of documents involves challenging memory space requirements. In this work, we propose TF-SIDF, a novel solution for extracting relevant words from streams of documents with a high number of terms. TF-SIDF relies on the Count-Min Sketch data structure, which allows to estimate the counts of all the terms in the stream. Results of the experiments conducted with two dataset show that this sketch-based algorithm achieves good approximations of the TF-IDF weighting values (as a rule, the top terms with highest TF-IDF values remaining the same), while substantial savings in memory usage are observed. It is also observed that the performance is highly correlated with the sketch size, and that wider sketch configurations are preferable given the same sketch size.

Large amounts of information are available on the Internet in the form of natural language text that can be processed as a stream of documents. Users need solutions that summarize the vast volume of data. Word clouds are a popular graphical representation approach that allows them to obtain such a quick visual summary. Nevertheless, the exact solution to this problem has high memory requirements, and is not scalable as the collection size grows up. In this work, we provide a method for approximate online computation of word clouds to summarize the contents of each label or category of a given text stream, using only the most relevant terms of each document according to some weighting function. We experimentally show that our method, based on sketching techniques, obtains a good performance while using a restricted quantity of memory.

Email foldering is a challenging problem mainly due to its high dimensionality and dynamic nature. This work presents a comparison of several feature extraction/ selection and adaptive strategies aimed at coping with these two main difficulties. To this end, several studies have been carried out using the ENRON email dataset in order to test how different configuration settings for feature extraction/selection and adapting processes can affect the classification performance.

Real-time classification of massive email data is a challenging task that presents its own particular difficulties. Since email data presents an important temporal component, several problems arise: emails arrive continuously, and the criteria used to classify those emails can change, so the learning algorithms have to be able to deal with concept drift. Our problem is more general than spam detection, which has received much more attention in the literature.

Two extensive research areas in Machine Learning are classification and prediction. Many approaches have been focused in the induction of ensemble to increase learning accuracy of individual classifiers. Recently, new approaches, different to those that look for accurate and diverse base classifiers, are emerging. In this paper we present a system made up of two layers: in the first layer, one ensemble classifier process every example and tries to classify them; in the second layer, one individual classifier is induced using the examples that are not unanimously classified by the ensemble. In addition, the examples that reach to the second layer incorporate new information added in the ensemble. Thus, we can achieve some improvement in the accuracy level, because the second layer can do more informed classifications. In the experimental section we present some results that suggest that our proposal can actually improve the accuracy of the system.

In order to discover behavior patterns, current algorithms only analyze historical data in terms of performance data or fault events, ignoring the temporal correlation among different types of information, including the configuration changes. A method is presented that can discover recurrent patterns from multiple flows of events, such as alarms and configuration events, as well as discrete information, such as traffic and usage, taking into account static and dynamic information concerning observed objects and their environments. This method can filter out theoretically useless patterns, using a novel technique for detecting chaos in sequences of events. The prediction accuracy of the discovered patterns has been measured using objects with dynamic behavior controlled by known and complex differential equations. The proposed mining method has been used for discovering and predicting alarms in a computer network composed of several Internet servers taking into account the alarm and configuration events history, as well as static information about these servers.