Paulo J. Azevedo

Multiresolution motif discovery in time series

Time series motif discovery is an important problem with applications in a variety of areas that range from telecommunications to medicine. Several algorithms have been proposed to solve the problem. However, these algorithms heavily use expensive random disk accesses or assume the data can fit into main memory. They only consider motifs at a single resolution and are not suited to interactivity. In this work, we tackle the motif discovery problem as an approximate Top-K frequent subsequence discovery problem. We fully exploit state of the art iSAX representation multiresolution capability to obtain motifs at different resolutions. This property yields interactivity, allowing the user to navigate along the Top-K motifs structure. This permits a deeper understanding of the time series database. Further, we apply the Top-K space saving algorithm to our frequent subsequences approach. A scalable algorithm is obtained that is suitable for data stream like applications where small memory devices such as sensors are used. Our approach is scalable and disk-efficient since it only needs one single pass over the time series database. We provide empirical evidence of the validity of the algorithm in datasets from different areas that aim to represent practical applications.

Comparing Rule Measures for Predictive Association Rules

We study the predictive ability of some association rule measures typically used to assess descriptive interest. Such measures, namely conviction, lift and i¾?2are compared with confidence, Laplace, mutual information, cosine, Jaccard and i¾?-coefficient. As prediction models, we use sets of association rules. Classification is done by selecting the best rule, or by weighted voting. We performed an evaluation on 17 datasets with different characteristics and conclude that conviction is on average the best predictive measure to use in this setting. We also provide some meta-analysis insights for explaining the results.

Mining approximate motifs in time series

The problem of discovering previously unknown frequent patterns in time series, also called motifs, has been recently introduced. A motif is a subseries pattern that appears a significant number of times. Results demonstrate that motifs may provide valuable insights about the data and have a wide range of applications in data mining tasks. The main motivation for this study was the need to mine time series data from protein folding/unfolding simulations. We propose an algorithm that extracts approximate motifs, i.e. motifs that capture portions of time series with a similar and eventually symmetric behavior. Preliminary results on the analysis of protein unfolding data support this proposal as a valuable tool. Additional experiments demonstrate that the application of utility of our algorithm is not limited to this particular problem. Rather it can be an interesting tool to be applied in many real world problems.

Protein sequence pattern mining with constraints

Considering the characteristics of biological sequence databases, which typically have a small alphabet, a very long length and a relative small size (several hundreds of sequences), we propose a new sequence mining algorithm (gIL). gIL was developed for linear sequence pattern mining and results from the combination of some of the most efficient techniques used in sequence and itemset mining. The algorithm exhibits a high adaptability, yielding a smooth and direct introduction of various types of features into the mining process, namely the extraction of rigid and arbitrary gap patterns. Both breadth or a depth first traversal are possible. The experimental evaluation, in synthetic and real life protein databases, has shown that our algorithm has superior performance to state-of-the art algorithms. The use of constraints has also proved to be a very useful tool to specify user interesting patterns.

Distribution rules with numeric attributes of interest

In this paper we introduce distribution rules, a kind of association rules with a distribution on the consequent. Distribution rules are related to quantitative association rules but can be seen as a more fundamental concept, useful for learning distributions. We formalize the main concepts and indicate applications to tasks such as frequent pattern discovery, sub group discovery and forecasting. An efficient algorithm for the generation of distribution rules is described. We also provide interest measures, visualization techniques and evaluation.

Protein sequence classification through relevant sequence mining and bayes classifiers

We tackle the problem of sequence classification using relevant subsequences found in a dataset of protein labelled sequences. A subsequence is relevant if it is frequent and has a minimal length. For each query sequence a vector of features is obtained. The features consist in the number and average length of the relevant subsequences shared with each of the protein families. Classification is performed by combining these features in a Bayes Classifier. The combination of these characteristics results in a multi-class and multi-domain method that is exempt of data transformation and background knowledge. We illustrate the performance of our method using three collections of protein datasets. The performed tests showed that the method has an equivalent performance to state of the art methods in protein classification.

Mining association rules for label ranking

Recently, a number of learning algorithms have been adapted for label ranking, including instance-based and tree-based methods. In this paper, we propose an adaptation of association rules for label ranking. The adaptation, which is illustrated in this work with APRIORI Algorithm, essentially consists of using variations of the support and confidence measures based on ranking similarity functions that are suitable for label ranking. We also adapt the method to make a prediction from the possibly conflicting consequents of the rules that apply to an example. Despite having made our adaptation from a very simple variant of association rules for classification, the results clearly show that the method is making valid predictions. Additionally, they show that it competes well with state-of-the-art label ranking algorithms.

An experiment with association rules and classification: post-bagging and conviction

In this paper we study a new technique we call post-bagging, which consists in resampling parts of a classification model rather then the data. We do this with a particular kind of model: large sets of classification association rules, and in combination with ordinary best rule and weighted voting approaches. We empirically evaluate the effects of the technique in terms of classification accuracy. We also discuss the predictive power of different metrics used for association rule mining, such as confidence, lift, conviction and χ2. We conclude that, for the described experimental conditions, post-bagging improves classification results and that the best metric is conviction.

Post-processing Operators for Browsing Large Sets of Association Rules

Association rule engines typically output a very large set of rules. Despite the fact that association rules are regarded as highly comprehensible and useful for data mining and decision support in fields such as marketing, retail, demographics, among others, lengthy outputs may discourage users from using the technique. In this paper we propose a post-processing methodology and tool for browsing/visualizing large sets of association rules. The method is based on a set of operators that transform sets of rules into sets of rules, allowing focusing on interesting regions of the rule space. Each set of rules can be then seen with different graphical representations. The tool is web-based and uses SVG. Association rules are given in PMML

Evaluating deterministic motif significance measures in protein databases

BackgroundAssessing the outcome of motif mining algorithms is an essential task, as the number of reported motifs can be very large. Significance measures play a central role in automatically ranking those motifs, and therefore alleviating the analysis work. Spotting the most interesting and relevant motifs is then dependent on the choice of the right measures. The combined use of several measures may provide more robust results. However caution has to be taken in order to avoid spurious evaluations.ResultsFrom the set of conducted experiments, it was verified that several of the selected significance measures show a very similar behavior in a wide range of situations therefore providing redundant information. Some measures have proved to be more appropriate to rank highly conserved motifs, while others are more appropriate for weakly conserved ones. Support appears as a very important feature to be considered for correct motif ranking. We observed that not all the measures are suitable for situations with poorly balanced class information, like for instance, when positive data is significantly less than negative data. Finally, a visualization scheme was proposed that, when several measures are applied, enables an easy identification of high scoring motifs.ConclusionIn this work we have surveyed and categorized 14 significance measures for pattern evaluation. Their ability to rank three types of deterministic motifs was evaluated. Measures were applied in different testing conditions, where relations were identified. This study provides some pertinent insights on the choice of the right set of significance measures for the evaluation of deterministic motifs extracted from protein databases.