Quispe Vargas, Walter

Profile Picture

Filters

2006 - 200912010 - 20191
Reset filters

Settings

Citations

Publication Search Results

Now showing 1 - 2 of 2
  • Thumbnail Image
    Publication
    Sieve bootstrap en series de tiempo de nubosidad en el caribe
    (2006) Quispe Vargas, Walter; Ramírez Beltrán, Nazario D.; College of Arts and Sciences - Sciences; Saito, Tokuji; Acuña Fernández, Edgar; Department of Mathematics; Aponte Avellanet, Nilda E.
    The Sieve Bootstrap is a resampling method, designed to deal with autocorrelated data, specifically a sequence of information taken at equal time intervals. Formally, the Sieve Bootstrap approximates a linear process by a sequence of autoregressive processes of order p= p(n), where p(n)→∞, [p(n)/n]→0 as the sample size n→∞ for a time series that is n expressed by an autoregressive AR(p(n)) model, it should be noted that the bootstrap is constructed over the residuals. In this thesis, we apply the Sieve Bootstrap to the construction of prediction intervals of cloudiness time series on Caribbean, which were obtained from the data base of level D2-DATA of the International Satellite Cloud Climatology Project. Results obtained show that the Sieve Bootstrap method provides a better prediction interval coverage. However the Box Jenkins technique shows a significance reduction in the length of the prediction intervals.
  • Thumbnail Image
    Publication
    Temporal outlier detection using dynamic Bayesian networks and probabilistic association rules
    (2019-12-10) Quispe Vargas, Walter; Acuña Fernández, Edgar; College of Engineering; Rolke, Wolfgang; Schutz Schmuck, Marko; Aparicio Carrasco, Roxana; Department of Electrical and Computer Engineering; Bartolomei Suárez, Sonia M.
    Temporal datasets provide records of the evolution and dependencies of random variables over time. Recently, there has been an increase in the application of temporal datasets in areas such as intrusion detection, fraud detection, activity recognition, etc. Interesting temporal outliers are anomalies that incorporate important or new information and contradict the causal probabilistic relationship in the domain knowledge described in a temporal dataset. One main objective of Data Mining is to discover interesting temporal anomalous patterns. Moreover, provide contextualization of the interestingness of the reported outliers. Most of the methods used to discover temporal outliers are reduction-based, losing valuable information in the discovery process. On the other hand, there are scarce studies about the interestingness of reported temporal outliers. Even less, to provide contextualization of the anomaly causes.
This thesis deals with the problem of discovering these interesting temporal outliers in datasets. We present probabilistic association rules as measures to discover interesting temporal outliers based on domain knowledge that has been learned and represented by a Dynamic Bayesian Network. Dynamic Bayesian networks are models to represent complex stochastic processes, to establish probabilistic dependencies in the feature space over time, and to capture the background knowledge in a causal relationship between features. The two probabilistic association rules: i) low support & high confidence, and ii) high support & low confidence, were used to identify scenarios where the discrepancies between prior and conditional probabilities are significant. Our novel approach coalesces both methods. It allows us to discover interesting temporal outliers and provide contextualization in the form of relational subspaces, under the proposed methodology called “Domain Specific Temporal Anomalous Patterns.” The evaluation of the proposed methodology was done on synthetic and real temporal datasets on the unsupervised and supervised scenario. The experimental results on temporal datasets show that our approach can detect genuine temporal outliers and provide relational subspaces to explain the probable causes of the reported outliers, with reasonable efficiency measures. In this way, our technique becomes a state of the art method to discover interesting temporal outliers in temporal datasets. Designed to provide contextual information of the reported outliers; this, in turn, can be used to improve our understanding of the domain knowledge and the underlying temporal data generating process.