doc. Ing. Kamil Dedecius, Ph.D.

doc. Ing. Kamil Dedecius, Ph.D.

RNDr. Vladimíra Sečkárová, Ph.D.

Department of Applied Mathematics

This work explores theory of Sentiment Analysis and investigates the possibilities of utilization of sentiments left on social media platforms for forecasting of UFC fight outcomes. Further, it utilizes already proposed approaches for usage of past statistics for prediction of future fights. This work proves that sentiments left online can be of great value for UFC outcome forecasting, as it achieves high accuracy for the case study which is conducted with the usage of posts from Twitter social media platform.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Radomír Žemlička

Department of Theoretical Computer Science

This thesis analyses the national time series of newly infected people by COVID-19. The data are taken from the World Health Organization. In the preprocessing phase are the national time series scaled to respect the size of the population. The dimension is reduced using the Piecewise aggregate approximation and just the trend component of the time series is taken into account. In the thesis, there are defined four measures of time series (dis)similarity like Dynamic Time Warping (DTW), Edit Distance With Real Penalty (ERP), Longest Common Subsequence Similarity LCSS, and Discrete Fréchet distance. In the following phase, the preprocessed data are clustered using the agglomerative hierarchical clustering algorithm with the use of the average linkage that exploits the defined measures. In the last phase, the resulting count of clusters is chosen for each metric using the dendrogram. In the conclusion of this thesis, there are the resulting plots, which are further discussed together with the properties of the distance measures.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Tomáš Vlk

Department of Applied Mathematics

This thesis focuses on the problem of tracking multiple targets, which is still a very challenging discipline. It deals with the uncertainty about the measurement-target association in a noisy-cluttered-environment. Furthermore, targets can appear and disappear more or less randomly. The solution described in this thesis is based on the Bayesian inference and consists of the probability hypothesis density (PHD) filters. This approach involves modeling the relevant sets of targets and measurements as random finite sets. The probability density recursions (PHDs) are used to propagate the posterior intensity, which is a first-order statistic of a random finite set of targets, over time.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Karel Klouda, Ph.D.

Department of Applied Mathematics

Streaming data has recently attracted attention in numerous fields such as IoT, social networks, and e-Commerce. Streaming data is the data that is continuously generated by some sources and there is a need for their real-time processing. Given that the frequency of a stream may be very high, the methods for its processing must be computationally cheap and have low memory requirements. Autoregressive (AR) models are one of the fundamental approaches to time series modelling. The idea behind the autoregressive models is that the current value of the series is linearly dependent on its most recent past values. Thanks to its structure, AR models are able to effectively process streaming data even in situations when more complex models cannot achieve the desired performance for their computational and memory burden. This thesis investigates AR processes and their properties, outlines the theory behind them, and provides an illustrative example of AR modelling on real data related to the COVID-19 pandemic course in Czech Republic.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Ondřej Tichý, Ph.D.

Department of Applied Mathematics

Logistic regression is a classification method used in machine data processing. This thesis deals with the optimization of the Bayesian estimate of the parameters of the logistic model using Laplace approximation. The state-of-the-art methods usually rely on suboptimal Laplace-type approximations. This thesis goes one step further. It investigates the impact of a more precise approximation on the estimation quality. The results obtained by computationally intensive optimization are compared with the traditional less intensive but also more imprecise.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Radomír Žemlička

Department of Applied Mathematics

In the years 2020-2021, the whole world is experiencing difficulties caused by the global epidemic of the new SARS-CoV-2 virus. This thesis aims at the statistical analysis of the time series related to the COVID-19 pandemic in the Czech Republic. It explores the usability of the Facebook Prophet and Seasonal Autoregressive Integrated Moving Average (SARIMA) models to analyze and predict the development of pandemic processes. Moreover, this thesis connects the global changes that occurred in the selected time series with the provided government restrictions (lockdown, wearing a face mask, and so on).

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

RNDr. Vladimíra Sečkárová, Ph.D.

Department of Applied Mathematics

Recommender systems are now an almost everyday part of our lives. Searching for a new book based on personal reading preferences takes considerable effort. To simplify the process, the majority of online retailers are utilizing the power of recommender systems to promote more products to their customers. In this thesis, we show the state-of-the-art recommender systems and analyze the different approaches for the recommendation process. Based on this analysis, we implement a recommendation system for books. Our system combines multiple recommendation algorithms and provides the user with the possibility to control the niche level of the recommendations.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Ondřej Tichý, Ph.D.

Department of Applied Mathematics

This thesis explores the problem of distributed Kalman filtering under partially heterogeneous models. A modification to the existing diffusion Kalman filter is proposed, enabling the employment of partially heterogeneous models in the diffusion networks. The performance of the less complex models is futher improved by the implementation of a node failure detection heuristic, resetting the failling nodes, and giving them a chance at a recovery.

Thesis on DSpace

Ing. Kamil Dedecius, Ph.D.

Ing. Tomáš Vondra, Ph.D.

Department of Applied Mathematics

This thesis deals with electricity price forecasting based on weather conditions. The research of this topic is of great interest because accurate forecasting is essential for both the electricity producers and consumers. Due to the inability of storing a large amount of electricity efficiently and the necessity to balance between production and consumption, the short-term electricity price tends to be very volatile, which makes it difficult to predict it accurately. There are many factors that affect the electricity price. Especially in Central and Western Europe, very important factor is the weather, due to increasing number of solar and wind power plants. This thesis aims to analyze available data about weather forecasts and the electricity prices and create predictive models for this data.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

doc. Ing. Tomáš Čejka, Ph.D.

Department of Applied Mathematics

Network anomaly detection is crucial for identifying issues in computer networks, with particular significance in large-scale global networks like the Worldwide LHC Computing Grid, that is used for processing data from CERN experiments. This thesis aims to enhance or replace the existing heuristic network anomaly detection algorithm, primarily designed to identify routing changes based on traceroute measurements between global datacenters, generated through the perfSonar software. The work demonstrates the informational richness of traceroute measurements and proposes several Bayesian inference-based models to improve anomaly detection in traceroute data.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

RNDr. Vladimíra Sečkárová, Ph.D.

Department of Applied Mathematics

Time series classification is used in a wide variety of real-world problems, from health analysis to food quality assurance. Despite its usage, the extent of understanding the reasons for the classification determined by the model is insufficient. This thesis proposes a novel type of explainable classifier based on ARIMA models, which are heavily used for time series forecasting. The time series dataset is converted to ARIMA coefficients and classified using random forests classifier. The explainability is provided by established explanation methods for tabular data. Besides the ARIMA models, the thesis additionally provides a method for explaining raw time series data, built upon the LIME method.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Tomáš Vlk

Department of Applied Mathematics

As is well-known, state-space models are widely used to model different observable processes in various fields, from finance to biology. Based on the observed sequence, the main interest is to deduce the sequence of hidden states that gave rise to them. The primary goal of this thesis is to study the performance and robustness of various Bayesian algorithms used for online state filtering under both well-specified and misspecified state-space models. This is due to the fact that in practice, for various reasons, there is often a lack of knowledge of process noise or measurements. This thesis, in particular, touches on the rather popular case where the measurement model is misspecified. As part of this work, the emphasis is placed on an introduction to the filters from the most popular family, the Kalman family, and consider them from a Bayesian perspective, also not without the particle filter, which is regarded as more stable in the context of application to misspecified models. Attention will also be paid, and experiments will be conducted with so-called approximate Bayesian filters, which allow complete ignorance of the noise distribution of measurements. The result of the thesis should be the coverage of the theory of all of the above and the results of experiments, not excluding the definition of plans for the future development of this topic.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Ondřej Tichý, Ph.D.

Department of Applied Mathematics

This thesis explores the problem of distributed Bayesian sequential estimation of unknown state-spacemodels with unknown processes and measurement noise covariance matrices. This is a frequent problem in real-world scenarios, where the information about noise covariance matrices for specific sensors may not be available. The solution proposed in this thesis is built upon the variational Bayesian paradigm, which is used for the estimation of the states, as well as the unknown measurement noise covariance matrix. From performance improvements, the measurements and posterior estimates are shared between the adjacent node in the network. It also shows a way of optimizing the process noise covariance matrix.

Thesis on DSpace

doc. Ing. Kamil Dedecius, Ph.D.

Ing. Tomáš Vlk

Department of Applied Mathematics

The thesis focuses on the multi-target tracking problem in cluttered environments with the uncertainty on the number of targets and using Bayesian inference. In this work, we mainly focus on the popular Gaussian mixture probability hypothesis density (GM-PHD) filter and introduce a technique to include additional information for cases when a sensor fails to detect targets due to environmental or physical limitations. We provide all required theoretical background from the basics of probability theory to the discussion of various multi-target tracking methods and the introduction of the Finite Set Statistics (FISST) framework. We also conclude with an extensive performance measurement and analysis of results, where we demonstrate that the proposed fusion technique significantly improves tracking results. Finally, we discuss the limitations of the filter and propose possible measures to overcome them.

Thesis on DSpace

Ing. Kamil Dedecius, Ph.D.

Department of Applied Mathematics

The Poisson regression is a popular generalized linear model used to model discrete count variables. This thesis is focused on the problem of its sequential estimation under potentially slowly time-varying regression coefficients. A convenient approximation by normal distribution is used to do so in the Bayesian setting. Also, a calibration technique is discussed to enhance the estimation quality. Finally, a use case of the proposed approach in the signal processing domain is suggested, in particular, its application in diffusion networks to perform distributed collaborative estimation.

Thesis on DSpace

Ing. Kamil Dedecius, Ph.D.

Department of Applied Mathematics

This thesis looks into the ways of enhancing the local predictions in agent networks using the information shared by their neighbouring nodes. Agents locally employ different models, thus the diversity of information to share is limited. Multiple variants of improving the predictions are proposed, some of them taking the prediction certainty into consideration. These are then implemented and compared on both simulated and real-world data.

Thesis on DSpace