Browsing by master's degree program "Master's Programme in Data Science"

Northern peatlands are a large source of methane (CH4) to the atmosphere and can vary strongly depending on local environmental conditions. However, few studies have mapped fine-grained CH4 fluxes at the landscape-level. The aim of this study was to predict land cover and CH4 flux patterns in Pallastunturi, Finland, in a study area dominated by forests, peatlands, fells, and lakes. I used random forest models to map land cover types and CH4 fluxes with multi-source remote sensing data and upscaled CH4 fluxes based on land cover maps. The random forest classifier reliably detected the same land cover patterns as the CORINE Land Cover maps. The main differences between the land cover maps were forest type classification, misclassification between neighboring peatland types, and detection of sparsely vegetated areas on fells. The upscaled CH4 fluxes of sinks were very robust to changes in land cover classification, but shrub tundra and peatland CH4 fluxes were sensitive to the level of detail in the land cover classification. The random forest regression performed well (NRMSE 6.6%, R2 82%) and predicted similar CH4 flux patterns as the upscaled CH4 flux maps, despite predicting larger areas that act as CH4 sources than the upscaled CH4 flux maps. The random forest regressor also better predicted CH4 fluxes in peatlands due to added information about soil moisture content from the remote sensing data. Random forests are a good model choice to detect landscape patterns and predict CH4 patterns in northern peatlands based on remote sensing and topographic data.

The importance of protecting sensitive data from information breaches has increased in recent years due to companies and other institutions gathering massive datasets about their customers, including personally identifiable information. Differential privacy is one of the state-of-the-art methods for providing provable privacy to these datasets, protecting them from adversarial attacks. This thesis focuses on studying existing differentially private random forest (DPRF) algorithms, comparing them, and constructing a version of the DPRF algorithm based on these algorithms. Twelve articles from the late 2000s to 2022, each implementing a version of the DPRF algorithm, are included in the review of previous work. The created algorithm, called DPRF_thesis , uses a privatized median as a method for splitting internal nodes of the decision trees. The class counts of the leaf-nodes are made with the exponential mechanism. Tests on the DPRF_thesis algorithm were run on three binary classification UCI datasets, and the accuracy results were mostly comparable with the two existing DPRF algorithms DPRF_thesis was compared to. ACM Computing Classification System (CCS): Computing methodologies → Machine learning → Machine learning approaches → Classification and regression trees Security and privacy → Database and storage security → Data anonymization and sanitization

In the context of data science and machine learning, feature selection is a widely used technique that focuses on reducing the dimensionality of a dataset. It is commonly used to improve model accuracy by preventing data redundancy and over-fitting, but can also be beneficial in applications such as data compression. The majority of feature selection techniques rely on labelled data. In many real-world scenarios, however, data is only partially labelled and thus requires so-called semi-supervised techniques, which can utilise both labelled and unlabelled data. While unlabelled data is often obtainable in abundance, labelled datasets are smaller and potentially biased. This thesis presents a method called distribution matching, which offers a way to do feature selection in a semi-supervised setup. Distribution matching is a wrapper method, which trains models to select features that best affect model accuracy. It addresses the problem of biased labelled data directly by incorporating unlabelled data into a cost function which approximates expected loss on unseen data. In experiments, the method is shown to successfully minimise the expected loss transparently on a synthetic dataset. Additionally, a comparison with related methods is performed on a more complex EMNIST dataset.

Currency risk is an important yet neglected consideration for investors holding internationally diversified investment portfolios. The foreign exchange market is an extremely liquid and efficient market, with daily transaction volumes exceeding the equivalent of several trillion euros. International investors have to decide upon the level of exposure on various currency risks typically by hedging some or all of the underlying currency exposure with currency derivative contracts. The currency overlay refers to an approach where the aggregate currency exposure from the investment portfolio is managed with a separate derivatives strategy, aimed at improving the overall portfolio’s risk adjusted returns. In this thesis, we develop a novel systematic, data-driven approach to manage the currency risk of investors holding diversified bond-equity portfolios, accounting for both risk minimization and expected returns maximization objectives on the portfolio level. The model is based upon modern portfolio theory, leveraging findings from prior literature in covariance modelling and expected currency returns. The focus of this thesis is in ensuring efficient risk diversification through the use of accurate covariance estimates fed by high-frequency data on exchange rates, bonds and equity indexes. As to the expected returns estimate, we identify purchasing power parity (PPP) and carry signals as credible alternatives to improve the expected risk-adjusted returns of the strategy. A block bootstrap simulation methodology is used to conduct empirical tests on different specifications of the developed dynamic overlay model. We find that dynamic risk-minimizing strategies significantly decrease portfolio risk relative to either unhedged or fully hedged portfolios. Using high-freqency data based returns covariance estimates is likely to improve portfolio diversification relative to a simple daily data-based estimator. The empirical results are much less clear in terms of risk adjusted returns. We find tentative evidence that the tested dynamic strategies improve risk adjusted returns. Due to the limited data sample used in this study, the findings regarding expected returns are highly uncertain. Nevertheless, considering evidence from prior research covering much longer time-horizons, we expect that both the risk-minimizing and returns maximizing components of the developed model are likely to improve portfolio-level risk adjusted returns. We recommend using the developed model as an input to support the currency risk management decision for investors with globally diversified investment portfolios, along with other relevant considerations such as solvency or discretionary market views.

For machine learning, quantum computing provides effective new computation methods. The number of states a quantum register may express is exponential compared to the classical register of the same size, and this expressivity may be used in machine learning. It has been shown that in less than exponential time, a theoretical fault-tolerant quantum computer may perform computations that cannot be run on a classical computer in a feasible time. In machine learning, however, it has been shown that a classical machine learning method may learn a target model defined by an arbitrary quantum circuit if given a sufficient number of training data. In other words, a machine learning method that utilizes quantum computing may gain a quantum prediction advantage over its classical counterpart if the number of training data is low. However, this result does not address the noise of contemporary quantum computers. In this thesis, we use a simulation of a quantum circuit to test how a gradually increased noise affects the ability of a hybrid quantum-classical machine learning system to retain the quantum prediction advantage. With a simulated quantum circuit, we embed classical data rows into the quantum Hilbert space that acts as a feature space known from classical kernel theory. We project the data back to classical space, yielding a projected dataset that differs from the original. Using kernel matrices of the original and projected datasets, we then create adversarial binary labeling. With few training data, this adversarial labeling is easy for a classical neural network to learn using the projected features but impossible for using the original data. We show that this quantum prediction advantage diminishes as a function of the error rate introduced in the simulation of the data-embedding quantum circuit. Our results suggest the noise threshold for a feasible system lies slightly above the ones of contemporary hardware, indicating our experiment should be tested on actual quantum hardware. We derive a parameter optimization scheme for an arbitrary hardware implementation such that it may be concluded whether the quantum hardware in question may produce a quantum advantage dataset beyond the simulation capability of classical computers.

Large graphs often have labels only for a subset of nodes. Node classification is a semi-supervised learning task where unlabeled nodes are assigned labels utilizing the known information of the graph. In this thesis, three node classification methods are evaluated based on two metrics: computational speed and node classification accuracy. The three methods that are evaluated are label propagation, harmonic functions with Gaussian fields, and Graph Convolutional Neural Network (GCNN). Each method is tested on five citation networks of different sizes extracted from a large scientific publication graph, MAG240M-LSC. For each graph, the task is to predict the subject areas of scientific publications, e.g., cs.LG (Machine Learning). The motivation of the experiments is to give insight on whether the methods would be suitable for automatic labeling of scientific publications. The results show that label propagation and harmonic functions with Gaussian fields reach mediocre accuracy in the node classification task, while GCNN had a low accuracy. Label propagation was computationally slow compared to the other methods, whereas harmonic functions were exceptionally fast. Training of the GCNN took a long time compared to harmonic functions, but computational speed was acceptable. However, none of the methods reached a high enough classification accuracy to be utilized in automatic labeling of scientific publications.

Deep neural networks are widely used in natural language processing. Large language models, trained with large corpora, enable improved information extraction from data that is too large for human processing. This thesis reviews the performance of a deep learning natural language processing pipeline in detecting and removing (anonymising) personal information. Methods for fast and accurate ano- or pseudonymisation of data containing sensitive information are vital to research and development in science and industry, as legislation demands extensive procedures concerning handling of data with direct or indirect personal information. We propose a method that achieves state of the art results on noisy data, and good performance on a contemporary benchmark. Our comparison of anonymisation performance is one of the first for Finnish free texts.

This thesis examines components of an emerging topic: the interplay between immigration background and partnering in the modern Finnish context. It poses the question: how do various computational methods capture the ways that immigrant background status alters (1) the time to first union formation and (2) subsequent first union dissolutions in Finland from 1987-2020? Using longitudinal Finnish register data, the study focuses on all women residents in Finland observed from age 18 onwards during the specified period, categorized by their intergenerational immigration status. The study examines the relationship between immigration status and the two events of interest using the nonparametric Kaplan-Meier survivor function, semiparametric Cox Proportional Hazards model, and parametric survival model fitted with generalized gamma distribution. The strengths, limitations, and findings from each analytic method are compared. The results suggest three main findings: firstly, there is a clear gradient in the risk of first union formation and dissolution across women with different immigrant backgrounds in Finland, with Native Finnish women experiencing the highest risk, followed by 2.5 generation women (women with one Native Finnish parent and one 1st generation immigrant parent), 1st generation immigrant women, and finally 2nd generation women (women with two 1st generation immigrant parents). Secondly, factors including educational attainment, region of origin, rural/urban residence, and partnership homogamy based on region of origin contribute to differences in the risks for both union formation and union dissolution. Finally, despite the unique assumptions and constraints of each method, results remain consistent across all models, indicating that a variety of computational methods can provide robust insights into the complex interplay between immigration and first union dynamics in Finland. In light of the growing immigrant population and the potential influence of their first union dynamics on population change, these findings suggest alignment with segmented assimilation theory, highlighting a non-linear assimilation process influenced by socio-economic status and socio-cultural resources. The observed differences between the 2.5 and 2nd generations raise intriguing questions about the experiences of immigrant children in Finland. The 2nd generation's particularly low risk of first union formation indicates potentially unique acculturation stressors that warrant further investigation.

Ultrasonic guided lamb waves can be used to monitor structural conditions of pipes and other equipment in industry. An example is to detect accumulated precipitation on the surface of pipes in a non-destructive and non-invasive way. The propagation of Lamb waves in a pipe is influenced by the fouling on its surface, which makes the fouling detection possible. In addition, multiple helical propagation paths around pipe structure provides rich information that allows the spatial localization of the fouled area. Gaussian Processes (GP) are widely used tools for estimating unknown functions. In this thesis, we propose machine learning models for fouling detection and spatial localization of potential fouled pipes based on GPs. The research aims to develop a systematic machine learning approach for ultrasonic detection, interpret fouling observations from wave signals, as well as reconstruct fouling distribution maps from the observations. The lamb wave signals are generated in physics experiments. We developed a Gaussian Process Regression model as a detector, to determine whether each propagation path is going across the fouling or not, based on comparison with clean pipe. This binary classification can be regarded as one case of the different fouling observations. Latent variable Gaussian Process models are deployed to model the observations over the unknown fouling map. Then Hamiltonian Monte Carlo sampling is utilized to perform full Bayesian inference for the GP hyper-parameters. Thus, the fouling map can be reconstructed based on the estimated parameters. We investigate different latent variable GP models for different fouling observation cases. In this thesis, we present the first unsupervised machine learning methods for fouling detection and localization on the surface of pipe based on guided lamb waves. In these thesis we evaluate the performance of our methods with a collection of synthetic data. We also study the effect of noise on the localization accuracy.

Predicting patient deterioration in an Intensive Care Unit (ICU) effectively is a critical health care task serving patient health and resource allocation. At times, the task may be highly complex for a physician, yet high-stakes and time-critical decisions need to be made based on it. In this work, we investigate the ability of a set of machine learning models to algorithimically predict future occurrence of in hospital death based on Electronic Health Record (EHR) data of ICU-patients. For one, we will assess the generalizability of the models. We do this by evaluating the models on hospitals the data of which has not been considered when training the models. For another, we consider the case in which we have access to some EHR data for the patients treated at a hospital of interest. In this setting, we assess how EHR data from other hospitals can be used in the optimal way to improve the prediction accuracy. This study is important for the deployment and integration of such predictive models in practice, e.g., for real-time algorithmic deterioration prediction for clinical decision support. In order to address these questions, we use the eICU collaborative research database, which is a database containing EHRs of patients treated at a heterogeneous collection of hospitals in the United States. In this work, we use the patient demographics, vital signs and Glasgow coma score as the predictors. We devise and describe three computational experiments to test the generalization in different ways. The used models are the random forest, gradient boosted trees and long short-term memory network. In our first experiment concerning the generalization, we show that, with the chosen limited set of predictors, the models generalize reasonably across hospitals but that only a small data mismatch is observed. Moreover, with this setting, our second experiment shows that the model performance does not significantly improve when increasing the heterogeneity of the training set. Given these observations, our third experiment shows that