Browsing by master's degree program "Datatieteen maisteriohjelma"

Official Statistics (OS) are crucial in facilitating informed and reliable decision-making. However, while the demand for diverse and precise information surges, challenges in obtaining accurate data emerge. The declining response rates to statistical surveys and escalating data collection costs further exacerbate the situation, particularly in surveys measuring rare events. This thesis explores the application of the Bayesian framework to statistical production. The concept of OS and the fundamental principles that guide their production are introduced. The suitability of the Bayesian approach for OS production is assessed from theoretical, philosophical, and practical standpoints. The core of statistical inference is explored, and the differences between Bayesian and frequentist approaches are compared. General tools for Bayesian inference and their practical utilization are presented, focusing especially on the graphical representation of a probabilistic model. Furthermore, a progressive construction of the proposed baseline model for analyzing Recreational Fishing Survey data is illustrated, with attention given to the issue of selection bias. The Bayesian Finnish Recreational Fishing Statistics 2020, with concise content produced through the developed model and the genuine data collected for OS, is also launched. While the thesis underscores that the proposed model should be regarded as a basis for further development, the results indicate that reasonable assessments can be obtained even with a simple Bayesian model. Overall, this thesis emphasizes the importance of adopting Bayesian thinking in statistical analysis to enhance knowledge-driven policy-making and adapt to the evolving information needs.

The aim of the research presented in this dissertation is to construct a model for personalised item recommendations in an online setting using a reinforcement learning approach, specifically Thompson sampling, which is part of the family of multi-armed bandit algorithms. Moreover the setting involves an online shopfront where arriving customers get viewed with the recommended item, and make purchasing decisions of them. The recommendations are conducted by the multi- armed bandit algorithm which "plays" different arms, represented by the items, while learning, exploring and exploiting the underlying distributions of the data that is obtained. Thompson sampling and the theory behind it is introduced thoroughly and comparison against other bandit algorithms as well as a multinomial logistic regression model is conducted both on real-life data collected over time from an online environment and a dummy data set. The experiments focus on the applicability of bandits in the setting, dealing with challenges that a bandit algorithm may face and the strengths they have over more traditional and well known models such as the logistic regression model in the setting at hand.

DP-SGD (Differentially Private Stochastic Gradient Descent) is the gold standard approach for implementing privacy in deep learning settings. DP-SGD achieves this by clipping the gradients during training and then injecting noise. The algorithm aims to limit the impact of any data point from the training dataset on the model. Therefore, the gradient of an individual sample will give information until a certain point, thus limiting the chances of an inference attack discovering the data used for the model training. While DP-SGD ensures the privacy of the model, there is no free lunch, and it has its downsides in terms of the utility-privacy trade-off and an increase in computational resources for training. This thesis aims to evaluate different DP-SGD implementations in terms of performance and computational efficiency. We will compare the use of optimized clipping algorithms, different GPU processing unit architectures, speed-up by compilation, lower precision in the data representation, and distributed training. These strategies effectively reduce the computational cost of adding privacy to the deep learning training compared to the non-private baseline.

Credit rating is one of the core tools for risk management within financial firms. Ratings are usually provided by specialized agencies which perform an overall study and diagnosis on a given firm’s financial health. Dealing with unrated entities is a common problem, as several risk models rely on the ratings’ completeness, and agencies can not realistically rate every existing company. To solve this, credit rating prediction has been widely studied in academia. However, research in this topic tends to separate models amongst the different rating agencies due to the difference in both rating scales and composition. This work uses transfer learning, via label adaptation, to increase the number of samples for feature selection, and appends these adapted labels as an additional feature to improve the predictive power and stability of previously proposed methods. Accuracy on exact label prediction was improved from 0.30, in traditional models, up to 0.33 in the transfer learning setting. Furthermore, when measuring accuracy with a tolerance of 3 grade notches, accuracy increased almost 0.10, from 0.87 to 0.96. Overall, transfer learning displayed better out-of-sample generalization.

In bioinformatics, new genomes are sequenced at an increasing rate. To utilize this data in various bioinformatics problems, it must be annotated first. Genome annotation is a computational problem that has traditionally been approached by using statistical methods such as the Hidden Markov model (HMM). However, implementing these methods is often time-consuming and requires domain knowledge. Neural network-based approaches have also been developed for the task, but they typically require a large amount of pre-labeled data. Genomes and natural language share many properties, not least the fact that they both consist of letters. Genomes also have their own grammar, semantics, and context-based meanings, just like phrases in the natural language. These similarities give motivation to the use of Natural language processing (NLP) techniques in genome annotation. In recent years, pre-trained Transformer neural networks have been widely used in NLP. This thesis shows that due to the linguistic properties of genomic data, Transformer network architecture is also suitable for gene predicting. The model used in the experiments, DNABERT, is pre-trained using the full human genome. Using task-specific labeled data sets, the model is then trained to classify DNA sequences into genes and non-genes. The main fine-tuning dataset is the genome of the Escherichia coli bacterium, but preliminary experiments are also performed on human chromosome data. The fine-tuned models are evaluated for accuracy, F1-score and Matthews correlation coefficient (MCC). A customized estimation method is developed, in which the predictions are compared to ground-truth labels at the nucleotide level. Based on that, the best models achieve a 90.15% accuracy and an MCC value of 0.4683 using the Escherichia coli dataset. The model correctly classifies even the minority label, and the execution times are measured in minutes rather than hours. These suggest that the NLP-based Transformer network is a powerful tool for learning the characteristics of gene and non-gene sequences.

The objective of this thesis was to explore the article corpus of a domain-specific Finnish-language newspaper to generate a new content tag word set to replace an existing one of poor quality. The articles used as the dataset in this study had previously human-assigned content tags, but in absence of a proper tagging strategy and guidelines, the assigned tags were found to include too much variation to be useful, e.g., as search terms. No supervised learning models were used, since there was no good quality training material available specific to the topics of the data. Instead, we experimented in generating new tagsets with two unsupervised methods alongside a few variations based on nouns and proper nouns in the text content of the articles. A proper tag set would be useful in tagging future articles automatically or in drafting guidelines for manual tagging of future articles by the journalists. A limited survey among subject matter experts and other responders was conducted in order to evaluate the results generated by the methods. In general, the results were not encouraging, with the most basic model TF-IDF clearly performing better than the other models across all responders. Further examination after topic modeling using Latent Dirichlet Allocation (LDA) revealed, that somewhat better scores could be found among some topics. A further manual task of assessing and naming of the topics followed by a new tagging effort within topics was suggested as a next step in order to overcome the deficiencies in the presented methods.

With the growing concerns over data privacy and new regulations like the General Data Protection Regulation (GDPR), there has been increased attention to privacy-preserving synthetic data generation methods. However, the usability of these methods has received limited attention. This thesis focuses on the usability challenges associated with privacy-preserving synthetic data generation methods based on probabilistic graphical models. This thesis addresses usability challenges related to running time efficiency, applicability with continuous data and query selection with experiments conducted on different datasets. This thesis aims to bridge the gap between cutting-edge privacy-preserving synthetic data generation methods and their practical implementation with real-world datasets. Proposed solutions have the potential to make these methods more accessible and usable, thereby facilitating their broader adoption. The thesis concludes by summarizing the key results and emphasizing the importance of addressing usability challenges in privacy-preserving synthetic data generation methods.

Valeuutiset ovat viime vuosina nousseet merkittäväksi yhteiskunnallisen keskustelun aiheeksi niin Suomessa kuin ulkomaillakin. Esimerkiksi vuoden 2016 yhdysvaltojen presidentinvaalien aikana jotkin valeuutiset levisivät laajemmalle kuin suosituimmat valtamediauutiset, ja valeuutisten onkin arveltu vaikuttaneen merkittävästi Trumpin voittoon kyseisissä vaaleissa. Aiemmasta suomalaisesta tutkimuksesta tiedetään, että Suomessa valeuutiset eivät aina sisällä suoraan virheellistä tietoa, ja tämän vuoksi suomalaisia valemedioita kutsutaan myös vastamedioiksi. Tiedetään myös, että suomalaisissa vastamediauutisissa kehystetään usein valtamedian uutisia tukemaan vastamedian omaa agendaa. Kehystämisellä tarkoitetaan viestinnän tutkimuksessa prosessia, jolla valikoinnin, poissulkemisen ja esimerkiksi metaforien ja iskulauseiden avulla muokataan mediaesityksen tulkintaa. Kehyksen käsite sekä kehysanalyysi ovat saaneet alkunsa sosiaalipsykologiasta ja levinneet sittemmin mediatutkimukseen. Laskennallisesti kehysanalyysiä on tehty sekä ohjatuilla että ohjaamattomilla koneoppimismenetelmillä, mutta yksikään näistä menetelmistä ei ole vakiintunut kehyksen operationalisoinnin monikäsitteisyyden vuoksi. Tämän tutkielman tarkoituksena on selvittää, millaisilla prosesseilla suomalainen vastamedia uudelleenkehystää valtamedian uutisia, sekä soveltaa ohjattua koneoppimista eri kehystämisen tapojen tunnistamiseen. Tutkimuskysymyksiin vastaamiseksi kerättiin kattava aineisto eräästä suomalaisesta vastamediasta, ja eroteltiin aineistosta valtamedialinkin sisältävät artikkelit. Tämän jälkeen identifioitiin laadullisesti kolme tapaa jolla vastamedia kehystää valtamedian uutisia: kritisoimalla valtamediaa, kopioimalla sisältöä sekä hyödyntämällä valtamedialähdettä argumentoinnin välineenä. Tässä tutkielmassa rakennetaan ohjattu koneoppimismalli kolmen edellä luetellun kehystämisen prosessin identifiointiin. Malli rakennettiin luokittelemalla 1000 artikkelin satunnaisotos valtamedialähteen sisältävästä aineistosta kolmeen edellä lueteltuun kehystämisen prosessin kategoriaan. Tämän jälkeen luokitellusta datasta eristettiin erilaisia piirteitä ja rakennettiin näiden pohjalta luokittelija. Työssä vertailtiin erilaisia satunnaismetsäluokittelijoita sekä tukivektorikoneita, joista eräs satunnaismetsäluokittelija suoriutui luokittelutehtävästi parhaiten. Luokittelijaa ei kuitenkaan voida pitää tarpeeksi tarkkana useimpiin käytännön hyvin korkeaa tarkkuutta vaativiin sovelluksiin. Luokittelijan merkittävimpinä pitämistä piirteistä saadaan kuitenkin uutta tietoa siitä, miten eri sanoja ja tekstin muotoilutyylillisiä keinoja käytetään eri kehystämistavoissa. Esimerkiksi artikkeleissa käytettyjen linkkien määrä sekä alaotsikkojen määrä nousivat luokittelijalle merkittävimpien piirteiden joukkoon. Tuloksista voidaan päätellä, että laskennallisessa mediatutkimuksessa sanojen lisäksi on hyödyllistä eristää myös artikkeliin liittyvää muotoiludataa. Toinen keskeinen tulos on, että ohjattua koneoppimista voidaan hyödyntää erilaisten median lähteeseen suuntautuvien orientaatioiden tunnistamiseen.

Electronic Health Records (ERH) data is one datatype that is continuously produced by hospitals all over the world. There is multiple different applications for this data; one of them is MIMIC dataset which applications are focused on machine learning, however, there is lack of visualisation of this dataset. Topic of this thesis is visualising ERH and more specifically the MIMIC dataset to help medical students gather insights into individual patient data. Considerations implementing this dataset to an individual patient visualisation are also examined. In thesis I first familiarized myself with interactive visualisation design principles with Munzners method. In the first stage I built a preliminary visualisation and gathered feedback with user inter- views. At the second stage I built a final visualisation and gathered final feedback and insights with user interviews. Visualisation consists of a dot graph for medication, line graph for patient response and table with diagnoses. The insights were analysed with Framework Method for qualitative data analysis. From the feedback for the visualisation and the insights gathered it is concluded that there are some issues that need to be taken in account when considering this dataset for individual patient visualisations. Insights from this visualisation were mostly factual and this type of visualisation seems to be slightly too challenging to medical students for more deep insights. Medication graph part of the final visualisation was seen as possible addition to current patient visualisations used in hospitals and the idea of using this visualisation as a teaching tool could only be considered with further additions and a different dataset.