Browsing by Title

Taloudellisten, institutionaalisten ja teknologisten ympäristöjen kiihtyvä muutostahti on luonut tarpeen tehdä oikeita valintoja menestymisen ja kehityksen takeeksi. Parhaimman valinnan tekee henkilö, jolla on eniten tietoa ja varmuutta tiedon paikkansapitävyydestä. Vaihtoehtoisesti päätöksenteon epävarmuustekijöitä voidaan hallita eliminointimenetelmillä, joiden hyödyntäminen voi myös johtaa parempiin päätöksiin. Epävarmuuden minimoimisen edellytyksenä on niin ikään pohjatietojen parantaminen. Tästä tarpeesta ovat nousseet tiedonlouhintamenetelmät. Tiedonlouhintamenetelmiä on kehittynyt valtava määrä vastaamaan kysynnän luomia tarpeita. Päätöksentekopuu on eräs tällainen analysointimenetelmä ja päätöksentekopuun pohjalta on luotu koneoppimismenetelmä satunnaismetsä. Satunnaismetsä on tutkimusten mukaan tällä hetkellä paras saatavilla oleva luokittelumenetelmä ja valikoitunut tämän tutkielman aiheeksi. Luvussa 2 luomme pohjaa satunnaismetsä-menetelmän ymmärtämiseksi. Lähdemme liikkeelle koneoppimisesta ja datalouhintamenetelmistä, joilla alustamme päätöksentekopuutyökalun. Käy ilmi, että on olemassa luokittelu-, regressio- ja luokitteluregressiopuita, ja että tässä tutkielmassa keskitymme luokittelupuihin. Tämän jälkeen esittelemme päätöksentekopuun metodologiaa. Luvussa 3 esittelemme tutkielman kannalta päätöksentekopuiden tärkeimmät validointimenetelmät, sillä datalouhinnassa analysointimenetelmien validoiminen on yhtä tärkeää kuin itse analysoiminen. Esittelemme mallien validointiin liittyviä käsitteitä, kuten tarkkuus, yleistysvirhe ja ylisovittuminen. Käymme läpi yleisimpiä tapoja validoida malleja, sekä näytämme esimerkkien kautta työkalut, joita käytämme tutkielmassa. Näitä ovat tarkkuus, väärinluokittelumatriisi, ROC-, kumulatiivinen saanti- ja nostokäyrä. Luvussa 4 esittelemme satunnaismetsä-koneoppimismenetelmän. Käymme ensiksi läpi joukko-oppimisen metodologiaa, jonka jälkeen käsittelemme satunnaismetsän algoritmin. Osoitamme teoreettisesti, miksi satunnaismetsä on parempi luokittelija kuin esimerkiksi päätöksentekopuu näyttämällä, että satunnaismetsän puumäärän kasvaessa satunnaismetsän yleistysvirhe suppenee kohti nollaa. Analysoimme teoreettisesti satunnaismetsän hyötyjä ja haittoja. Satunnaismetsän hyötyjä ovat sen tarkkuus, nopeus, ymmärrettävyys, toimivuus valtavilla datamäärillä, sekä kykeneväisyys analysoida tietojoukon merkittävimpiä muuttujia. Haittoja ovat, että satunnaismetsä ei suoriudu yhtä hyvin regressio-ongelmissa kuin esimerkiksi logistinen regressiomalli, sekä huono sovellettavuus pieniin tietojoukkoihin. Luvussa 5 sovellamme opittuja taitoja suppeaan tietojoukkoon. Tarkoituksenamme on arvioida RStudion ja SAS Enterprise Minerin satunnaismetsä-pakettien toimivuutta tunnetulla syötejoukolla. Analysoimme satunnaismetsäin suoriutumista luokittelutehtävässä ja vertailemme tuloksia päätöksentekopuuhun ja regressiomalliin. Hyödynnämme luvussa 3 opittuja validointimenetelmiä. Käy ilmi, että RStudion ja SAS Enterprise Minerin satunnaismetsä-paketit toimivat hyvin, ja että satunnaismetsä suoriutuu pienenkin tietojoukon luokittelussa malleista parhaiten. Luvussa 6 sovellamme satunnaismetsää yrityksen tarjoamaan haasteeseen, jossa tarkoitus on selittää ja ennustaa uusien asiakkaiden tietyn asiakassegmentin asiakasvaihtuvuutta. Käytämme yhtiön tarjoamia tietokantoja ja SAS Enterprise Miner-työkalua. Suoritamme vertailun satunnaismetsä- ja päätöksentekopuu-mallien välillä käyttämällä luvussa 3 esiteltyjä validointimenetelmiä ja analysoimme tulokset. Käy ilmi, että ROC-käyrien ja tarkkuuden perusteella satunnaismetsä suoriutuu sekä luokittelussa että ennustamisessa paremmin kuin päätöksentekopuu. Luvussa 7 pohdimme, millaisissa puitteissa satunnaismetsä soveltuu yrityksen liiketoimintaprosessiin. Käymme läpi vaatimuksia, joita satunnaismetsän soveltaminen asettaa, sekä mitä lisäarvoa satunnaismetsä menetelmänä tuo yritykselle. Tulos on, että satunnaismetsä soveltuu hyvin yrityksille, jotka hyödyntävät SAS-työkaluja ja tuo lisäarvoa analysointitehtäviin olemalla ymmärrettävä malli, mutta kuitenkin monipuolinen, nopea ja tarkka.

The number of IoT and sensor devices is expected to reach 25 billion by 2030. Many IoT appli- cations, such as connected vehicle and smart factory that require high availability, scalability, low latency, and security have appeared in the world. There have been many attempts to use cloud computing for IoT applications, but the mentioned requirements cannot be ensured in cloud environments. To solve this problem, edge computing has appeared in the world. In edge environments, containerization technology is useful to deploy apps with limited resources. In this thesis, two types of high available Kubernetes architecture (2 nodes with an external DB and 3 nodes with embedded DB) were surveyed and implemented using K3s distribution that is suitable for edges. By having a few experiments with the implemented K3s clusters, this thesis shows that the K3s clusters can provide high availability and scalability. We discuss the limitations of the implementations and provide possible solutions too. In addition, we provide the resource usages of each cluster in terms of CPU, RAM, and disk. Both clusters need only less than 10% CPU and about 500MB RAM on average. However, we could see that the 3 nodes cluster with embedded DB uses more resources than the 2 nodes + external DB cluster when changing the status of clusters. Finally, we show that the implemented K3s clusters are suitable for many IoT applications such as connected vehicle and smart factory. If an application that needs high availability and scalability has to be deployed in edge environments, the K3s clusters can provide good solutions to achieve the goals of the applications. The 2 nodes + external DB cluster is suitable for the applications where the amount of data fluctuate often, or where there is a stable connection with the external DB. On the other hand, the 3 nodes cluster will be suitable for the applications that need high availability of the database even in poor internet connection. ACM Computing Classification System (CCS) Computer systems organization → Embedded and cyber-physical systems Human-centered computing → Ubiquitous and mobile computing

Traditional natural language processing has been shown to have excessive reliance on human-annotated corpora. However, the recent successes of machine translation and speech recognition, ascribed to the effective use of the increasingly availability of web-scale data in the wild, has given momentum to a re-surging interest in attempting to model natural language with simple statistical models, such as the n-gram model, that are easily scaled. Indeed, words and word combinations provide all the representational machinery one needs for solving many natural language tasks. The degree of semantic similarity between two words is a function of the similarity of the linguistic contexts in which they appear. Word representations are mathematical objects, often vectors, that capture syntactic and semantic properties of a word. This results in words that are semantic cognates having similar word representations, an important property that we will widely use. We claim that word representations provide a superb framework for unsupervised learning on unlabelled data by compactly representing the distributional properties of words. The current state-of-the-art word representation adopts the skip-gram model to train shallow neural networks and presents negative sampling, an idea borrowed from Noise Contrastive Estimation, as an efficient method of inducing embeddings. An alternative approach contends that the inherent multi-contextual nature of words entails a more Canonical Correlation Analysis-like approach for best results. In this thesis we develop the first fully Bayesian model to induce word embeddings. The prominent contributions of this thesis are: 1. A crystallisation of the best practices from previous literature on word embeddings and matrix factorisation into a single hierarchical Bayesian model. 2. A scalable matrix factorisation technique for structured sparse data. 3. Representation of the latent dimensions as continuous Gaussian densities instead of as point estimates. We analyse a corpus of 170 million tokens and learn for each word form a vectorial representation based on the 8 surrounding context words with a negative sampling rate of 2 per token. We would like to stress that while we certainly hope to beat the state-of-the-art, our primary goal is to develop a stochastic and scalable Bayesian model. We evaluate the quality of the word embeddings against the word analogy tasks as well as other such tasks as word similarity and chunking. We demonstrate competitive performance on standard benchmarks.

This Master’s thesis studies the mechanisms connected to negative changes in educational outcomes in upper comprehensive schools in Helsinki. What are the factors associated with negative changes in educational outcomes of individual students during the transition from 7th to 9th grade? There is an increased socioeconomic and ethnic segregation in Helsinki Metropolitan Area, and the differences between schools’ levels of success have also been growing throughout the 21st century. There is little research on combining schools and city development in Finland. The aim is to examine is there an association between decreasing individual educational outcomes and socio-spatial or school segregation, and to look at what is the role of individual factors and social context in decreased educational outcomes. Studying pupils and schools is a good way to capture local processes of differentiation and neighbourhood effect, because children and youth are especially prone to neighbourhood and school effects due to their ongoing process of socialization, localized lives in their neighbourhood and shared institutions, such as school. This study is conducted quantitatively, and the main method in this study is hierarchical linear regression. The data is from Metropolitan Longitudinal Finland research, which studies the success and wellbeing of pupils in upper comprehensive schools in the Helsinki Metropolitan area. The study was conducted during the Fall of 2011 and the Spring of 2014 tracking the same cohort when the pupils were in their 7th and 9th grades. The results suggest that there are no differences found between schools, but some of the qualities describing neighborhoods indicate that some neighbourhood effect might be found. There are indications that pupils with decreased educational outcomes are more likely to study in schools that are located in low income areas than higher income areas. Also, for pupils with decreased educational outcomes, attending a school that is located in Northern or Southeastern Great districts is more likely than attending a school in Eastern Great district. Based on the results, pupils with negative change in educational outcomes are more likely to spend time with friends of own area than with school friends. Boys have a bigger risk for a negative change in educational outcomes than girls, and the change of school is connected to decreased educational outcomes. Mother’s education and immigration background was not found to have connection with decreased educational outcomes. Decreased educational outcomes have a connection with a low parents’ pedagogical ethos, but no connection with peers’ pedagogical ethos was found. The results are significant from the perspective of urban and educational politics and planning. The indications that the educational outcomes in upper comprehensive schools in Helsinki are differentiated in neighborhood level for example between Great districts, and in individual level between genders, challenge the goals of equal educational opportunities. Also, urban planning should be targeted to prevent socio-spatial differentiation of neighborhoods, in order to combat differentiation in schools’ composition of pupils. In future research, the starting level of educational success could be studied more closely- does decrease in educational outcomes implicate different educational paths for pupils that start with high starting level than pupils that have lower starting level in the beginning? This study provided information that there are no differences between schools found currently, but the processes of differentiation are not stable, so the processes should be observed continuously.

Compression methods are widely used in modern computing. With the amount of data stored and transferred by database systems constantly increasing, the implementation of compression methods into database systems has been studied from different angles during the past four decades. This thesis studies the scientific methods used in relational database research. The goal of the thesis is to evaluate the methods employed and to gain an understanding into how research into the subject should be conducted. A literature review is conducted. 14 papers are identified for review and their methodology is described and analysed. The papers reviewed are used to answer four research question and classified according to insights gained during the review process. There are similarities in methods of different papers that can be described to use as a starting point for conducting research in the field of database compression.

Strongly coupled matter called quark–gluon plasma (QGP) is formed in heavy-ion collisions at RHIC [1, 2] and the LHC [3, 4]. The expansion of this matter, caused by pressure gradients, is known to be hydrodynamic expansion. The computations show that the expanding QGP has a small shear viscosity to entropy density ratio (η/s), close to the known lower bound 1/4π [5]. In such a medium one expects that jets passing through the medium would create Mach cones. Many experimental trials have been done but no proper evidence of the Mach cone has been found [6, 7]. Mach cones were thought to cause double-bumps in azimuthal correlations. However these were later shown to be the third flow harmonic. In this thesis a new method is proposed for finding the Mach cone with so called event-engineering. The higher order flow harmonics and their linear response are known to be sensitive to the medium properties [8]. Hence a Mach cone produced by high momentum jet would change the system properties and, thus, the observable yields. Different flow observables are then studied by selecting high energy jet events with different momentum ranges. Considered observables for different momenta are then compared to the all events. Found differences in the flow harmonics and their correlations for different jet momenta are reported showing evidence of Mach cone formation in the heavy-ion collisions. The observations for different jet momenta are then quantified with χ 2 -test to see the sensitivities of different observables to momentum selections.

This thesis studies the application of language models to improve search in an online shop specialising in wholesale builder–trade product and service sales. The first aim was to determine if a Finnish language model could capture the meaning behind the search query words to improve the match between the queries and the product descriptions. Secondly, it was investigated if it was possible to train the model to recognise what products the users wanted to find with the search terms they used. Finally, it was investigated if it was possible to use the model for search ranking. Three models were trained using FinBERT as a model checkpoint and domain-specific product and clickthrough data for fine-tuning the models. The first two models were trained to classify online store products into product categories. The task was completed with a 0.98 F measure score for the model with 55 target categories and a 0.81 F measure score for the model with 762 target categories. The third model was trained to evaluate the relevance probabilities of search query-product pairs. The model generally determined more products as relevant than the current search engine solution. The F measure score for the model was 0.90, and in qualitative evaluation, the predictions made by the model made semantically sense. The restrictions for the practical use of the third model for search ranking come from the prediction inference needing to be faster to make search ranking predictions for many products.

Atmospheric secondary organic aerosol formation is not completely understood, aerosol processes and chemistry are complex as well as their interactions and evolution throughout the atmosphere. In this study the potential aerosol mass chamber was used to simulate these processes. Potential aerosol mass being the maximum mass that can be produced from precursor gases being oxidised. UV lamps situated inside the chamber produce oxidants (O3, OH, HO2) at the same ratio as the atmosphere and the lamp intensity can be changed to simulate different days of photochemical aging. Laboratory characterisation measurements were undertaken by adjusting various experimental parameters (level of oxidants, relative humidity, temperature, type of precursor and its concentration). The aim of these measurements is to understand how the PAM chamber simulates atmospheric aerosol aging and photochemical processes. The PAM chamber was observed to simulate atmospheric SOA formation well. As the amount of oxidants was increased the amount of formed SOA increased initially. After a certain point, it was observed that fragmentation started to decrease the mass even though the oxidation level of SOA increased further.

Connected devices of IoT platforms are known to produce, process and exchange vast amounts of data, most of it sensitive or personal, that need to be protected. However, achieving minimal data protection requirements such as confidentiality, integrity, availability and non-repudiation in IoT platforms is a non-trivial issue. For one reason, the trillions of interacting devices provide larger attack surfaces. Secondly, high levels of personal and private data sharing in this ubiquitous and heterogeneous environment require more stringent protection. Additionally, whilst interoperability fuels innovation through cross-platform data flow, data ownership is a concern. This calls for categorizing data and providing different levels of access control to users known as global and local scopes. These issues present new and unique security considerations in IoT products and services that need to be addressed to enable wide adoption of the IoT paradigm. This thesis presents a security and privacy framework for the Web of IoT platforms that addresses end-to-end security and privacy needs of the platforms. It categorizes platforms’ resources into different levels of security requirements and provides appropriate access control mechanisms.

Eutrophication and harmful substances of anthropogenic origin threaten the state of the Baltic Sea and especially its geochemistry and oxygen levels near the seafloor. Water exchange between the Baltic Sea and the Atlantic Ocean can affect oxygen circulation and sedimentation rates, but they are considered very sporadic and it is unclear how the water circulation and flow rates affect element concentrations and sediment deposition in the near seafloor environments. One of the less studied basins is the Western Gulf of Finland and its seafloor environment. During the 2019 voyage, the seafloor located to the south of the city of Hanko was investigated through bathymetric sounding tools and other measurements in which element concentration and sediment deposition rate data was acquired. The sounding revealed a large channel cutting the seafloor which was hypothesized to influence the nearbottom conditions. The obtained data consisted of samples from 13 short, 40 cm sediment cores which were analysed for 137Cs activity, organic content, and grain size distribution. The goal of the thesis was to determine the intensity of water exchange taking place in the seabed channels between the mid-Baltic Sea and the Western Gulf of Finland and investigate the effect of the seafloor channel and flow rates on sediment and element deposition, their relationships, and how they affect the overall conditions in the study area. These relationships were analyzed through spatial and statistical methods by utilizing GIS-tools to interpolate the data obtained from the study locations by using the Inverse Distance Weighting (IDW) method, and by multielement analyses in the R-environment, namely Principal Component Analysis (PCA) and Partial Least Regression (PLS) to analyze grain size and element concentration correlations and combine them with obtained flow rate data. The results showed strong correlation in flow rate intensities between the Western Gulf of Finland and the mid-Baltic Sea, and they are strongly linked with sedimentation and element deposition rates. However, no long-term trend was identified for the seafloor channel velocity frequencies. The Cs-activity shows stronger sedimentation activity on the western side of the seafloor channel. The overall element and sediment deposition in the study area was largely controlled by monthly and seasonal current velocity fluctuations among other processes. The element concentration comparison showed weakened oxygen conditions in the study area with increased eutrophication and carbon burial since the 1950s. The principal Component Analysis showed smaller grain sizes (0.15 - 2 mm) having a stronger influence on the datasets with Mo, N, and C providing largest variation in the data. Interpolation showed oxygen, pH, and H2S to have more fluctuation in the study area, which can indicate changes in the vertical gradients in each sample point. It could also be determined that other measured concentrations, such as temperature, turbidity, and salinity do not respond very sensitively to water inflow fluctuations or sedimentation rate changes. The results indicate that harmful substances and eutrophication are most likely going to increase in the near-bottom environment in the Western Gulf of Finland, contributed by anthropogenic activity. Water exchange is likely to become more and more uneven, thus affecting the flow rate effects to the sediment deposition in the Baltic Sea. Further studies are needed to link these processes to large-scale global changes and the general state of changes happening in the Baltic Sea and its surrounding areas. The seafloor of the Western Gulf of Finland could also be studied further to gain a better understanding of longer timescale changes on the seafloor channel currents, and element and sediment deposition rates.