<p>QC 20110912</p> / MODyS

Physical storage is not enough to handle the amount of data created each day. As a result of the pandemic the need to create and share information over the Internet has increased. As a result, the importance of cloud services in our society increases, as they offer storage solutions that their customers do not have to maintain themselves. However, there are several security risks and threats, as for everything that is handled over the Internet. This study highlights these threats and describes the current countermeasures. A survey is conducted to examine private individuals' use of cloud services and the security functions offered. A comparative analysis found differences between the four largest cloud storage services through which security features are offered, the effectiveness of which is also compared with the survey results. The study found that both cloud service providers and their customers each have an important role to play in maintaining security. / Fysisk lagring är inte tillräckligt för att hantera den mängd data som skapasvarje dag. Till följd av pandemin har även behovet av att skapa och delainformation över Internet ökat. Därav ökar betydelsen av molntjänster i vårtsamhälle, eftersom de erbjuder lagringslösningar som deras kunder intesjälva behöver uppehålla. Däremot följer en mängd säkerhetsrisker och hot,likt allt som hanteras över Internet. Den här studien belyser dessa hot ochbeskriver åtgärder. Genom en enkät undersöks även privatpersonersanvändning av molntjänster samt säkerhetsfunktionerna som erbjuds. Enkomparativ analys fann olikheter mellan de fyra största molntjänsternagenom vilka säkerhetsfunktioner som erbjuds, vars effektivitet ävenjämfördes gentemot enkätresultaten. Studien fann att både leverantörerna avmolntjänsterna samt dess kunder har var sin viktig roll i att bibehållasäkerheten

The threat of violent extremism is considered by authorities as one of the largest today. Political extremism has increased over the years and the number of politically motivated terrorist incidents in the Western world is higher than religiously motivated. The role of digital platforms when it comes to recruitment and radicalization is widely debated. This study specifically examines TikTok which is one of the fastest growing and leading digital platforms. TikTok's recommendation system differs from other social media by being the main product for the platform. Experiments show how TikTok's algorithms recommend hateful content exponentially, and that the different types of hateful content that are examined have a high correlation to one another. In some cases, extremist content was also recommended. Furthermore, it is stated that TikTok does not take enough action when videos and/or profiles are being reported for hateful and/or extremist content. The literature study also showed that the Swedish Police Authority's handling of hate crimes on the Internet is insufficient despite pressure to improve from the government. / Hotet från våldsbejakande extremism betraktas av myndigheter som ett av de största idag. Politisk extremism har under åren ökat och antalet politiskt motiverade terrorincidenter i västvärlden är högre än religiöst motiverade. Rollen som digitala plattformar har när det kommer till rekrytering och radikalisering är väldigt omdiskuterad. Den här studien undersöker specifikt TikTok som är en av de ledande digitala plattformarna och snabbast växande. TikToks rekommendationssystem skiljer sig från andra sociala medier genom att vara den huvudsakliga produkten för plattformen. Genom experiment redovisas hur TikToks algoritmer rekommenderar hatiskt innehåll exponentiellt, och att de olika typerna av hatiskt innehåll som undersöks har hög korrelation med varandra. I vissa fall rekommenderades även extremistiskt innehåll. Vidare konstaterades att TikToks agerande vid anmälningar av videor och/eller profiler med hatiskt och/eller extremistiskt innehåll är bristfällig. Litteraturstudien visade även att polismyndighetens hantering av hatbrott på Internet är otillräcklig trots påtryckningar om förbättring av regeringen.

Networks are useful when modeling interactions in real-world systems based on relational data. Since networks often contain thousands or millions of nodes and links, analyzing and exploring them requires powerful visualizations. Presenting the network nodes in a map-like fashion provides a large scale overview of the data while also providing specific details. A suite of algorithms can compute an appropriate layout of all nodes for the visualization. However, these algorithms are computationally expensive when applied to large networks because they must repeatedly derive relations between every node and every other node, leading to quadratic scaling. Also, the available implementations compute the layout from the raw data instead of the network, making customization difficult. In this thesis, I introduce a modular algorithm that removes the need to consider all node pairs by approximating groups of pairwise relations. The groups are determined by clustering the network into densely connected groups of nodes with a community-detection algorithm. The implementation accepts a network as input and returns the layout coordinates, enabling modular and straightforward integration in a data analysis pipeline. The approximations improve the new algorithm's scaling to an order of 2N1.5 compared to the original N2. For a network with one million nodes, this scaling improvement gives a 500-fold performance boost such that a computation that previously took one week now completes in about 20 minutes.

Power generation and consumption are directly affected by weather, and the weather variables are fluctuating, which means that generation and consumption also fluctuate. These fluctuations can damage equipment or cause other problems in an electrical network. Having accurate predictions to adjust the infrastructure ahead of time can therefore mitigate these problems. Extensive research has been done on short-term forecasting using machine learning (ML) in electrical networks where some of them also use weather data. However, often the research studies do not cover many of the algorithms that are known to be suitable for short-term forecasting in an electrical network, or the importance of the features is not shown. Our study aims to compare a wide range of ML algorithms that are known to be good for short-term forecasting with time-series data. Additionally, we want to investigate which of the available features, including weather data, have the greatest impact on the results of the models. The research consists of a literature study to find the most suitable algorithms for time-series short-term forecasting in an electrical network using weather information. To find the best algorithms, several algorithms are trained and tested on the full data set. Then, the algorithms are retrained on the same data set after feature selection and compared to the models trained on the full data set. The results of our experiment are that BiLSTM and CNN are the best models and that the features most closely related to the target are the most important while the weather features still have some impact.

Today in Sweden there are thousands of sensors used to check the state of train vehicles to detect faults. Almost all these sensors get separate measurements for every axle on a train so if an error is detected its location is defined by an axle number within a train. This axle number needs to be matched with a certain vehicle to be able to easily locate it and remove or at least check the vehicle. It is, therefore, necessary to be able to break down trains into vehicles from timestamps readings. This project explores the possibilities of using machine learning to classify the train vehicles based on timestamp readings made by RFID detector setups.Throughout the project, several algorithms were attempted with different structures and different ways of using the timestamp data. In the end, the MLP-neural network structure was most promising and a model that could predict 91% of the trains correctly was created. This model showed that machinelearning was a promising way to classify vehicles from axle timestamp readings. The model also worked for some of the faulting sensors. It worked since it did not require the entire RFID detector setup to be fully functional, which was an unexpected extra positive outcome of the project

En grundläggande faktor till en effektiv och lyckad integration i samhället och arbetsmarknaden, är spraket. Genom att bryta den existerande språkbarriären kan både samhället och individen gynnas markant. Rapporten syftar således till att undersöka om språkmodeller kan effektivisera inlärningen av språk. Genom att utgå från ramverket BERT jämförs två olika språkmodellers prestanda avseende att klassificera en meningsföljd som korrekt respektive inkorrekt. Här används en grundmodell, baserad på den förtränade modellen KB-BERT, samt en finjusterad modell som tränats på dialoger från företaget Lingio. Resultaten tyder på att den finjusterade BERT modellen har god potential som sprakmodell och därmed kan appliceras på NLP-uppgifter hos applikationer ämnade för inlärning av språk. Denna slutsats stöds även av argument framförda under en intervju med Lingios CTO. Det krävs dock ytterligare undersökning för att kunna fastställa om BERT är applicerbar på fler uppgifter, såsom att förutsäga nästkommande ord samt möjligheten att generera svar utifrån en fråga. Arbetet år utvecklat tillsammans med Lingio, VINNOVA och KTH, och hoppas kunna bidra med underlag for vidareutveckling av deras projekt inom området. / One of the most essential factors for an effective and successful integration into society and the labor market is language. Both society and the individual can benefit significantly by breaking the existing language barrier that prevents integration. Therefore, the purpose of this report is to study if language models can shape language learning into becoming more effective. Based on the BERT framework, the performance of two different language models are compared, in regards to classifying a sequence of two sentences as correct or incorrect. This is done using a model from KB-BERT that has only been pre-trained, and a fine-tuned version that has been trained on dialogues from the company Lingio. The results indicate that the fine-tuned BERT model has a good potential to be used as a language model and is therefore applicable on NLP-tasks in applications for language learning. This conclusion is also supported by arguments which were stated during an interview with Lingio’s CTO. However, further investigation is required to determine whether BERT is applicable on other tasks, such as predicting the next word in a sentence or the ability to generate answers based on questions. The work has been developed together with Lingio, VINNOVA and KTH, and can hopefully act as a basis for further development of their project within the field.

Att identifiera finansiella svårigheter vid bedömning av ett företags ekonomiska situation är väsentligt för att kreditgivare ska undvika kreditförluster. En viktig del av kreditbedömningen är att analysera sannolikheten för att ett företag kommer gå i konkurs eller inte. Att identifiera en förhöjd konkursrisk ar därmed en faktor som kan hjälpa kreditgivare att fatta mer varsamma investeringsbeslut. Arbetet ämnar därför att undersöka hur väl fyra olika maskininlärningsalgoritmer kan predicera okad risk för konkurs utifrån finansiell bolagsdata. Modellerna som används är logistisk regression, Support Vector Machine, Decision Trees och Random Forest. Då datan var obalanserad där antalet icke-konkurser var överrepresenterad fick modellerna tränas och testas på flera olika fördelade dataset och de slutgiltiga resultaten bygger på ett dataset som är balanserat. Modellerna utvärderades med hjälp av en förväxlingsmatris och evalueringsmatten korrekthet, precision, täckning och F-score. Ju mer balanserad datan blev desto bättre blev resultaten men trots detta skiljde sig resultaten mellan modellerna. Studien visade att logistisk regression presterade sämst av samtliga modeller med ett F-score på 60%. Random Forest var den modell som hade bast prediktiv förmåga med ett F-score på 77%. Vid studerande av särdragen visade det sig bland annat att förändring i antalet anställda, soliditet och eget kapital har en förklaringsgrad till konkurs och är något som bör tas i beaktande vid kreditbedömning. Andra faktorer, såsom vilken industri ett företag tillhör, bör även det ha en betydelse vid kreditbedömning då olika branscher tenderar att ha fler konkurser än andra. / When evaluating a company’s financial situation it is essential to identify financial distress in order for creditors to avoid credit losses. An important part of credit assessment is to analyze the probability that a company will go bankrupt or not. Analyzing an increased risk of bankruptcy is thus a factor that can help lenders make more prudent investment decisions. Accordingly, this study aims to investigate how well four different machine learning algorithms can predict increased risk of bankruptcy based on financial company data. The models used are Logistic Regression, Support Vector Machine, Decision Trees and Random Forest. As the data was imbalanced where the number of non-bankruptcies was overrepresented, the models were trained on several different distributed datasets and the final results are based on a dataset that is balanced. The models were evaluated using a confusion matrix and the evaluation metrics accuracy, precision, recall and F-score. The more balanced the data was, the better the results were but despite this the results differed between the models. The study showed that logistic regression performed the worst of add models with an F-score of 60%. Random Forest was the model with the best predictive ability with an F-score of 77%. When investigating the features, change in number of employees, equity ratio and equity turned out to have a degree of explanation for bankruptcy and should be taken into account when assessing credit. Other factors, such as which industry a company belongs to, should also be a factor taken into account as some industries tend to have more bankruptcies than others.

The influencer marketing industry has seen a tremendous growth in recent years, yet the effectiveness of this marketing form is still largely unexplored. This report aims to explore how various performance measures are linked to the reach of social media pages, utilizing the linear regression model. Three different data sets were collected manually, or using web scraping. By splitting these data sets to training- and test data we examined the degree to which the linear regression model can predict the actual reach, the page views and the weekly growth of an influencer. We concluded that there is a statistically significant correlation between multiple performance metrics of a social media page and the actual reach or the page views of that account. This study is however limited by its narrow data set and time frame, warranting future research in order to further establish the degree of this correlation. The results of this study can benefit companies in their process of selecting influencers to collaborate with, as well as determining the expected return on investment for that particular collaboration. This can in turn lead to a more efficient, authentic and transparent marketplace, and to consumers being less exposed to advertisement from misleading and malicious influencers. / Under de senaste åren har marknadsföringsindustrin med influencers växt drastiskt, ändå är effektiviteten hos denna marknadsföringsform relativt outforskad. Denna rapport avser använda linjär regression för att utforska hur olika prestationsmått är kopplade till räckvidden hos profiler på sociala medier. De olika datamängderna samlades manuellt, eller med hjälp av web scraping. Genom att dela upp datamängderna i träningsdata och testdata undersökte vi i hur hög grad den linjära regressionsmodellen kan förutsäga faktisk räckvidd, sidvisningar och profilens tillväxt under en vecka.  Vi drog slutsatsen att det finns en statistisk signifikant korrelation mellan flera prestationsmått för en profilsida, och antalet sidvisningar for det kontot. Studien är emellertid begränsad av sin datamängd och tidsspann, något som motiverar framtida studier for att ytterligare etablera korrelationsgraden.  Studiens resultat kan gynna företag i deras process att välja vilka influencers de vill samarbeta med, såväl som i deras process att bestämma den förväntade avkastningen för ett specifikt samarbete. Detta kan i sin tur bidra till en mer effektiv, autentisk och transparent marknad, något som också gör att konsumenten ¨ blir mindre exponerad for marknadsföring från vilseledande och illvilliga influencers.

Detta arbete undersöker möjligheten att estimera intäkter för företag med hjälp av maskininlärning. Datan som modellerna utgår ifrån består av punkter från bolagens balansräkningar och annan offentlig data. Eftersom frågeställningen som arbetet utreder ar outforskad sedan tidigare ligger arbetets huvudsakliga fokus på att utforska vilka metoder som är mest lämpliga för uppgiften samt vilka särdrag i datasetet som har störst inverkan på modellerna. I arbetet utreds frågan med hjälp av fyra olika modeller; Random Forest regression, XGBoost, Minstakvadratmetoden och Lasso. Modellerna utvärderades med kvantitativa mättal såsom R2-varde och absoluta genomsnittliga procentuella felet (MAPE). Den algoritm och slutgiltiga modell som presterade bast utifrån dessa mått var Random Forest regression med genomsnittligt R2-score på 0,8197 och MAPE-score på 0.3864. Denna studie drar slutsatsen att ensemble metoder som XGBoost och Random Forest troligtvis ar mer lämpliga att använda för denna typ av studier i jämförelse med simplare regressionsmodeller såsom Minstakvadratmetoden och Lasso. Avslutningsvis dras slutsatsen att modellerna kan bidra till beslutsunderlaget vid utvärdering av bolag för vilka intäkterna är ok ända. / This work examines the possibility of estimating revenue for companies using machine learning. The data on which the models are based consists of points from the companies’ balance sheets and other public data. Since the research area is unexplored prior to this study, the main focus of this thesis is to explore which methods are most suitable for the task and which features in the dataset have the greatest impact on the models. In the study, the issue is investigated with the help of four different models; Random Forest regression, XGBoost, ordinary least squares method and Lasso. The models were evaluated with quantitative measures such as R2 score and mean absolute percentage error (MAPE). The algorithm and final model that performed best based on these measures were Random Forest regression with an average R2 score of 0,8197 and MAPE score of 0.3864. This study concludes that ensemble methods such as XGBoost and Random Forest are probably more suitable to use for this type of study compared to simpler regression models such as least squares method and Lasso. In conclusion, the models can contribute to the initial financial analysis of companies for which the income is unknown.