Master Theses

Master Theses


Master Theses

/ Current and completed projects


Alsvåg, S.,  Addressing the Next-Poster Problem: A Hybrid Recommender System for Streaming Platforms, together with TV 2, UiB, Bergen 2024. Supervisor: Full. Prof. Dr. Christoph Trattner

Recommendation strategies in the Movie domain is varied, but has been shown to aid users in finding content they like. On video streaming-platforms such as TV 2 Play, the user is exposed to several different arenas where they can find something they would like to watch, be that the landing-page of the streaming site, or a suggestion for something more to watch after they concluded a movie or series. These are all areas where Recommender strategies can recommend something based on either the preferences of the user, or in the case of the concluded movie or series, something more to watch based on what they just watched. This latter aspects, being what I refer to as the next-poster problem in this thesis, is not a largely explored area of research, where previous actors have simply utilized the already established Collaborative Filtering (CF) model concerned with the user’s preferences without considering what the user just watched. Here I show that a solution to the next-poster problem is to combine the CF model with a Sequence Aware approach based on Markov Chains, finding an increase in implied user satisfaction over the baseline CF approach. Through an online evaluation on the streaming platform TV 2 Play, I show that using a Hybrid approach to solve the next-poster problem rather than a traditional CF model leads to a lessening in user engagement such as CTR, but an increase in the clicks resulting in a user actually watching the content, this being our implied user satisfaction. Further as a result of this online evaluation, I am able to show that its possible to find the best configuration for a Hybrid model based on Sequence Aware and CF approaches deployed in a real life scenario, through offline evaluation. The results allows me to showcase the importance of considering Sequence of items when recommending for the next-poster problem, and to show that an offline evaluation can imply results in a real world scenario, when considering the Movie domain. Although an improvement, this thesis also shows that there are many more avenues to consider for the next-poster problem.

Bergh S.,  Personalized Recommendations of Upcoming Sport Events, together with TV 2, UiB, Bergen 2023. Supervisor: Mehdi Elahi.

Recommender systems have emerged as essential tools for enhancing user engagement and content discovery in various domains, including the sports industry. In the context of sports viewing, personalized recommendations have become increasingly significant, enabling users to easily connect with their favorite sports teams, explore new content, and broaden their viewing preferences. Collaborative filtering (CF) stands out as a popular recommendation algorithm that analyzes the similarities and patterns in user-item interactions. By examining the behavior and preferences of a group of users, CF identifies similar users and recommends items that have been positively received by those with similar tastes. Applying CF to sports recommendations presents an opportunity to introduce users to new sports events enjoyed by their peers. However, recommending upcoming live sports events introduces unique challenges, such as limited availability and the need to strike a balance between catering to users’ favorite sports and introducing them to new content. This master thesis aims to address these challenges through the development of a personalized recommendation system for upcoming sports events using CF. The system will analyze user viewing history to provide tailored recommendations that facilitate content discovery and enable users to easily locate their preferred sports events. The research objectives include identifying the most suitable collaborative filtering model for sports content recommendation, investigating the factors that influence sports fans’ preferences for specific types of live sports events, and evaluating the effectiveness of personalized recommendations compared to non-personalized approaches. The proposed system is implemented and A/B tested on TV 2 Play, one of Norway’s largest digital streaming platforms, with the ultimate goal of enhancing user experience and engagement by delivering personalized and relevant recommendations for sports content. This research contributes to the field by proposing a novel collaborative filtering recommender for sports based on user viewing sessions, exploring effective strategies for recommending upcoming live sports events, and assessing the system’s performance in terms of accuracy and user satisfaction.

Forstner S., Designing a Dyslexia-Friendly Interaction with News Articles, UiB, Bergen 2023. Supervisor: Bjørnar Tessem.

This master’s thesis presents a research project focused on improving the accessibility of news articles for people with dyslexia. A significant portion of online news content is text-based, which poses challenges for individuals with reading difficulties. While there are existing digital tools for text simplification and summarization, there is a lack of solutions specifically designed for dyslexic readers, and even fewer adapt to languages other than English. The project addresses this gap with the development of a prototype that generates simplified versions of Norwegian news articles. Adopting a user-centered perspective, qualitative research methods and literature research were conducted to provide a basis for the design process. The later developed prototype aims to provide a flexible model that can accommodate the different reading experiences and perspectives of dyslexic people. Its main functionality lies in visual and content-related text modifications. Together with results from corresponding user tests, the high-fidelity prototype provides detailed findings on how news articles can be made more accessible. They offer insights into the requirements and needs of news consumers with dyslexia and explore the potential of automatic text simplification in this context. The results can benefit companies, institutions, and organizations that are seeking to provide accessible news content, eliminating the need for manual simplification of every article. Moreover, the research conducted in this project can support further studies on design and digital accessibility solutions.

Hansen K., Nyheter fortalt slik unge voksne liker det, UiB, Bergen 2023. Supervisor: Brita Ytre-Arne. 

Nyhetspodkastene Aftenposten Forklart og NRK Oppdatert ligger uke etter uke på topplisten over de mest lyttede podkastene i Norge. Til tross for nyhetsformatets økende popularitet, har det til nå vært lite forskning på hva det er med formatet som appellerer til unge voksne. Dette masterprosjektet undersøker nyhetspodkaster, og hva det er med formatet som appellerer til unge voksne. Ved å intervjue 11 informanter som lytter til nyhetspodkaster ukentlig, vil denne masteroppgaven undersøke hvilke faktorer som er relevante for unge voksne sin opplevelse av nyhetspodkaster, og hvordan det er å få nyheter i podkastformat sammenlignet med andre nyhetsmedier. Funnene i denne undersøkelsen viser at nyhetspodkaster er en engasjerende måte å få med seg nyheter, og bruken av fortellingsteknikker i formidlingen er en hovedgrunn til at informantene velger å lytte. Sammenlignet med tradisjonelle nyhetsmedier viser funnene fra studien at nyhetspodkaster hjelper lytterne med å få en bedre oversikt og forståelse av dagsaktuelle nyheter.

Jahr O., Creating an Agglomerative Clustering Approach Using GDELT, UiB, Bergen 2023. Supervisor: Andreas L. Opdahl.

GDELT is a project with a large scale, continuously updated databank that provides a real-time image of the global news picture by outputting these as files that can be downloaded and used by anyone. However, this data is of low granularity, and each source of data does not provide much information on its own. This thesis attempts to leverage the large amount of data available by utilizing a Hierarchical Agglomerative Cluster method to identify news articles that report about the same real life event. To do this, the thesis also explores if the GDELT data is granular enough to be used without extensive preprocessing, and if a distance metric for the cluster algorithm can be created. The findings show promising results when regarded with qualitative measures, but the quantitative measures are not yet optimized. Inherent flaws in GDELT and clustering algorithms are a hurdle to be overcome before the real potential of GDELT’s data can be unleashed, and this thesis will explore some of these difficulties and make recommendations for how to circumvent them in future works.

Kirkhorn T., Generative språkmodeller for automatisert tekstforenkling: Tilpasning av nyhetsartikler for lesbarhet hos personer med dysleksi, together with NRK and NTB, UiB, Bergen 2023. Supervisor: Bjørnar Tessem.

I denne masteroppgaven utforskes utviklingen av en prototype for automatisert tekstforenkling med fokus på tilpasning av nyhetsartikler for personer med dysleksi. Prosjektet involverer flere iterasjoner for å finne den mest effektive tilnærmingen for å forenkle tekstene ved hjelp av generative språkmodeller. Prototypen inneholder tre nivåer av forenklet tekst, der nivå null representerer den originale teksten og nivå en til tre gradvis forenkler innholdet. Testing og evaluering av prototypen gjennomføres ved å utføre brukertester med målgruppen og samle tilbakemeldinger fra eksperter innen feltet. Disse tilbakemeldingene er verdifulle i forbedringen av prototypen og tekstforenklingen. En viktig del av arbeidet er eksperimentering med ulike prompts og instruksjoner for å finne den mest effektive måten å be språkmodellen forenkle artiklene på. Observeringer indikerer at formuleringen av promptene har en innvirkning på resultatene, og at det er utfordringer knyttet til språket i de genererte versjonene. Dette vil kreve manuell redigering av tekstene for å oppnå ønsket lesbarhet og kvalitet. Funnene indikerer at språkmodellen fortsatt har begrensninger når det gjelder å generere grammatisk korrekte setninger og å velge passende ord i konteksten ved omformulering av en eksisterende tekst. Videre identifiseres utfordringer knyttet til maksgrensen for tokens, som begrenser promptlengden, artikkellengden og den forenklede teksten. Det er viktig å påpeke at denne forskningen blir gjennomført innenfor de nåværende rammene for språkmodeller for å teste reformulering av norske nyhetsartikler. Med fremtidig utvikling av språkmodeller og økende tokenkapasitet, samt inkorporering av retningslinjer for klarspråk direkte i modellene, kan vi forvente ytterligere forbedringer i automatisert tekstforenkling.

Klingenberg P., Using content- and behavioural data for recommendations in the Norwegian news market, together with BT, UiB, Bergen 2023. Supervisor: Mehdi Elahi.

In the news domain, the technological development has contributed to always accessible platforms with frequent publishment of articles. This leads to a vast amount of articles, in addition to a multitude of choices. Interactions executed by users (clicks, views or ratings) and content of articles (text, section, etc.) on such platforms can be utilized to create personalised recommendations. However, there can arise a situation where there is a lack of data, i.e., when a new article is published or there is a new user, known as the cold start problem. In addition, personalised recommendations in the news industry, where user behaviour are the basis of recommendations, are more complex in comparison with, for example, the movie domain due to the timeliness with short duration and relevance of articles. Recommendations provided by content data may mitigate this issue. This thesis investigates the viability of using both content- and user-behavioural data to generate recommendations in the Norwegian media market in collaboration with one of Norway’s largest newspapers, i.e., Bergens Tidende (BT). The first technique investigated is collaborative-based filtering, built on the recommender models Alternating Least Squares, Bayesian Personalized Ranking and Logistic Matrix Factorization, which is specially tailored to use user behavioural data as input. The second technique investigated is content-based filtering, built on a state-of-the-art architecture named BERT, specially trained to draw the semantic content of sentences in the Norwegian language. A comprehensive offline evaluation of both techniques is executed using dimension reduction and a diverse selection of accuracy metrics. In addition, an online evaluation of a large-scale A/B test of the content-based filtering technique against a former recommendation technique in BT is performed. The evaluation methods used in the online experiment are descriptive data- and Bayesian analysis. The results of the collaborative filtering technique have shown increased performance in different stages of filtering, in addition to the importance of hyperparameter tuning. The results have shown promising performance based on the content-based filtering technique, indicating that this attracts users’ interest at a greater scale, even the groups that usually show less interaction.

Moholdt E., Detecting Out-of-Context Image-Caption Pairs in News: A Counter-Intuitive Method, UiB, Bergen 2023. Supervisor: Duc-Tien Dang-Nguyen.

The growth of misinformation and re-contextualized media in social media and news along leads to an increasing need for fact-checking methods. Concurrently, the recent advancement in generative models enables a new method of spreading misinformation via AI manipulated images and videos. With the advent of deep-learning technologies used to create ‘deepfakes’ and generative models that can produce near realistic images from text prompts, fabricated media is getting increasingly hard to detect. The media covered in this project is re-contextualised media referred to as ‘cheapfakes’, which consists of image and its associated captions. For example, an image might appear in different online sources with different caption claims. The field of generative models is currently experiencing rapid growth. While text-to-image generative models can potentially be misused for ‘deepfake’ creation and spreading misinformation, this thesis present a positive application. In this thesis we present a novel approach using generative image models to our advantage for cheapfake detection. We present two new datasets with a total of 6800 images generated using two different generative models including (1) DALL-E 2, and (2) Stable-Diffusion. We propose text-to-image generative models can be employed to detect out-of-context media. The similarity or dissimilarity of the generated images versus the original image may serve as an indicator of opposing or misleading captions containing out-of-context news. We evaluate and employ a handful of methods for computing image similarity. Our cheapfake detection model is the first of its kind that utilizes generative models, and we achieve a 68% accuracy score for cheapfake detection, proving that image generation models can be utilized to efficiently detect cheapfake media. Our models similarity measures also align with human perception of image similarity. Moreover, we outline several opportunities for optimization. We are confident that the method proposed in this thesis can further research on generative models in the field of cheapfake detection. We are confident that the resulting datasets can be used to train and evaluate new models aimed at detecting cheapfake media and would further research in this area.

Nilsen A., Investigating the Effects of Instagram Filters on Perceived Trust in Online News Posts, UiB, Bergen 2023. Supervisor: Alain Starke, Christoph Trattner. 

The increasing overgrowth of information is only getting harder to navigate through and the spread of fake news and misinformation is concerning. With the shift towards digital deliv- ery of news and concerns about the accuracy and reliability of information shared on social media, it is important to understand the factors that contribute to trust in social media news. Motivated by these challenges, this study aimed to investigate what effect Instagram filters have on users’ perceived trust in online news posts. Trust ratings of four different articles with four different image filters, including the original image, were collected through an on- line user study. Also, the role of general trust and familiarity with the topic and the context of the different topics were explored. We did an online experiment with 204 participants recruited from a crowdsourcing platform. Participants were asked to answer six questions per online news post shown. Our analysis revealed that while Instagram filters overall may not affect perceived trust, specific visual characteristics of the filters such as brightness and contrast may play a role. Additionally, individual differences in general trust and attitude to- wards the topic may influence the users’ perception of trust. The study also found that there may be differences in perceived trust across different news topics. Thus, there could be other factors influencing the users’ participants of trust.

Rosnes D., Evaluating Feature-Specific Similarity Metrics using Human Judgments for Norwegian News, UiB, Bergen 2023. Supervisor: Christoph Trattner, Alain Starke.


This master’s thesis delves into the measurement of similarity between news articles within the Norwegian news domain. Four central questions form the basis of the thesis: the identification of information cues utilized by readers, the effectiveness of specific similarity metrics, the comparison with other domains, and the exploration of differences in human similarity ratings between national and local news. Key findings include that a Sentence-BERT metric, applied to the body text, best represented human similarity judgments. Compared to other news domains, the Norwegian news domain showed stronger correlations for a majority of the metrics. A minimal contrast was observed between human ratings for local and national news, with local news considered slightly more similar. This disparity between local and national levels, however, did not markedly impact how metrics represented human similarity judgments. The findings from this thesis may provide valuable insights for enhancing news recommendation systems within the news sector.

Ulvolden H., Complexity-based Ordering of Suffixes in Norwegian, UiB, Bergen 2023. Supervisor: Koenraad de Smedt. 

Denne studien undersøker bruken av 18 avledningssuffikser i norsk bokmål, og har som mål å replikere studien gjort av Hay and Plag (2004). Studien deres forsøkte å forklare hvordan restriksjoner på affikskombinasjoner bestemmes av såkalt kompleksitetsbasert sortering. Den viser at et suffiks har større sannsynlighet for å plasseres etter et annet suffiks hvis det er mer produktivt (i stand til å produsere nye ord) og mer parserbart (lettere for en språkbruker å separere fra baseordet sitt). På samme måte har jeg brukt korpusmateriale til å finne ut hvilke kombinasjoner av to suffikser som finnes i norsk. Deretter har jeg brukt korpus til å samle informasjon om suffiksenes produktivitet og parserbarhet. Produktivitet måles ved å sammenligne proporsjonen av hapaxer (ord som kun dukker opp én gang i et korpus) med det totale antallet ting som inneholder nevnt suffiks. Parserbarhet måles ved å sammenligne frekvensen av avledninger med frekvensen av baseord (f.eks. ærlighet og ærlig), som viser hvor sannsynlig det er at en språkbruker oppfatter dem som enkelte ord istedenfor et baseord med et suffiks. Suffikskombinasjoner ble organisert i et hierarki basert på hvilke suffikser som kan opptre etter andre. Resultatene viser at i motsetning til i engelsk kan noen kombinasjoner også opptre i motsatt rekkefølge, og affikser kan derfor ikke organiseres i et like strengt hierarki som i engelsk. Resultatene viser også en korrelasjon mellom produktivitet og parserbarhet, dvs. at et produktivt suffiks også har lettere for å separeres fra baseordet sitt. Vi ser også en viss sammenheng mellom disse to faktorene og suffikshierarkiet. Selv om det er vanskelig å si hvor mye dette henger sammen, er det tydelig at mange av de fleksible suffiksene som kan plasseres etter andre også er mer produktive og parserbare.

Vlasenko A., Multi-List Recommendations for Personalizing Streaming Content, together with TV 2, UiB, Bergen 2023. Supervisor: Christoph Trattner. 

The decision behind choosing a recommender system that yields accurate recommendations yet allows users to explore more content has been a topic of research in the last decades. This work attempts to find a recommender system for TV 2 Play, a movie streaming platform, that would perform well on implicit feedback data and provide multi-lists as recommenda- tions. Several approaches are examined for suitability, and Collaborative Filtering and Multi- Armed Bandits are decided upon. The models for each approach are built using the pipeline utilized by TV 2 Play. The models are then compared in performance on several evaluation metrics in the first stage of offline testing, yielding Alternating Least Squares and Bayesian Personalized Ranking as the best-performing models. The second stage of offline testing includes testing the two models and their variants with the BM25 weighting scheme applied against each other. The unweighted Bayesian Personalized Ranking model has shown the highest user-centric metrics while maintaining relatively high recommendation-centric met- rics, which led to that model being tested in online settings against the algorithm currently used by TV 2 Play team. The online testing has revealed that our model underperforms compared to the TV 2 Play model when used on the kids’ page but produces equally good results on the movies page. The results can be attributed to the differences in behavioral content consumption patterns between users.

Westli W., Minimum ressursbruk. Maksimum utbytte., together with TV 2, UiB, Bergen 2023. Supervisor: Lars Nyre.

The media industry is constantly changing and it is therefore important for companies to keep up with technology. Until 2021, TV 2 had had the same logo for almost 30 years, and as TV 2 CEO Olav Sandnes said during a general meeting in 2021, “…TV 2 in recent years (…) has taken big steps in the digital and within streaming without the brand having kept up.” (Hauger, 2021). At the same time as the public’s external profile has changed, the internal following must also keep up. This master’s thesis is a discussion of how student projects should be set up in order to get as good a starting point and result as possible when collaborating with a media company. The focus is on planning student prototypes to turn them into products. I use theories taken from Lim his article “The Anatomy of Prototypes” about how prototypes are built and the process around them, while I use the approaches and methods from Eric Ries’ “The Lean Startup” and Jake Knapp’s “Sprint”. The analysis work is set up so that I analyze the tools we have created for the prototype “TV 2 Pippen Relevans”, before I then go into what is needed to create a prototype that can become a product.

Postponed access: the abstract will be accessible after 2024-06-01.

Ekren K., Experiments on Satire Detection for Norwegian News Articles, together with Web64, UiB, Bergen 2022. Supervisor: Samia Touileb.


The spread and amount of misinformation is increasing. The World Economic Forum (WEF) has listed it as one of the main threats to our society (Howell, 2013). Satire is one of the problems when it comes to misinformation, more specifically news satire. News satire is a genre of satire that resembles the characteristics of true journalistic reporting, while keeping the main objective of satire that is: use of a combination of humor and irony, usually with exaggeration, to expose and make fun of political or newsworthy issues.

In this thesis we present to the best of our knowledge, the first attempt to automat- ically detecting satire in Norwegian news articles. Automatically identifying satirical news pieces can aid in minimizing the potential deceptive impact of satire. To this end, we employ three classification methods, namely Naïve Bayes, SVM (support-vector machines) and logistic regression, based on TF-IDF (Term Frequency Inverse Docu- ment Frequency) feature weights. All three machine learning models achieved similar results.

In total, our dataset incorporates 6322 articles containing a balanced collection of satirical and non-satirical news texts from various domains (3161 satirical and 3161 non-satirical). Using this corpus we proposed three cross-domain satire detection tasks, one considering only the use of headlines, one considering only the use of article texts, and lastly one considering full articles, including both headlines and texts.

After experimenting on the test sets, we achieved the top accuracy score of 98% using only text as input, and the combination of title and text as input, with SVM. We observe that satire detection on news headlines was significantly more challenging, the top accuracy score being 76% using SVM. We believe that this shows that the automatic detection of satire using only headlines is quite challenging. Especially when using simple machine learning approaches, and we believe that this might be due to the length of headlines and the need for more context.


Irony is a complex phenomenon of human communication and due to its con- textual nature has been notoriously difficult for machine learning algorithms to detect. With an established practical definition of irony based in the environ- ment of Facebook comment sections. Used together with a Norwegian language pre-trained BERT model converted to a long version that supports longer text inputs, and a Norwegian Facebook comment dataset with contextual article and reply comment text included. It was found that the long BERT model trained on the context included inputs dataset outperformed the short BERT models trained on datasets of the same and more comments, but without the contextual infor- mation encoded.

Jakobsen D., Visual analysis, recommendation and personalization, UiB, Bergen 2022. Supervisor: Mehdi Elahi.


Advertisement is one of the primary sources of revenue for companies. Advertisement can result in increased sales for the companies by promoting their products or by hosting the advertisement of the other companies. While there have been many studies investigating the effectiveness of the advertisement in various domains, there are still challenges to be addressed. Notable challenges are overabundance, privacy, spamming, or irrelevant placement of advertisement. The results of such challenges can be poor quality of advertisement irrelevant to the audience interest or irrelevant in the context of the content in which it has been shown. A colourful and shiny image in an advertisement, promoting a luxury product that is placed together with a news article discussing the growing poverty in certain regions of the world can be an example. This can lead to dissatisfaction of the audience when experiencing a particular irrelevant advertisement. Personalisation techniques can be useful in addressing such challenges. For example it can be used to match the advertisement shown in news outlets to the article. Such techniques can analyse the content of the news articles and find right advertisement that is better suited to the article. However, this requires rich data of displayed advertisement which are not always available. That can be one of the reasons why viewing an irrelevant advertisement is still part of our everyday experience. In this thesis, I have explored the potential of using automatic visual analysis to obtain better representation of the advertisement and to improve the relevance of advertisement to the audience. For such analysis, a number of visual features has been extracted from the images of the advertisements and analysed to better understand their correlations with audience behaviour online when interacting with advertisement campaigns. I have received the data set from Amedia, one of the largest media companies in Norway. The data set includes campaigns and audience behaviour, and the images associated with the campaigns. Four analyses have been conducted in the thesis including, exploratory analysis and correlation analysis. Machine learning models have been built (using CatBoost) based on the audience behaviours and visual features extracted from advertisement. In addition a method called SHAP (SHapley Additive exPlanations) has been used to find explanation of model prediction. The contribution of this thesis includes preprocessing and exploratory analyses of the industry data, a novel data set containing the visual features extracted from the images of the advertisement, a Python prototype, and the results of analyses based on this prototype.

Johansen S., Expanding Digital Workspaces Using Cross-Device Interactive Applications, UiB, Bergen 2022. Supervisor: Morten Fjeld.


In this thesis I explore the use of a cross-device interactive application and how single individuals could make use of multiple connected devices by themselves for their work. I have created a functional prototype that is used to evaluate the workload of a single device versus the use of multiple devices with the use of in-built cross-device interaction capability. The process of this study consisted of research, prototyping, programming the application, and performing a study to analyze the data gathered. Due to the covid-19 situation, the project had major re-routes in 2021 to handle its effects, resulting in an additional year of work to complete it.

Khutarniuk Y., Cross-Lingual Approaches to Identifying Argument Components and Relations in Norwegian Reviews, UiO, Oslo 2022. Supervisors: Samia Touileb, Lilja Øvrelid.


Argument mining is the process of automatic extraction of certain argu- mentation structures from data. Argument mining consists of several stages such as argument component detection, argument component clas- sification, and argumentative discourse analysis. The lack of training data in low resource languages is a common issue in argument mining applica- tions. In this work we analyse the possibilities for the application of zero- shot and few-shot language transfer models trained on the language ma- terial in a resource-rich language (English) for the tasks of argument com- ponent detection, and argument component classification in a low-resource language (Norwegian) with the aim to find out if these techniques can help overcome the challenge of no available training data. In addition, we com- pare models based on different transformer architectures and experiment with additional hand-crafted features.

Olsen D., Movie Recommendation based on Stylistic Visual Features, together with TV 2, UiB, Bergen 2022. Supervisor: Mehdi Elahi.


When a new movie is added to the catalogue of a recommendation-empowered movie streaming platform, the system exploits various types of data (e.g., clicks, views, and ratings) in order to generate personalized recommendations for the users. However, in the absence of sufficient data, undesired situations can arise where the system may fail to include the new movie in the recommendation list. This is known as the Cold Start problem. A solution can be using content features attributed to the movies (e.g., tags, genre, and description). However, such features require expensive editorial efforts and it is not necessarily available in good quantity or quality.

This thesis investigates the viability of using novel stylistic visual features as meta- data to incorporate in the movie recommendation process. The visual features represent the stylistic properties of the movies and can have a wide range of forms, e.g., color palette, contrast, and brightness. The stylistic visual features can be automatically ex- tracted, and hence, do not require any (manual) human annotation. Accordingly, the thesis proposes a novel technique for generating recommendation based on such visual features and describes the technical details for different stages of the process. The tech- nique has been evaluated in both offline and online experiments and different scenarios, i.e., cold start and warm start. The online experiment has been conducted in collabo- ration with TV 2, one of Noways largest digital streaming platforms adopting an A/B testing methodology. The proposed technique includes utilizing the extracted visual features when used individually (in a similarity based recommendation process), and when combined with other types of data (in a hybrid recommendation process). The results of the experiments have been promising and shown that the stylistic visual fea- tures can be beneficial particularly in the hybrid recommendation process in the cold start scenario.

Sjøen V., Digital Image Forensics In The Wild: Social Media Platforms, UiB, Bergen 2022. Supervisor: Duc Tien Dang Nguyen.

Solberg V., News Recommendation based on Human Similarity Judgment, together with Bergens Tidende, UiB, Bergen 2022. Supervisors: Christoph Trattner, Alain Starke.


Similar item recommendation is one of the most popular types of recommender systems. As the name implies, the objective is to recommend items that are similar to a reference item. The news domain is one of the many that employ this form of recommendation, which utilize similarity functions in order to calculate the similarity. In this study, human judgments of article similarity were acquired using an online user study in which each of the 173 participants evaluated the similarity of 12 pairs of articles. Each of the 12 article pairs had their own unique characteristics. One pair would be made up of two completely dissimilar articles, while the other pairs had either a shared topic, a named entity in common, publication dates in close proximity, or some combination of these three characteristics. Half the pairs contained articles from the News (i.e., recent events) category, while the other half contained Sport articles. The similarity of the same article pairings was then calculated utilizing various similarity functions, and the correlation between human judgment and function scores was computed. This thesis found that the correlation ranged from weak to strong, depending on the function. The thesis also found that the correlation is largely dependent on the whether the articles have certain characteristics in common. On average, the functions correlated more strongly to human judgment if the articles belonged the category News (i.e., recent events) than Sport. The functions were also better at predicting human similarity when the articles in question were relatively similar to one another. The novel work presented in this thesis shows that the correlation between human judgment and similarity functions can be stronger than previous work has suggested, if news articles are paired in a meaningful way.

Stenerud B., Reaksjonsvideoer. En studie av unge voksnes opplevelser med mediefenomenet reaksjonsvideoer, UiB, Bergen 2022. Supervisor: John Magnus Dahl.


Denne oppgaven undersøker hvordan unge voksne benytter seg av mediefenomenet reaksjonsvideoer. Ettersom plattformer som YouTube og Twitch blir mer populære og sentrale i folks hverdag vil man naturligvis også bli smittet av trender. Den siste tiden har det blitt populært med reaksjonsvideoer som handler om at streamere eller innholdsprodusenter filmer seg selv reagere til hva enn som skjer på skjermen deres. Et eksempel på dette er Asmongold som har rundt 200 000 seere på hans livestream av at han selv reagerer og kommenterer på rettsaken til Johnny Depp og Amber Heard. Fenomenet vil bli sett i lys av opinionsledere og parasosiale relasjoner. Gjennom kvalitative dybdeintervjuer tar studiet for seg 10 informanters opplevelser, samt hva de får ut av fenomenet. Studien ser og på hvordan informantene selv tolker og forklarer fenomenet, og hvordan de bruker reaksjonsvideoer annerledes fra «vanlige» videoer. Funnene i denne studien tyder på at informantene bruker reaksjonsvideoer som en måte å få mer informasjon om det man ser på, og for å få inn flere perspektiver. Reaksjonsvideoer bidrar og til økt selskap mens informantene ser på videoene. På denne måten opplever informantene at de ser på en video «med» noen som de kan føle at de kjenner. Man kan også se at informantene bruker reaksjonsvideoer, og kommentarfelt, som en måte å bekrefte egne tanker og meninger, samt at slike videoer fungerer som en måte å se på andre gjenoppleve det man selv har gjort.

Årmot I., Smarttelefonen som ein del av konsertopplevinga, UiB, Bergen 2022. Supervisor: John Magnus Dahl.


Smarttelefonen er ein integrert del av kvardagslivet vårt og me brukar den til fleire praktiske formål, og upraktiske. I kvardagen blir smarttelefonen hylla fordi den har fleire bruksområder, men samtidig kan den bli slakta for evne si til å forstyrre merksemda vår, eller forstyrre andre som irriterer seg over andre sin bruk i situasjonar dei skulle helst vore forutan. Denne masteroppgåva prøvar å kome med eit bidrag til studiar om smarttelefonbruk i eit medieerfaringsperspektiv. Med inspirasjon frå irritasjonen enkelte publikummarar uttrykkjer i kronikkar og lesarinnlegg om andre publikummarar sin bruk av smarttelefonen gjennom ein heil konsert, undersøkar denne oppgåva korleis unge vaksne frå alderen 20 til 26 år brukar smarttelefonen i ein konsertsamanheng. Ved å utføre deltakande observasjon av konsertane til Daniel Kvammen og Kjartan Lauritzen på USF Verftet i Bergen, etterfølgt av kvalitative intervju i etterkant av utvalde publikummarar, prøvar denne oppgåva å svare på korleis smarttelefonen er ein del av ei medieerfaring av å reise på konsert. I eit medieerfaringsperspektiv undersøker oppgåva korleis informantar brukar smarttelefonen på konsertane til Daniel Kvammen og Kjartan Lauritzen. Altså kva tar dei bilete og videoar av? Kvifor deler dei bileta og videoane med andre? Kva slags betydingar har det for informantane å dele konsertbileta og -videoane? Funna i analysen kan indikere at informantane ikkje er opptekne av å bruke smarttelefonen heile tida på konsertane til Daniel Kvammen og Kjartan Lauritzen ettersom det kan forstyrre ein tilstand av flow. Likevel brukte informantane smarttelefonen for å ta bilete og videoar av konsertane, og med tanke på at smarttelefonen kan verke å vere naturalisert i liva til informantane i den grad at dei ikkje tenkjer over bruken lengre, kan det tyde på at smarttelefonen er ein del av flowen. Naturaliseringa av smarttelefonen i liva til informantane kan også gjere det naturleg for dei å dele bileta og videoane med andre om at dei opplev noko utanom det vanlege i kvardagen, og dei informantane deler med er i hovudsak vener og familie. Å dele bilete og videoar frå konsertane kan ha ein verdi ved at det opnar opp for at dei kan halde ved like eksisterande forhold og pleie nære relasjonar som informantane ikkje ser så ofte i kvardagen.


Aarnes P., Named Entity Recognition in Speech to Text Transcripts, together with TV 2, UiB, ongoing. Supervisor: Samia Touileb.

Bødtker D., Nudging Stories in News Search: The Role of Visual Attractiveness and Presentation Order Effects on Online News Consumption, UiB, ongoing. Supervisors: Alain Starke, Christoph Trattner.

Espeseth F., Media Analytics for Personalization and Advertisement, together with Amedia, UiB, ongoing. Supervisors: Mehdi Elahi, Igor Pipkin.

Rosenvinge F., Automated Identification of Severe Errors in Speech to Text Transcripts, together with TV 2, UiB, ongoing. Supervisor: Samia Touileb.

Stokke E., Automated Speech to Text (ASR) Quality Metric, together with TV 2, UiB, ongoing. Supervisor: Samia Touileb.

Find us

Lars Hilles gate 30
5008 Bergen

Contact us



Responsible Editor:
Centre Director Prof. Dr. Christoph Trattner


Subscribe to our monthly Newsletter by sending mail to


Hosted by 




Copyright © University of Bergen 2024