Browsing by Title

A tag is a freely chosen keyword that a user attaches to an item. Offering a simple, cheap, and natural way to describe content, tagging has become popular in contemporary web applications. The tag genome is a data structure that contains item-tag relevance scores, i.e., continuous scale numbers from 0 to 1 indicating how relevant a tag is for an item. For example, the tag romantic comedy has a relevance score of 0.97 for the movie Love Actually. With sufficient data, a tag genome dataset can be constructed for any domain. To the best of available knowledge, there are tag genome datasets for movies and books. The tag genome for movies is used in a movie recommender and for various purposes in recommender systems research, such as detecting filter bubbles and serendipity. Creating a diverse tag genome dataset requires an effective machine learning solution, as manual assessment of item-tag relevance scores is impractical. The current state-of-the-art solution, called TagDL, uses features extracted from user-generated tags, reviews, and ratings to employ a multilayer perceptron architecture to predict the item-tag relevance scores. This study aims to enhance TagDL by extracting more features from the embeddings of textual content, namely tags, user reviews, and item titles, using Bidirectional Encoder Representations from Transformers (BERT). The results show that features based on BERT embeddings have a potential positive impact on item-tag relevance score prediction. However, the results do not generalize to both tag genome datasets, improving the results only for the movie dataset. This may indicate that the new features have a stronger impact if the amount of available training data is smaller, as with the movie dataset. Moreover, this thesis discusses future work ideas and implementation possibilities.

Large language models (LLMs) have been proven to be state-of-the-art solutions for many NLP benchmarks. However, LLMs in real applications face many limitations. Although such models are seen to contain real-world knowledge, it is kept implicitly in their parameters that cannot be revised and extended unless expensive additional training is performed. These models can hallucinate by confidently producing human-like texts which might contain misleading information. The knowledge limitation and the tendency to hallucinate cause LLMs to struggle with out-of-domain settings. Furthermore, LLMs lack transparency in that their responses are products of big black-box models. While fine-tuning can mitigate some of these issues, it requires high computing resources. On the other hand, retrieval augmentation has been used to tackle knowledge-intensive tasks and proven by recent studies to be effective when coupled with LLMs. In this thesis, we explore Retrieval-Augmented Generation (RAG), a framework to augment generative LLMs with a neural retriever component, in a domain-specific question answering (QA) task. Empirically, we study how RAG helps LLMs in knowledge-intensive situations and explore design decisions in building a RAG pipeline. Our findings underscore the benefits of RAG in the studied situation by showing that leveraging retrieval augmentation yields significant improvement on QA performance over using a pre-trained LLM alone. Furthermore, incorporating RAG in an LLM-driven QA pipeline results in a QA system that accompanies its predictions with evidence documents, leading to a more trustworthy and grounded AI applications.

Automatic bird sound recognition has been studied by computer scientists since late 1990s. Various techniques have been exploited, but no general method, that could even nearly match the performance of a human expert, has been developed yet. In this thesis, the subject is approached by reviewing alternative methods for cross-correlation as a similarity measure between two signals in template-based bird sound recognition models. Template-specific binary classification models are fit with different methods and their performance is compared. The contemplated methods are template averaging and procession before applying cross-correlation, use of texture features as additional predictors, and feature extraction through transfer learning with convolutional neural networks. It is shown that the classification performance of template-specific models can be improved by template refinement and utilizing neural networks’ ability to automatically extract relevant features from bird sound spectrograms.

Cancer is a worldwide health problem; in 2018 9.6 million people died of cancer, meaning that about 1 in 6 deaths was caused by it. The challenge with cancer drug therapy has been the development of cancer drugs that are effective against cancer but are not harmful to the healthy cells. One of the solutions to this has been antibody-drug conjugates (ADCs), where a cytotoxic drug is bound to an antibody. The antibody binds to specific antigen present on the surface of the cancer cell, thus working as a vessel to carry the drug specifically to the cancer cells. Monomethyl auristatin E (MMAE) and monomethyl auristatin F (MMAF) are mitosis preventing cancer drugs. The auristatins are pentapeptides that were developed from dolastatin 10. MMAE consist of monomethyl valine (MeVal), valine (Val), dolaisoleiune (Dil), dolaproine (Dap) and norephedrine (PPA). MMAF has otherwise similar structure, but norephedrine is replaced by phenylalanine (Phe). They prevent cell division and cancer cell proliferation by binding to microtubules and are thus able to kill any kind of cell. By attaching the auristatin to an antibody that targets cancer cells, they can effectively be used in the treatment of cancer. MMAE and MMAF exist as two conformers in solution, namely as cis- and trans-conformers. The trans-conformer resembles the biologically active conformer. It was recently noted that in solution 50-60 % of the MMAE and MMAF-molecules exist in the biologically inactive cis-conformer. The molecule changes from one conformer to the other by the rotation of an amide bond. However, this takes several hours in body temperature. As the amount of the cis-conformer is significant, the efficacy of the drug is decreased, and the possibility of side effects is increased. It is possible that the molecule leaves the cancer cell in its inactive form, migrates to healthy cells and tissue, and transforms to the active form there, damaging the healthy cell. The goal of this study was to modify the structure of the auristatins so that the cis/trans-equilibrium would change to favor the biologically active trans-conformer. The modifications were done virtually, and the relative energies were computed using high-level quantum chemical methods, at density functional theory (DFT), 2nd order perturbation theory (MP2) and coupled cluster levels. Intramolecular interactions were analyzed computationally, employing symmetry-adapted perturbation theory and the non-covalent interactions analysis. The results suggest that simple halogenation of the benzene ring para-position is able to significantly shift the cis/trans-equilibrium to favor the trans-conformer. This is due to changes in intramolecular interactions that favor the trans-conformer after halogenation. For example, the NCI analysis shows that the halogen atom invokes stabilizing intramolecular interactions with the Dil amino acid; there is no such interaction between the para-position hydrogen and Dil in the original molecules. We also performed docking studies that show that the halogenated molecules can bind to microtubules, thus confirming that the modified structures have potential to be developed into new, more efficient and safe cancer drugs. The most promising drug candidates are Cl-MMAF, F-MMAF, and F-MMAE where 94, 90, and 79 % of the molecule is predicted to exist in the biologically active trans-conformer, respectively.

Profit Software's Profit Life and Pension (PLP) is an investment insurance management system. This means that PLP handles investment insurances from the moment they are sold to when they eventually expire. For a system that handles money, it is important that it can be trusted. Therefore, testing is a required part of PLP's development. This thesis is an investigation into PLP's testing strategy. In this thesis we analyse PLP's current testing strategy to find flaws and impediments. We then offer improvement suggestions to the identified problem areas as well as suggest additions which we found could be beneficial.

Web development is in great demand these days. Constantly developing technologies enables to create impressive websites and mitigates the amount of development work. However, it is useful to consider the performance aspect, which affects directly to user experience. Performance in this context means website’s load times. Front end web development typically involves using Cascading Style Sheets (CSS) which is a style sheet language and a web technology that is used to describe the visual presentation of a website. This research consist of a literature review part, which contains background knowledge about how web browsers work, performance in general, performance metrics along with CSS performance optimization and an empirical part, which includes different benchmarks presented in major software industry conferences for testing the performance of a certain CSS feature, that have a possibility to improve the performance of the website. The loading times obtained from the benchmarks are reviewed and compared with each other. In addition, a few techniques are presented that do not have their own benchmark, but which may have an effect on performance. To highlight the results, CSS performance is usually not the biggest bottleneck of performance on a website, since the overall style calculation takes about a quarter of the total runtime calculation on average. However, utilizing some particular techniques and managing to shrink the style calculation costs can be valuable. Based on the benchmarks on this research, using shadow DOM and scoped styles have a positive effect on style performance. For layout, performance benefits can be achieved by utilizing CSS containment and concurrent rendering. From other practices, it can be concluded that removing unused CSS, avoiding reflow and repaint along with complex selectors and considering the usage of web fonts a better results can be achieved in terms of performance.

NP-hard optimization problems can be found in various real-world settings such as scheduling, planning and data analysis. Coming up with algorithms that can efficiently solve these problems can save various rescources. Instead of developing problem domain specific algorithms we can encode a problem instance as an instance of maximum satisfiability (MaxSAT), which is an optimization extension of Boolean satisfiability (SAT). We can then solve instances resulting from this encoding using MaxSAT specific algorithms. This way we can solve instances in various different problem domains by focusing on developing algorithms to solve MaxSAT instances. Computing an optimal solution and proving optimality of the found solution can be time-consuming in real-world settings. Finding an optimal solution for problems in these settings is often not feasible. Instead we are only interested in finding a good quality solution fast. Incomplete solvers trade guaranteed optimality for better scalability. In this thesis, we study an incomplete solution approach for solving MaxSAT based on linear programming relaxation and rounding. Linear programming (LP) relaxation and rounding has been used for obtaining approximation algorithms on various NP-hard optimization problems. As such we are interested in investigating the effectiveness of this approach on MaxSAT. We describe multiple rounding heuristics that are empirically evaluated on random, crafted and industrial MaxSAT instances from yearly MaxSAT Evaluations. We compare rounding approaches against each other and to state-of-the-art incomplete solvers SATLike and Loandra. The LP relaxation based rounding approaches are not competitive in general against either SATLike or Loandra However, for some problem domains our approach manages to be competitive against SATLike and Loandra.

XML-tietomallin käyttö on yleistynyt mm. rakenteisissa dokumenteissa, verkkosovellusten toteuttamisessa ja Internetissä tapahtuvassa tiedonsiirrossa. Tämän myötä tarve XML-muotoisen tiedon pysyvään säilyttämiseen on kasvanut. Tähän tarkoitukseen on kehitetty XML-tietomallia tiedonsäilytys- ja käsittelymuotonaan käyttäviä XML-pohjaisia tietokantoja. XML-muotoiset dokumentit ovat usein rakenteeltaan monimuotoisia ja kooltaan suuria. Tämän vuoksi XML-tietokannanhallintajärjestelmä on suunniteltava ja toteutettava tehokkaaksi, jotta sen avulla voidaan kohtuullisin laitteistoresurssein ja lyhyin vasteajoin suorittaa suuriakin määriä tietokantakyselyitä ja -päivityksiä, jotka voivat myös olla monipuolisia ja rinnakkaisia ja kohdistua suureen määrään tietoa kerrallaan. Tässä työssä esitetään, miten XML-tietokannanhallintajärjestelmän suorituskykyä voidaan merkittävästi parantaa dokumenttien indeksoinnilla. Indeksoinnissa XML-dokumenttien elementeille luodaan yksikäsitteiset tunnisteet, joihin perustuen luodaan erilaisia indeksihakemistoja. Indeksoinnin avulla tieto voidaan tehokkaasti paikantaa tietokannan tietosivuilta ja siirtää tietokannanhallintajärjestelmän tietosivujen ja puskurin välillä, mikä nopeuttaa tietokannanhallintajärjestelmän toimintaa ja lisää sen kykyä käsitellä rinnakkaisia luku- ja kirjoituspyyntöjä. Indeksoinnin avulla voidaan myös tehostaa tietokannanhallintajärjestelmän kyselynkäsittelyalgoritmien toimintaa mahdollistamalla niiden käyttämien joukkoliitosoperaatioiden tehokas toteutus. BaseX- ja eXist ovat XML-pohjaisia tietokannanhallintajärjestelmiä, joissa käytettävissä on useita erilaisia indeksejä. Indeksien toteutus näissä järjestelmissä kuvataan, ja näiden järjestelmien tehokkuutta XML-dokumentteihin tehtävien tietokantakyselyiden suorituksessa mitataan ja arvioidaan tätä varten kehitetyn XMark-koetinkuorman avulla.

Independent Component Analysis (ICA) aims to separate the observed signals into their underlying independent components responsible for generating the observations. Most research in ICA has focused on continuous signals, while the methodology for binary and discrete signals is less developed. Yet, binary observations are equally present in various fields and applications, such as causal discovery, signal processing, and bioinformatics. In the last decade, Boolean OR and XOR mixtures have been shown to be identifiable by ICA, but such models suffer from limited expressivity, calling for new methods to solve the problem. In this thesis, "Independent Component Analysis for Binary Data", we estimate the mixing matrix of ICA from binary observations and an additionally observed auxiliary variable by employing a linear model inspired by the Identifiable Variational Autoencoder (iVAE), which exploits the non-stationarity of the data. The model is optimized with a gradient-based algorithm that uses second-order optimization with limited memory, resulting in a training time in the order of seconds for the particular study cases. We investigate which conditions can lead to the reconstruction of the mixing matrix, concluding that the method is able to identify the mixing matrix when the number of observed variables is greater than the number of sources. In such cases, the linear binary iVAE can reconstruct the mixing matrix up to order and scale indeterminacies, which are considered in the evaluation with the Mean Cosine Similarity Score. Furthermore, the model can reconstruct the mixing matrix even under a limited sample size. Therefore, this work demonstrates the potential for applications in real-world data and also offers a possibility to study and formalize identifiability in future work. In summary, the most important contributions of this thesis are the empirical study of the conditions that enable the mixing matrix reconstruction using the binary iVAE, and the empirical results on the performance and efficiency of the model. The latter was achieved through a new combination of existing methods, including modifications and simplifications of a linear binary iVAE model and the optimization of such a model under limited computational resources.

Kimberlites are a primary source of diamonds. However, not all kimberlites contain diamonds, let alone in the abundance that enable economically profitable exploitation. To assess the diamond potential of kimberlites, the study of kimberlite/diamond indicator minerals (i.e. mantle-derived xenocrysts) can be utilized. In this study the diamond potential of Liqhobong kimberlite cluster in Lesotho, Southern Africa, was studied. Indicator mineral grains (chromian diopside, garnet, chromite, ilmenite) from Liqhobong kimberlites were analysed for major and minor elements using electron microprobe. Results were used to examine formation conditions (P/T) and chemical characteristics of indicator minerals, to define local geotherm and diamond window and to provide general overview to diamond potential of the Liqhobong kimberlite cluster. Based on single clinopyroxene thermobarometry (using Cr-diopside grains as a thermobarometer), the local Liqhobong geotherm is estimated to be ~41 mW/m². It corresponds reasonably accurately to the reference geotherm 40 mW/m², which is widely considered being a decent “average” geotherm at Kaapvaal-Kalahari Craton area. Analysed xenocrysts are principally from the peridotitic source rock. Geochemistry of the indicator minerals show that there is a significant diamond potential in Liqhobong kimberlites. Specifically, Ni-in-garnet thermometry, using the formation depth of the G10 garnets combined to the relevant 40 mW/m² reference geotherm, displays the existence of a significant diamond window at the depth interval of ~140-230 km below the Liqhobong area.