Browsing by Subject "machine learning"

Flares are short, high-energy magnetic events on stars, including the Sun. Observations of young stars and red dwarfs regularly show the occurrence of flare events multiple orders of magnitude more energetic than even the fiercest solar storms ever recorded. As our technology remains vulnerable to disruptions due to space weather, the study of flares and other stellar magnetic activity is crucial. Until recently, the detection of extrasolar flares has required much manual work and observation resources. This work presents a mostly automatic pipeline to detect and estimate the energies of extrasolar flare events from optical light curves. To model and remove the star's background radiation in spite of complex periodicity, short windows of nonlinear support vector regression are used to form a multi-model consensus. Outliers above the background are flagged as likely flare events, and a template model is fitted to the flux residual to estimate the energy. This approach is tested on light curves collected from the stars AB Doradus and EK Draconis by the Transiting Exoplanet Survey Satellite, and dozens of flare events are found. The results are consistent with recent literature, and the method is generalizable for further observations with different telescopes and different stars. Challenges remain regarding edge cases, uncertainties, and reliance on user input.

Along with the rapid scale-up of biological knowledge bases, mechanistic models, especially metabolic network models, are becoming more accurate. On the other hand, machine learning has been widely applied in biomedical researches as a large amount of omics data becomes available in recent years. Thus, it is worth to conduct a study on integration of metabolic network models and machine learning, and the method may result in some biological discoveries. In 2019, MIT researchers proposed an approach called 'White-Box Machine Learning' when they used fluxomics data derived from in silico simulation of a genome-scale metabolic (GEM) model and experimental antibiotic lethality measurements (IC50 values) of E. coli under hundreds of screening conditions to train a linear regression-based machine learning model, and they extracted coefficients of the model to discover some metabolic mechanism involving in antibiotic lethality. In this thesis, we propose a new approach based on the framework of the 'White-Box Machine Learning'. We replace the GEM model with another state-of-the-art metabolic network model -- the expression and thermodynamics flux (ETFL) formulation. We also replace the linear regression-based machine learning model with a novel nonlinear regression model – multi-task elastic net multilayer perceptron (MTENMLP). We apply the approach on the same experimental antibiotic lethality measurements (IC50 values) of E. coli from the 'White-Box Machine Learning' study. Finally, we validate their conclusions and make some new discoveries. Specially, our results show the ppGpp metabolism is active under antibiotic stress, which is supported by some literature. This implies that our approach has potential to make a biological discovery even if we don't know a possible conclusion.

Several nuclear power plants in the European Union are approaching the ends of their originally planned lifetimes. Extensions to the lifetimes are made to secure the supply of nuclear power in the coming decades. To ensure the safe long-term operation of a nuclear power plant, the neutron-induced embrittlement of the reactor pressure vessel (RPV) must be assessed periodically. The embrittlement of RPV steel alloys is determined by measuring the ductile-to-brittle transition temperature (DBTT) and upper-shelf energy (USE) of the material. Traditionally, a destructive Charpy impact test is used to determine the DBTT and USE. This thesis contributes to the NOMAD project. The goal of the NOMAD project is to develop a tool that uses nondestructively measured parameters to estimate the DBTT and USE of RPV steel alloys. The NOMAD Database combines data measured using six nondestructive methods with destructively measured DBTT and USE data. Several non-irradiated and irradiated samples made out of four different steel alloys have been measured. As nondestructively measured parameters do not directly describe material embrittlement, their relationship with the DBTT and USE needs to be determined. A machine learning regression algorithm can be used to build a model that describes the relationship. In this thesis, six models are built using six different algorithms, and their use is studied in predicting the DBTT and USE based on the nondestructively measured parameters in the NOMAD Database. The models estimate the embrittlement with sufficient accuracy. All models predict the DBTT and USE based on unseen input data with mean absolute errors of approximately 20 °C and 10 J, respectively. Two of the models can be used to evaluate the importance of the nondestructively measured parameters. In the future, machine learning algorithms could be used to build a tool that uses nondestructively measured parameters to estimate the neutron-induced embrittlement of RPVs on site.

Modeling customer engagement assists a business in identifying the high risk and high potential customers. A way to define high risk and high potential customers in a Software-as-a-Service (SaaS) business is to define them as customers with high potential to churn or upgrade. Identifying the high risk and high potential customers in time can help the business retain and grow revenue. This thesis uses churn and upgrade prediction classifiers to define a customer engagement score for a SaaS business. The classifiers used and compared in the research were logistic regression, random forest and XGBoost. The classifiers were trained using data from the case-company containing customer data such as user count and feature usage. To tackle class imbalance, the models were also trained with oversampled training data. The hyperparameters of each classifier were optimised using grid search. After training the models, performance of the classifiers on a test data was evaluated. In the end, the XGBoost classifiers outperformed the other classifiers in churn prediction. In predicting customer upgrades, the results were more mixed. Feature importances were also calculated, and the results showed that the importances differ for churn and upgrade prediction.

Standard machine learning procedures are based on assumption that training and testing data is sampled independently from identical distributions. Comparative data of traits in biological species breaks this assumption. Data instances are related by ancestry relationships, that is phylogeny. In this study, new machine learning procedures are presented that take into account phylogenetic information when fitting predictive models. Phylogenetic statistics for classification accuracy and error are proposed based on the concept of effective sample size. Versions of perceptron training and KNN classification are built on these metrics. Procedures for regularised PGLS regression, phylogenetic KNN regression, neural network regression and regression trees are presented. Properties of phylogenetic perceptron training and KNN regression are studied with synthetic data. Experiments demonstrate that phylogenetic perceptron training improves robustness when the phylogeny is unbalanced. Regularised PGLS and KNN regression are applied to mammal dental traits and environments to both test the algorithms and gain insights in the relationship of mammal teeth and the environment.

Response Surface Models (RSM) are cheap, reduced complexity, and, usually, statistical models that are fit to the response of more complex models to approximate their outputs with higher computational efficiency. In atmospheric science, there has been a continuous push to reduce the amount of training data required to fit an RSM. With this reduction in costly data gathering, RSMs can be used more ad hoc and quickly adapted to new applications. However, with the decrease in diverse training data, the risk increases that the RSM is eventually used on inputs on which it cannot make a prediction. If there is no indication from the model that its outputs can no longer be trusted, trust in an entire RSM decreases. We present a framework for building prudent RSMs that always output predictions with confidence and uncertainty estimates. We show how confidence and uncertainty can be propagated through downstream analysis such that even predictions on inputs outside the training domain or in areas of high variance can be integrated. Specifically, we introduce the Icarus RSM architecture, which combines an out-of-distribution detector, a prediction model, and an uncertainty quantifier. Icarus-produced predictions and their uncertainties are conditioned on the confidence that the inputs come from the same distribution that the RSM was trained on. We put particular focus on exploring out-of-distribution detection, for which we conduct a broad literature review, design an intuitive evaluation procedure with three easily-visualisable toy examples, and suggest two methodological improvements. We also explore and evaluate popular prediction models and uncertainty quantifiers. We use the one-dimensional atmospheric chemistry transport model SOSAA as an example of a complex model for this thesis. We produce a dataset of model inputs and outputs from simulations of the atmospheric conditions along air parcel trajectories that arrived at the SMEAR II measurement station in Hyytiälä, Finland, in May 2018. We evaluate several prediction models and uncertainty quantification methods on this dataset and construct a proof-of-concept SOSAA RSM using the Icarus RSM architecture. The SOSAA RSM is built on pairwise-difference regression using random forests and an auto-associative out-of-distribution detector with a confidence scorer, which is trained with both the original training inputs and new synthetic out-of-distribution samples. We also design a graphical user interface to configure the SOSAA model and trial the SOSAA RSM. We provide recommendations for out-of-distribution detection, prediction models, and uncertainty quantification based on our exploration of these three systems. We also stress-test the proof-of-concept SOSAA RSM implementation to reveal its limitations for predicting model perturbation outputs and show directions for valuable future research. Finally, our experiments affirm the importance of reporting predictions alongside well-calibrated confidence scores and uncertainty levels so that the predictions can be used with confidence and certainty in scientific research applications.

Bayesian networks (BN) are models that map the mutual dependencies and independencies between a set of variables. The structure of the model can be represented as a directed acyclic graph (DAG), which is a graph where the nodes represent variables and the directed edges between variables represent a dependency. BNs can be either constructed by using knowledge of the system or derived computationally from observational data. Traditionally, BN structure discovery from observational data has been done through heuristic algorithms, but advances in deep learning have made it possible to train neural networks for this task in a supervised manner. This thesis provides an overview of BN structure discovery and discusses the strengths and weaknesses of the emerging supervised paradigm. One supervised method, the EQ-model, that uses neural networks for structure discovery using equivariant models, is also explored in further detail with empirical tests. Through a process of hyperparameter optimisation and moving to online training, the performance of the EQ-model is increased. The EQ-model is still observed to underperform in comparison to a competing score-based model, NOTEARS, but offers convenient features, such as dramatically faster runtime, that compensate for the reduced performance. Several interesting lines of further study that could be used to further improve the performance of the EQ-model are also identified.

In this thesis, we give an overview of current methodology in the field of uncertainty estimation in machine learning, with focus on confidence scores and their calibration. We also present a case study, where we propose a novel method to improve uncertainty estimates of an in-production machine learning model operating in an industrial setting with real-life data. This model is used by a Finnish company Basware to extract information from invoices in the form of machine-readable PDFs. The solution we propose is shown to produce confidence estimates, which outperform the legacy estimates on several relevant metrics, increasing coverage of automated invoices from 65.6% to 73.2% with no increase in error rate.

The aim of this thesis is to predict total career racing performance of Finnish trotter horses by using trotters early career racing performance and other early career variables. This thesis presents a brief introductory of harness racing and horses used in Finnish trotting sport. The data is presented and modified for predictions, with descriptive statistics of tables and visuals. The machine learning method of Random forests for regression is introduced and used in the predictions. After training the model, this thesis presents the prediction accuracy and variables of importance of the predictions of total career racing performance for both Finnhorse trotters and Finnish Standardbred trotter population. Finally, the writer discusses on the shortages and possible improvements for future research. The data for this thesis was provided by The Finnish trotting and breeding association (Suomen Hippos ry), which included all information of harness races from 1984 to the end of 2019, raced in Finland. From almost three million rows, the data was summarised to a data table of 46704 rows of trotters, that have started their career at earliest allowed three age groups. A total of 37 independent variables were used to predict three outcomes of total career earnings, total number of career starts and total number of career first placings, as separate models. The predictors are derived from other studies that estimate the environmental and genetic factors of racing performance of a trotter. The three models performed poor to moderate, with total earnings having the highest prediction accuracy. The model predicted quite well larger amounts of earnings, but was avid to predict some earnings when there in fact were none. Prediction accuracy of total number of starts was poor, especially when the true amount of starts was low. Model that predicted total number of career first placings performed the worst. This can partially be explained by the fact that winning is a rare event for a trotter in general. The models fit better for Finnish Standardbred trotters than for Finnhorse trotters. This thesis works as a good basis for future similar research, where massive amounts of data and machine learning is used to predict trotter’s career, racing performance or other factors. The results show that predicting total career racing performance as a classification problem could be a better fit than regression. These adequate classes, as well as possible better predictors and suitable imputes for missing values, should be consulted with an audience of superior knowledge in harness racing.