Aaltodoc Repository :: Browsing by Author "Jung, Alexander"

Browsing by Author "Jung, Alexander"

Now showing 1 - 20 of 80

The Actor-Dueling-Critic Method for Reinforcement Learning
(Multidisciplinary Digital Publishing Institute (MDPI), 2019-04-01) Wu, Menghao; Gao, Yanbin; Jung, Alexander; Zhang, Qiang; Du, Shitong; Department of Computer Science; Helsinki Institute for Information Technology (HIIT); Professorship Jung Alexander; Harbin Engineering University
Model-free reinforcement learning is a powerful and efficient machine-learning paradigm which has been generally used in the robotic control domain. In the reinforcement learning setting, the value function method learns policies by maximizing the state-action value (Q value), but it suffers from inaccurate Q estimation and results in poor performance in a stochastic environment. To mitigate this issue, we present an approach based on the actor-critic framework, and in the critic branch we modify the manner of estimating Q-value by introducing the advantage function, such as dueling network, which can estimate the action-advantage value. The action-advantage value is independent of state and environment noise, we use it as a fine-tuning factor to the estimated Q value. We refer to this approach as the actor-dueling-critic (ADC) network since the frame is inspired by the dueling network. Furthermore, we redesign the dueling network part in the critic branch to make it adapt to the continuous action space. The method was tested on gym classic control environments and an obstacle avoidance environment, and we design a noise environment to test the training stability. The results indicate the ADC approach is more stable and converges faster than the DDPG method in noise environments.
AIR: Aerial Inspection RetinaNet for Land Search and Rescue Missions
(2022-01-24) Pyrrö, Pasi; Naseri, Hassan; Perustieteiden korkeakoulu; Jung, Alexander
Search and rescue (SAR) missions have been carried out for centuries to aid those who are lost or in distress, typically in some remote areas, such as wilderness. With recent advances in technology, small unmanned aerial vehicles or drones have been used during SAR missions for years in many countries. The reason is that drones enable rapid aerial photographing of large areas with potentially difficult-to-reach terrain and can even match several land search parties in efficiency. However, there remains the issue of inspecting a vast amount of aerial drone images for tiny clues about the missing person location, which is currently a manual task done by humans in most cases. It turns out this inspection process is very slow, tedious and error-prone for most people and can significantly delay the entire aerial drone search operation. In this thesis, we propose a novel deep learning based object detection approach to automate this drone footage inspection task. As such, we use a data set called HERIDAL of aerial imagery from Mediterranean landscape to train our detector, and the goal is to outperform existing object detection methods on the HERIDAL test data. Consequently, we experiment with hyperparameter tuning, model architecture selection, online data augmentation, image tiling, confidence score threshold calibration and several other tricks to improve the test performance of our method. Finally, we present Aerial Inspection RetinaNet (AIR) as the outcome of these experiments, which is our solution to this aerial person detection (APD) problem in SAR. Moreover, we demonstrate state-of-the-art results for the AIR detector on the difficult HERIDAL benchmark in terms of both precision (~21 percentage points increase) and speed. In addition, we provide a new formal definition for the APD problem in SAR missions related to the HERIDAL data set. That is, we define a novel evaluation scheme, which ranks detectors in terms of real-world SAR localization requirements, which are much looser than in typical object detection tasks. Moreover, we devise an estimator for average human detection performance via a meta-analysis study, which can be used as an initial baseline for APD method performance. Lastly, we propose a novel bounding box aggregation method for robust, approximate object localization: the merging of overlapping bounding boxes (MOB) algorithm.
Anomaly Detection Algorithms and Techniques for Network Intrusion Detection Systems
(2020-08-18) Mishin, Mikhail; Kahles, Julen; Perustieteiden korkeakoulu; Jung, Alexander
In recent years, many deep learning-based models have been proposed for anomaly detection. This thesis presents a comparison of selected deep autoencoding models and classical anomaly detection methods on three modern network intrusion detection datasets. We experiment with different configurations and architectures of the selected models, as well as aggregation techniques for input preprocessing and output postprocessing. We propose a methodology for creating benchmark datasets for the evaluation of the methods in different settings. We provide a statistical comparison of the performance of the selected techniques. We conclude that the deep autoencoding models, in particular AE and VAE, systematically outperform the classic methods. Furthermore, we show that aggregating input network flow data improves the overall performance. In general, the tested techniques are promising regarding their application in network intrusion detection systems. However, secondary techniques must be employed to reduce the high numbers of generated false alarms.
Anomaly Detection on Osmosis Trades
(2023-10-09) García Gutierrez, Álvaro; Avilés, Rafael; Perustieteiden korkeakoulu; Jung, Alexander
This report focuses on anomaly detection within Osmosis DEX swap transactions, the largest decentralized exchange in the Cosmos ecosystem. The core objectives of this study are two: first, to evaluate the feasibility of detecting outliers in Osmosis DEX transactions using the available data, and second, to assess the real-world impact of these identified anomalies on the proposed market. To achieve these objectives a pipeline, from data indexing and preprocessing to model training and deployment has been designed and implemented. For that, some Big Query tables are created following different approaches depending on the use case and the data sources available, but always ensuring the quality and efficiency of the pipeline. Based on the obtained data, a variety of anomaly detection techniques have been explored, including One-Class SVM, Isolation Forest, and KMeans among other models. After the evaluation, the dense autoencoder has emerged as the most effective approach for detecting anomalies in this specific context. The dense autoencoder has a Silhouette Score of 0.909 when the maximum is 1. However, the true strength of this model arises when assessing the impact of identified outliers on market metrics such as volatility, price evolution and volume. The Mann-Whitney U Test and the Kolmogorov-Smirnov Test have been evaluated, and their results demonstrate the statistical influence of these outliers on the market. Although the main objectives have been achieved, the report concludes by outlining future directions and opportunities for improvement. These include cost optimization in the data pipeline, refinement of evaluation metrics, and further research into factors influencing market behavior.
Anti-Money Laundering system based on customer behavior
(2019-08-19) Torres Porta, Miguel; Jahkola, Olli; Perustieteiden korkeakoulu; Jung, Alexander
Money Laundering is a big problem that concerns many governments and institutions. Vast amounts of money from illicit activities are laundered and go through the financial system. The criminals behind these crimes are still unpunished. In addition to this, the money launderers are improving their techniques to counteract the efforts of the legal institutions that are fighting against these crimes. Those who are working to uncover these crimes should also improve the way to detect money laundering and therefore, to identify suspicious transactions. That is the reason behind the interest of companies to use innovative techniques such as Artificial Intelligence (AI) to improve the accuracy of the current methods used in the industry. The objective of this thesis is to research and develop a system capable of flag changes in the behaviour of the contacts. There are many techniques in the field of AI, concretely in the area of machine learning that can be used to classify the activities in asset management companies. The algorithm used to analyse the behaviour was K-Means. This algorithm can group the data points for a given set of parameters. These features were carefully selected to characterise the behaviour of the investors, flagging the clients who move from one cluster to another. The change might be seen as suspicious behaviour of money laundering. An expert should review this flagged client, and decide whether it is necessary to report the authorities or not. The main reason to use AI is to reduce the number of false positives that are reported to the experts. The purpose is to reduce the manual work needed to identify suspicious transactions, allowing the experts to be free to focus on the main things of their jobs.
Application of Contextual Bandits Models in a Supervised Learning Setting
(2020-08-18) Dikmen, Ceren; Zhao, Yu; Perustieteiden korkeakoulu; Jung, Alexander
Application of machine learning to link click predictions in Facebook Family of Apps advertising
(2021) Seppälä, Joonas; Gloukhotsev, Alexei; Jung, Alexander; Markkinoinnin laitos; Kauppakorkeakoulu; School of Business
Social media networks gather a major share of global marketing advertisement spend whilst simultaneously comprising a major share of time spent online by people with the means to participate in one. The marketing budgets globally are shifting towards digital where social media marketing is one of the key beneficiaries. One of the key metrics used to assess digital marketing effectiveness across all digital marketing mediums is clicks that has been investigated academically from several perspectives. A comparison of different supervised machine learning (ML) algorithms was performed to predict link clicks in Facebook Family of Apps advertising campaigns. Retrospectively collected data was from completed Facebook Family of Apps campaigns over the course of two years between February 2018 until February 2020 targeted to Finland. Four different ML algorithms were tested, including linear regression, decision tree, random forest, and artificial neural network (multi-layer perceptron). Estimation of predictive performance was done using repeated k-fold cross-validation. Comparison between different algorithms was done using root mean square error (RMSE) as the loss function and coefficient of determination (R-squared) as the model fit metric. The relative importances of the features were also estimated from the same cross-validation procedure through algorithm specific approaches. This study approached the prediction problem from the media buyer’s perspective through the use of features that are widely accessable to all the people working within this domain. The test results show that the artificial neural network algorithm performs the best out of the selected algorithms across the three different datasets producing both the smallest loss scores and highest R-squared values. The implications of the findings and suggestions for potential future studies are further discussed.
Applying machine learning methods to predict taxi pickups using historical taxi data
(2022-01-24) Channabasaiah, Akshay; Tian, Yu; Perustieteiden korkeakoulu; Jung, Alexander
Applying Machine Learning to Root Cause Analysis in Agile CI/CD Software Testing Environments
(2019-01-28) Kahles Bastida, Julen; Huuhtanen, Timo; Perustieteiden korkeakoulu; Jung, Alexander
This thesis evaluates machine learning classification and clustering algorithms with the aim of automating the root cause analysis of failed tests in agile software testing environments. The inefficiency of manually categorizing the root causes in terms of time and human resources motivates this work. The development and testing environments of an agile team at Ericsson Finland are used as this work's framework. The author of the thesis extracts relevant features from the raw log data after interviewing the team's testing engineers (human experts). The author puts his initial efforts into clustering the unlabeled data, and despite obtaining qualitative correlations between several clusters and failure root causes, the vagueness in the rest of the clusters leads to the consideration of labeling. The author then carries out a new round of interviews with the testing engineers, which leads to the conceptualization of ground-truth categories for the test failures. With these, the human experts label the dataset accordingly. A collection of artificial neural networks that either classify the data or pre-process it for clustering is then optimized by the author. The best solution comes in the form of a classification multilayer perceptron that correctly assigns the failure category to new examples, on average, 88.9\% of the time. The primary outcome of this thesis comes in the form of a methodology for the extraction of expert knowledge and its adaptation to machine learning techniques for test failure root cause analysis using test log data. The proposed methodology constitutes a prototype or baseline approach towards achieving this objective in a corporate environment.
Applying supervised machine learning for predicting maximal heart rate during aerobic exercise
(2024-05-23) Hirvonen, Sini; Jung, Alexander; Sähkötekniikan korkeakoulu; Turunen, Markus
Machine learning (ML) is a rapidly advancing field of study, with applications already reaching health and sport domains. The popularity of wearable activity monitors and increased data collection enable diverse applications. ML can contribute to a more informed, safe, and personalised approach to health, fitness and performance optimisation. This bachelor’s thesis examines the utilisation of ML in analysing data collected during aerobic exercise. The thesis investigates the use of supervised ML methods to predict maximal heart (HR$_{max}$) rate during aerobic exercise from external weather-related variables. The two methods compared were linear regression and random forest regression. The performance of these methods was evaluated by using k-fold cross-validation, and by examining the difference between the true and predicted values of HR$_{max}$. The used loss functions were mean squared error (MSE) and mean absolute error (MAE). Based on the obtained validation errors, random forest regression proved to be the better performing model. The computed test errors were 44,09 bpm with MSE, and 5,19 bpm with MAE. Of the weather-related features, air temperature affected the predictions most. Based on the obtained results, the reliability of the implemented ML model for predicting HR$_{max}$ is insufficient for practical applications. In addition, the results cannot be used to comprehensively explain the effect of weather conditions on heart rate response during aerobic exercise. However, the prediction accuracy and understanding regarding personalised health monitoring could be improved by developing the model further. Therefore, a working ML model could possibly enhance both safety and performance optimisation while prioritising an individual’s health.
Automatic Job Skill Taxonomy Generation For Recruitment Systems
(2019-06-17) Baad, Dipika; Auvinen, Tapio; Perustieteiden korkeakoulu; Jung, Alexander
The goal of this thesis is to optimize the job recommendation systems by automatically extracting the skills from the job descriptions. With rapid development in technology, new skills are continuously required. This makes the skill tagging of the job descriptions a more difficult problem since a simple keyword match from an already generated skill list is not suitable. A way of automatically populating the skills list to improve the job search engines is needed. This thesis focuses on solving this problem with the help of natural language processing and neural networks. Automatic detection of skills in the unstructured job description dataset is a complex problem as it involves being robust to the ambiguity of natural language and adapting to words not seen in the historical data. This thesis solves this problem by using recurrent neural network models for capturing the context of the skill words. Based on the context captured, the new system is capable of predicting if the word in the given text is a skill or not. Neural network models like Long short-term memory and Bi-directional Long short-term memory are used to capture the long term dependencies in the sentence to identify skills present in the job descriptions. Various natural language processing techniques were utilized to improve the input feature quality to the model. Results obtained from using context before and after the skill words have shown the best results in identifying skills from textual data. This can be applied to capture skills data from job ads as well as it can be extended to extract the skill features from resume data to improve the job recommendation results in the future.
Automating Root Cause Analysis via Machine Learning in Agile Software Testing Environments
(2019-04-01) Kahles, Julen; Torronen, Juha; Huuhtanen, Timo; Jung, Alexander; Department of Computer Science; Professorship Jung Alexander; Helsinki Institute for Information Technology (HIIT); Ericsson Finland
We apply machine learning to automate the root cause analysis in agile software testing environments. In particular, we extract relevant features from raw log data after interviewing testing engineers (human experts). Initial efforts are put into clustering the unlabeled data, and despite obtaining weak correlations between several clusters and failure root causes, the vagueness in the rest of the clusters leads to the consideration of labeling. A new round of interviews with the testing engineers leads to the definition of five ground-truth categories. Using manually labeled data, we train artificial neural networks that either classify the data or pre-process it for clustering. The resulting method achieves an accuracy of 88.9%. The methodology of this paper serves as a prototype or baseline approach for the extraction of expert knowledge and its adaptation to machine learning techniques for root cause analysis in agile environments.
Binary Classification of Histopathological Images by Pre-Trained modified VGG16 model
(2022-04-15) Lamoureux, Maxim; Jung, Alexander; Kemiantekniikan korkeakoulu; Hummel, Michael
Building Information Modeling Connection Recommendation Based on Machine Learning Using Multimodal Information
(2023-08-21) Zhou, Zixin; Schmidt, Fabian; Filoche, Pascal; Perustieteiden korkeakoulu; Jung, Alexander
The increasing complexity of construction projects gives rise to the need for an efficient way of designing, managing, and maintaining structures. Building Information Modeling (BIM) facilitates these processes by providing a digital representation of physical structures. Tekla Structures (TS) has emerged as a popular building information modeling software for structural design. In structural engineering, connections play an important role in joining various building objects. However, the efficient and accurate design of connections in TS remains a challenge due to the wide range of available connection types. Existing solutions for connection recommendation often rely on predefined rules, limiting their applicability and requiring time-consuming setup. Recent research has explored machine learning approaches for connection recommendation, but they suffer from scalability issues or high computational costs. This thesis addresses the connection type recommendation problem in TS as a classification task, leveraging the diverse representations of the BIM objects, including 2D images and attributes. This thesis improves existing approaches for single modality data, comparing XGBoost with random forest for attributes, while enhancing the previous CNN model for image classification. Furthermore, this thesis investigates the potential of combining images and attribute data for connection type classification, using two multimodal data fusion strategies: late fusion and intermediate fusion. The results show that XGBoost with metadata of the attribute dataset yields the best performance, with a maximum accuracy of 0.9283, and the experimented multimodal data fusion methods are unable to further optimise the classification results. The accuracy of attribute-based methods is improved by up to 0.6 percent. The improvement in CNN model can enhance the classification accuracy by up to 5 percent. By comparing various data sources and approaches, this thesis aims to provide a practical connection recommendation design, thereby laying a foundation for better connection design processes in construction projects.
Classification of Purchase Invoices to Analytic Accounts with Machine Learning
(2023-01-23) Johansson, Samuel; Heino, Petri; Sähkötekniikan korkeakoulu; Jung, Alexander
The process of analytic account selection for invoices is time-consuming and expensive. These problems can be diminished with an automated system for selecting the analytic accounts on the invoices. The aim of this thesis is to implement a software system for classifying purchase invoice accounting lines into analytic account in Odoo ERP system. The proposed software uses machine learning methodology as analytic account selection is essentially a multi-class text classification problem. The proposed accuracy goal set is 90% as it is an indicated threshold in a related invoice account selection problem. The thesis investigates the current text classification algorithms and compares them empirically with each other. The selected algorithms are Multinomial Naive Bayes, Logistic Regression, Linear and Non-linear Support Vector Machines (SVM), k-Nearest Neighbor, Decision Tree, Random Forest, Perceptron and Multilayer Perceptrons (MLP). Additionally a primitive Odoo module is created to use this software system in an Odoo server for automatic analytic account selection. The features of the data contain textual descriptions of an invoice line which require the text data preprocessing before transferred into the models. The results indicate the best performing models are Logistic Regression, linear SVM and MLP where the linear SVM is the most effective overall. This model is optimized even further where a test instance reached an accuracy of 88% at best. However empirically the macro averages (precision, recall and F1-score) for an average instance is around 0.80. The thesis presents a functional Odoo module where the training and prediction of analytic accounts can be done within an Odoo server. The Odoo user interface allows to train a model, show the scores of different classes and predict analytic accounts for lines.
Classifying Big Data over Networks Via the Logistic Network Lasso
(2019-02-19) Ambos, Henrik; Tran, Nguyen; Jung, Alexander; Department of Computer Science; Matthews, Michael B.; Professorship Jung Alexander; Helsinki Institute for Information Technology (HIIT); Department of Computer Science
We apply network Lasso to solve binary classification and clustering problems on network structured data. In particular we generalize ordinary logistic regression to non-Euclidean data defined over a complex network structure. The resulting logistic network Lasso classifier amounts to solving a convex optimization problem. A scalable classification algorithm is obtained by applying the alternating direction methods of multipliers.
Classifying Partially Labeled Networked Data VIA Logistic Network Lasso
(2020-05) Tran, Nguyen; Ambos, Henrik; Jung, Alexander; Department of Computer Science; Professorship Jung Alexander; Helsinki Institute for Information Technology (HIIT); Department of Computer Science
We apply the network Lasso to classify partially labeled data points which are characterized by high-dimensional feature vectors. In order to learn an accurate classifier from limited amounts of labeled data, we borrow statistical strength, via an intrinsic network structure, across the dataset. The resulting logistic network Lasso amounts to a regularized empirical risk minimization problem using the total variation of a classifier as a regularizer. This minimization problem is a nonsmooth convex optimization problem which we solve using a primal-dual splitting method. This method is appealing for big data applications as it can be implemented as a highly scalable message passing algorithm.
Classifying Restaurant Menu Items With Supervised Learning
(2021-01-25) Kokkonen, Teemu; Fedintsev, Alexander; Perustieteiden korkeakoulu; Jung, Alexander
Online food delivery is a rapidly growing business. The large selection of restaurants in these platforms poses a challenge as the individual dishes are described in the natural text and might not be tagged or categorized for machine use. This makes it difficult for the items to be used in platform-wide filtering, analysis, marketing, or search. This thesis uses a text classification model to classify dishes based on the natural text descriptions of restaurant dishes. The problem is explored using a dataset of restaurant menus from Wolt, one of the largest food delivery platforms in Finland. Dishes are classified using a multi-label approach using 20 different pre-determined labels such as pasta, dessert, and alcohol. A dataset involving 20 different labels is created for a supervised learning model. Machine learning pipeline is created and the performance of classifiers (support vector machines with different kernels, random forest classifier, multinomial Naive Bayes, linear regression, logistic regression, multi-layer perceptron neural network classifiers) is compared. The support vector machine with a linear kernel has the best subset accuracy (63.2%), F-measure (0.89), and recall (84.8%) scores among the tested classifiers. Multi-layer perceptron classifier is also performing well with an F-measurement of 0.867. Other classifiers such as random decision forest, logistic regression, or multinomial Naive Bayes were not as performant. Using pre-trained Word2Vec embeddings was not helpful in classification.
Clustering fitness tracker data to correct interdevice differences in energy expenditure estimation
(2023-05-15) Tikkanen, Riku; Jung, Alexander; Perustieteiden korkeakoulu; Jung, Alexander
An active lifestyle can prevent many chronic illnesses which has a positive impact not only on individuals but also on societies as a whole. With health care costs increasing, it is understandable that the promotion of a healthy lifestyle is beneficial for many parties. The promotion of activeness has been studied in the past and results suggest that with the help of wearables, people can be motivated to be more active. Unfortunately, systems built for physical activity promotion are built to work with only one specific brand and model of wearable because of inter-device differences. More precisely, different brands and models of wearables give different energy expenditure estimates for similar exercises and therefore results may vary. The primary goal of this thesis was to find out if clustering methods could be used to correct the differences in energy expenditure estimates from different devices for similar activities. If possible, the method could be used to build future activity recommendation systems that work with any device. The suggested procedure would be to cluster similar activities from different wearables in order to calculate an average energy expenditure estimate. The results suggest that the method could be plausible only if the distributions of users for different wearables are similar. While theoretically possible, it would require a lot of time and effort to gather enough users for all considered wearables.
A Comparative Analysis of Graph Signal Recovery Methods for Big Data Networks
(2017-10-23) Mara, Alexandru; Jung, Alexander; Perustieteiden korkeakoulu; Jung, Alexander
Graph signal processing, signal recovery, semi-supervised learning, traGraph-based signal recovery (GSR) techniques have been successfully used in different domains for labelling complete graphs from partial subsets of given labels. Much research has been devoted to finding new efficient approaches for solving this learning problem. However, we have identified a lack of research in empirically comparing different GSR methods on big data graphs. In this work, we implement highly scalable versions of five state-of-the-art methods, which we benchmark under identical conditions on a number of real and synthetic datasets. We perform a comprehensive evaluation of these methods in terms of accuracy, scalability, robustness to noise and graph topology as well as sampling set selection. We find that recently proposed methods based on TV minimization outperform more classical approaches that measure the graphs smoothness through the quadratic form. We draw other interesting conclusions and discuss merits and faults of each of the methods studied.

Browsing by Author "Jung, Alexander"

Results Per Page

Sort Options