Automatic anatomical landmark localization in medical imaging requires not just accurate predictions but reliable uncertainty quantification for effective clinical decision support. Current uncertainty quantification approaches often fall short, particularly when combined with normality assumptions, systematically underestimating total predictive uncertainty. This paper introduces conformal prediction as a framework for reliable uncertainty quantification in anatomical landmark localization, addressing a critical gap in automatic landmark localization. We present two novel approaches guaranteeing finite-sample validity for multi-output prediction: multi-output regression-as-classification conformal prediction (M-R2CCP) and its variant multi-output regression to classification conformal prediction set to region (M-R2C2R). Unlike conventional methods that produce axis-aligned hyperrectangular or ellipsoidal regions, our approaches generate flexible, non-convex prediction regions that better capture the underlying uncertainty structure of landmark predictions. Through extensive empirical evaluation across multiple 2D and 3D datasets, we demonstrate that our methods consistently outperform existing multi-output conformal prediction approaches in both validity and efficiency. This work represents a significant advancement in reliable uncertainty estimation for anatomical landmark localization, providing clinicians with trustworthy confidence measures for their diagnoses. While developed for medical imaging, these methods show promise for broader applications in multi-output regression problems.

Microservice applications are omnipresent due to their advantages, such as scalability, flexibility and consequentially resource cost efficiency. The loosely-coupled microservices can be easily added, replicated, updated and/or removed to address the changing workload. However, the distributed and dynamic nature of microservice architectures introduces a complexity with regard to monitoring and observability, which is paramount to ensure reliability, especially in critical domains. Anomaly detection has become an important tool to automate microservice monitoring and detect system failures. Nevertheless, state-of-the-art solutions assume the topology of the monitored application to remain static over time and fail to account for the dynamic changes the application, and the infrastructure it is deployed on, undergoes. This paper tackles these shortcomings by introducing a context-aware anomaly detection methodology using dynamic knowledge graphs to capture contextual features which describe the evolving state of the monitored system. Our methodology leverages resource and network monitoring to capture dependencies between microservices, and the infrastructure they are running on. In addition to the methodology for anomaly detection, this paper presents an open-source benchmark framework for context-aware anomaly detection that includes monitoring, fault injection and data collection. The evaluation on this benchmark shows that our methodology consistently outperforms the non-contextual baselines. These results underscore the importance of contextual awareness for robust anomaly detection in complex, topology-driven systems. Beyond these achieved improvements, our benchmark establishes a reproducible and extensible foundation for future research, facilitating the experimentation with broader ranges of models and a continued advancement in context-aware anomaly detection.

The inception of the Relational Graph Convolutional Network (R-GCN) marked a milestone in the Semantic Web domain as a widely cited method that generalizes end-to-end hierarchical representation learning to Knowledge Graphs (KGs). R-GCNs generate representations for nodes of interest by repeatedly aggregating parametrized, relation-specific transformations of their neighbors. However, in this work, it is posited that the R-GCN’s main contribution lies in this “message passing” paradigm, rather than the learned weights. To prove this, the “Random Relational Graph Convolutional Network” (RR-GCN) is introduced, which leaves all parameters untrained and thus constructs node embeddings by aggregating randomly transformed random representations from neighbors. Additionally, the advantage offered by learnable parameters for RR-GCN without completely losing the advantages of random transformations is explored. It is empirically shown that RR-GCNs can compete with fully trained R-GCNs in node classification.

Data visualization recommendation aims to assist the user in creating visualizations from a given dataset. The process of creating appropriate visualizations requires expert knowledge of the available data model as well as the dashboard application that is used. To relieve the user from requiring this knowledge and from the manual process of creating numerous visualizations or dashboards, we present a context-aware visualization recommender system (VisCARS) for monitoring applications that automatically recommends a personalized dashboard to the user, based on the system they are monitoring and the task they are trying to achieve. Through a knowledge graph-based approach, expert knowledge about the data and the application is included as contextual features to improve the recommendation process. A dashboard ontology is presented that describes key components in a dashboard ecosystem in order to semantically annotate all the knowledge in the graph. The recommender system leverages knowledge graph embedding and comparison techniques in combination with a context-aware collaborative filtering approach to derive recommendations based on the context, i.e., the state of the monitored system, and the end-user preferences. The proposed methodology is implemented and integrated in a dynamic dashboard solution. The resulting recommender system is evaluated on a smart healthcare use-case through a quantitative performance and scalability analysis as well as a qualitative user study. The results highlight the performance of the proposed solution compared to the state-of-the-art and its potential for time-critical monitoring applications.

Knowledge Graphs (KGs) are gaining popularity and are being widely used in a plethora of applications. They owe their popularity to the fact that KGs are an ideal form to integrate and retrieve data originating from various sources. Using KGs as input for Machine Learning (ML) tasks allows to perform predictions on these popular graph structures. However, KGs cannot directly be used as ML input in their graph representation, they first require to be transformed to a vector space representation through an embedding technique. As ML techniques are data-driven, they can generalize over unseen input data that deviates to some extent from the data they were trained upon. To fully exploit the generalization capabilities of ML algorithms when using embedded KGs as input, small changes in the KGs should also result in small changes in the embedding space. Various embedding techniques for graphs in general exist, however, they have not been tailored towards embedding whole KGs, while KGs can be considered a special kind of graph that adheres to a certain KG schema. This paper evaluates if these existing embedding techniques that embed the whole graphs can represent the similarity between KGs in their embedding space, allowing ML algorithms to generalize over their input. We compare the similarities between KGs in terms of changes in size, entity labels, and KG schema. We found that most techniques were able to represent the similarities in terms of size and entity labels in their embedding space, however, none of the techniques were able to capture the similarities in KG schema.

Artificial Intelligence for IT Operations (AIOps) addresses the rising complexity of cloud computing and Internet of Things by assisting DevOps engineers to monitor and maintain applications. Machine Learning is an essential part of AIOps, enabling it to perform Anomaly Detection and Root Cause Analysis. These techniques are often executed in centralized components, however, which requires transferring vast amounts of data to a central location. This increase in network traffic causes strain on the network and results in higher latency. This paper leverages edge computing to address this issue by deploying ML models closer to the monitored services, reducing the network overhead. This paper investigates two architectural approaches: a sidecar architecture and a federated architecture, and highlights their advantages and shortcomings in different scenarios. Taking this into account, it proposes a framework that orchestrates the deployment and management of distributed edge ML models. Additionally, the paper introduces a Python library to assist data scientists during the development of AIOps techniques and concludes with a thorough evaluation of the resulting framework towards resource consumption and scalability. The results indicate up to 98.3% reduction in network usage depending on the configuration used while maintaining a minimal increase in resource usage at the edge.

Traditionally, neural networks are viewed from the perspective of connected neuron layers represented as matrix multiplications. We propose to compose these weight matrices from a set of orthogonal basis matrices by approaching them as elements of the real matrices vector space under addition and multiplication. Making use of the Kronecker product for vectors, this composition is unified with the singular value decomposition (SVD) of the weight matrix. The orthogonal components of this SVD are trained with a descent curve on the Stiefel manifold using the Cayley transform. Next, update equations for the singular values and initialization routines are derived. Finally, acceleration for stochastic gradient descent optimization using this formulation is discussed. Our proposed method allows more parameter-efficient representations of weight matrices in neural networks. These decomposed weight matrices achieve maximal performance in both standard and more complicated neural architectures. Furthermore, the more parameter-efficient decomposed layers are shown to be less dependent on optimization and better conditioned. As a tradeoff, training time is increased up to a factor of 2. These observations are consequently attributed to the properties of the method and choice of optimization over the manifold of orthogonal matrices.

Water supply companies around the globe are struggling to meet the needs of an ever-increasing population, while climate change contributes to more drought. At the same time, up to 30% of the total amount of treated drinking water in the water supply system is lost due to leaks. An important strategy to reduce leak losses is using hydraulic modeling to localize leaks in a expert-driven manner. In this paper, we present a hybrid leak localization approach combining both hydraulic modeling and machine learning-based classification. A Gaussian Naive Bayes classifier is trained to localize leaks based on simulated pressures and historical pressure measurements. The simulated pressures are obtained using a hydraulic model of the water supply system. In our methodology, learned parameters of the classifier are inferred directly from processing the simulated and measured pressures, without the need for explicit training. We demonstrate the effectiveness of our leak localization approach by using real leak experiments, achieved by opening hydrants at different locations in an operational water supply system. State-of-the-art results are achieved, similar to an approach where explicit training is still needed.

Abstract Feature selection is a crucial step in developing robust and powerful machine learning models. Feature selection techniques can be divided into two categories: filter and wrapper methods. While wrapper methods commonly result in strong predictive performances, they suffer from a large computational complexity and therefore take a significant amount of time to complete, especially when dealing with high-dimensional feature sets. Alternatively, filter methods are considerably faster, but suffer from several other disadvantages, such as (i) requiring a threshold value, (ii) many filter methods not taking into account intercorrelation between features, and (iii) ignoring feature interactions with the model. To this end, we present powershap, a novel wrapper feature selection method, which leverages statistical hypothesis testing and power calculations in combination with Shapley values for quick and intuitive feature selection. Powershap is built on the core assumption that an informative feature will have a larger impact on the prediction compared to a known random feature. Benchmarks and simulations show that powershap outperforms other filter methods with predictive performances on par with wrapper methods while being significantly faster, often even reaching half or a third of the execution time. As such, powershap provides a competitive and quick algorithm that can be used by various models in different domains. Furthermore, powershap is implemented as a plug-and-play and open-source sklearn component, enabling easy integration in conventional data science pipelines. User experience is even further enhanced by also providing an automatic mode that automatically tunes the hyper-parameters of the powershap algorithm, allowing to use the algorithm without any configuration needed.

As companies rely on an ever increasing number of connected devices for their day to day operations, a need arises for automated anomaly detectors to constantly observe crucial device metrics in real time to prevent downtime and data loss. As production environments tend to monitor a huge amount of these metrics, it prevents current state-of-the-art techniques to be deployed as the required computational resources is too high. This paper proposes a lightweight anomaly detection method that can be deployed in these environments without a reduction in accuracy. The approach works fully online, and does not require an extensive history set to be kept in memory. The method is benchmarked on the publicly available Numenta dataset, as well as a network monitoring dataset from different environments provided by a network management solution vendor. These benchmarks show the proposed technique to be very competitive with the current state-of-the-art and exceeding it in production applicability.

Anomalies and faults can be detected, and their causes verified, using both data-driven and knowledge-driven techniques. Data-driven techniques can adapt their internal functioning based on the raw input data but fail to explain the manifestation of any detection. Knowledge-driven techniques inherently deliver the cause of the faults that were detected but require too much human effort to set up. In this paper, we introduce FLAGS, the Fused-AI interpretabLe Anomaly Generation System, and combine both techniques in one methodology to overcome their limitations and optimize them based on limited user feedback. Semantic knowledge is incorporated in a machine learning technique to enhance expressivity. At the same time, feedback about the faults and anomalies that occurred is provided as input to increase adaptiveness using semantic rule mining methods. This new methodology is evaluated on a predictive maintenance case for trains. We show that our method reduces their downtime and provides more insight into frequently occurring problems.

Solenoid valves (SV) are essential components of industrial systems and therefore widely used. As they suffer from high failure rates in the field, fault prognosis of these assets plays a major role for improving their maintenance and reliability. In this work, Bayesian convolutional neural networks are used to predict the remaining useful life (RUL) of SV, by training them on the valve's current signatures. Predictive performance is further improved upon by using salient physical features obtained from an electromechanical model as the network's training input. Results show that our designed network architecture produces well-calibrated uncertainty estimations of the RUL predictive distributions, which is an important concern in prognostic decision-making.

Time series analysis is becoming more popular in both research and industry. One recent innovation is the Ostinato algorithm, which finds the best preserved patterns that are repeated in a collection of series, i.e. consensus motifs and corresponding radii. However, Ostinato only works as a batch algorithm, can only find the top-k patterns, only finds patterns that are repeated in multiple series and has a runtime that depends on the input series and setup parameters. To tackle these limitations, we present two algorithms in this paper that can answer broader questions. First, we created an anytime version of Ostinato, called Anytime Ostinato, that finds the exact consensus radius for each subsequence, i.e. the radius profile, or can estimate these radii in a fraction of the time. Second, we designed a batch algorithm, called Single Series Ostinato, that finds the radius profile for a single series allowing us to detect repeating patterns in a single series, which is not possible for Ostinato. In this paper we explain both algorithms and apply them to the REFIT and PAMAP2 datasets respectively.

This paper contributes to the pursuit of leveraging unstructured medical notes to structured clinical decision making. In particular, we present a pipeline for clinical information extraction from medical notes related to preterm birth, and discuss the main challenges as well as its potential for clinical practice. A large collection of medical notes, created by staff during hospitalizations of patients who were at risk of delivering preterm, was gathered and analyzed. Based on an annotated collection of notes, we trained and evaluated information extraction components to discover clinical entities such as symptoms, events, anatomical sites and procedures, as well as attributes linked to these clinical entities. In a retrospective study, we show that these are highly informative for clinical decision support models that are trained to predict whether delivery is likely to occur within specific time windows, in combination with structured information from electronic health records.

As companies are increasingly measuring their products and services, the amount of time series data is rising and techniques to extract usable information are needed. One recently developed data mining technique for time series is the Matrix Profile. It consists of the smallest z-normalized Euclidean distance of each subsequence of a time series to all other subsequences of another series. It has been used for motif and discord discovery, for segmentation and as building block for other techniques. One side effect of the z-normalization used is that small fluctuations on flat signals are upscaled. This can lead to high and unintuitive distances for very similar subsequences from noisy data. We determined an analytic method to estimate and remove the effects of this noise, adding only a single, intuitive parameter to the calculation of the Matrix Profile. This paper explains our method and demonstrates it by performing discord discovery on the Numenta Anomaly Benchmark and by segmenting the PAMAP2 activity dataset. We find that our technique results in a more intuitive Matrix Profile and provides improved results in both usecases for series containing many flat, noisy subsequences. Since our technique is an extension of the Matrix Profile, it can be applied to any of the various tasks that could be solved by it, improving results where data contains flat and noisy sequences.

In order to minimize operation and maintenance costs and extend the lifetime of rotating machinery, damaging conditions and faults should be detected early and automatically. To enable this, sensor streams should continuously be monitored, processed, and interpreted. In recent years, infrared thermal imaging has gained attention for the said purpose. However, the detection capabilities of a system that uses infrared thermal imaging is limited by the modality captured by this single sensor, as is any single sensor-based system. Hence, within this paper a multisensor system is proposed that not only uses infrared thermal imaging data, but also vibration measurements for automatic condition and fault detection in rotating machinery. It is shown that by combining these two types of sensor data, several conditions/faults and combinations can be detected more accurately than when considering the sensor streams individually.

Abstract Machine clients are increasingly making use of the Web to perform tasks. While Web services traditionally mimic remote procedure calling interfaces, a new generation of so-called hypermedia APIs works through hyperlinks and forms, in a way similar to how people browse the Web. This means that existing composition techniques, which determine a procedural plan upfront, are not sufficient to consume hypermedia APIs, which need to be navigated at runtime. Clients instead need a more dynamic plan that allows them to follow hyperlinks and use forms with a preset goal. Therefore, in this paper, we show how compositions of hypermedia APIs can be created by generic Semantic Web reasoners. This is achieved through the generation of a proof based on semantic descriptions of the APIs' functionality. To pragmatically verify the applicability of compositions, we introduce the notion of pre-execution and post-execution proofs. The runtime interaction between a client and a server is guided by proofs but driven by hypermedia, allowing the client to react to the application's actual state indicated by the server's response. We describe how to generate compositions from descriptions, discuss a computer-assisted process to generate descriptions, and verify reasoner performance on various composition tasks using a benchmark suite. The experimental results lead to the conclusion that proof-based consumption of hypermedia APIs is a feasible strategy at Web scale.

Hyperspectral image (HSI) classification is one of the most widely used methods for scene analysis from hyperspectral imagery. In the past, many different engineered features have been proposed for the HSI classification problem. In this paper, however, we propose a feature learning approach for hyperspectral image classification based on convolutional neural networks (CNNs). The proposed CNN model is able to learn structured features, roughly resembling different spectral band-pass filters, directly from the hyperspectral input data. Our experimental results, conducted on a commonly-used remote sensing hyperspectral dataset, show that the proposed method provides classification results that are among the state-of-the-art, without using any prior knowledge or engineered features.

The ability to automatically determine the road type from sensor data is of great significance for automatic annotation of routes and autonomous navigation of robots and vehicles. In this paper, we present a novel algorithm for content-based road type classification from images. The proposed method learns discriminative features from training data in an unsupervised manner, thus not requiring domain-specific feature engineering. This is an advantage over related road surface classification algorithms which are only able to make a distinction between pre-specified uniform terrains. In order to evaluate the proposed approach, we have constructed a challenging road image dataset of 20,000 samples from real-world road images in the paved and unpaved road classes. Experimental results on this dataset show that the proposed algorithm can achieve state-of-the-art performance in road type classification.

Despite adding emotions to applications has proven to enhance the user experience, emotion recognition applications are still not widely available nor used. Within this paper, emotion recognition is done on Twitter tweets using six emotion classification algorithms that are compared on precision and timing. The paper shows that precision can be enhanced by 5.02% compared to the current state-of-the-art by improving the features. Furthermore, the presented algorithms work in real-time.

This paper proposes a novel multi-sensor fire detection method based on ordinary video images and the amplitude images of a time-of-flight camera. Using this multi-modal in formation, flame regions can be detected very accurately. Regions with high accumulative amplitude differences and high values in all detail images of the amplitude image its discrete wavelet transform, are labeled as candidate flame regions. Simultaneously, moving objects in the visual images are also investigated. Objects which possess the experimentally found low-cost flame features are also labeled as candidate flame region. Finally, if one of the visual and amplitude candidate flame regions overlap, fire alarm is given. Experiments show that the proposed detector has an average flame detection rate of 92% with no false positive detections.

As the complexity of business processes grows, the shift towards workflow-based programming becomes more attractive. The typical long-running characteristic of workflows imposes new challenges such as dynamic adaptation of running workflow instances. Windows Workflow Foundation (in short WF) was released by Microsoft as their solution for workflow-driven application development. Although WF contains features that allow dynamic workflow adaptation, the framework lacks an instrumentation framework to make such adaptations more manageable. Therefore, we built an instrumentation framework that provides more flexibility for applying workflow adaptation batches to workflow instances, both at creation time and during an instance's lifecycle. In this paper we present this workflow instrumentation framework and performance implications caused by dynamic workflow adaptation are detailed.