Andreas Hotho

Konstantin Kobs, Michael Steininger, Andrzej Dulny et al.

Recent studies have shown that many deep metric learning loss functions perform very similarly under the same experimental conditions. One potential reason for this unexpected result is that all losses let the network focus on similar image regions or properties. In this paper, we investigate this by conducting a two-step analysis to extract and compare the learned visual features of the same model architecture trained with different loss functions: First, we compare the learned features on the pixel level by correlating saliency maps of the same input images. Second, we compare the clustering of embeddings for several image properties, e.g. object color or illumination. To provide independent control over these properties, photo-realistic 3D car renders similar to images in the Cars196 dataset are generated. In our analysis, we compare 14 pretrained models from a recent study and find that, even though all models perform similarly, different loss functions can guide the model to learn different features. We especially find differences between classification and ranking based losses. Our analysis also shows that some seemingly irrelevant properties can have significant influence on the resulting embedding. We encourage researchers from the deep metric learning community to use our methods to get insights into the features learned by their proposed methods.

Julian Tritscher, Fabian Gwinner, Daniel Schlör et al.

Recent estimates report that companies lose 5% of their revenue to occupational fraud. Since most medium-sized and large companies employ Enterprise Resource Planning (ERP) systems to track vast amounts of information regarding their business process, researchers have in the past shown interest in automatically detecting fraud through ERP system data. Current research in this area, however, is hindered by the fact that ERP system data is not publicly available for the development and comparison of fraud detection methods. We therefore endeavour to generate public ERP system data that includes both normal business operation and fraud. We propose a strategy for generating ERP system data through a serious game, model a variety of fraud scenarios in cooperation with auditing experts, and generate data from a simulated make-to-stock production company with multiple research participants. We aggregate the generated data into ready to used datasets for fraud detection in ERP systems, and supply both the raw and aggregated data to the general public to allow for open development and comparison of fraud detection approaches on ERP system data.

Andrzej Dulny, Andreas Hotho, Anna Krause

Previous work on learning physical systems from data has focused on high-resolution grid-structured measurements. However, real-world knowledge of such systems (e.g. weather data) relies on sparsely scattered measuring stations. In this paper, we introduce a novel simulated benchmark dataset, DynaBench, for learning dynamical systems directly from sparsely scattered data without prior knowledge of the equations. The dataset focuses on predicting the evolution of a dynamical system from low-resolution, unstructured measurements. We simulate six different partial differential equations covering a variety of physical systems commonly used in the literature and evaluate several machine learning models, including traditional graph neural networks and point cloud processing models, with the task of predicting the evolution of the system. The proposed benchmark dataset is expected to advance the state of art as an out-of-the-box easy-to-use tool for evaluating models in a setting where only unstructured low-resolution observations are available. The benchmark is available at https://anonymous.4open.science/r/code-2022-dynabench/.

Konstantin Kobs, Michael Steininger, Andreas Hotho

Common Deep Metric Learning (DML) datasets specify only one notion of similarity, e.g., two images in the Cars196 dataset are deemed similar if they show the same car model. We argue that depending on the application, users of image retrieval systems have different and changing similarity notions that should be incorporated as easily as possible. Therefore, we present Language-Guided Zero-Shot Deep Metric Learning (LanZ-DML) as a new DML setting in which users control the properties that should be important for image representations without training data by only using natural language. To this end, we propose InDiReCT (Image representations using Dimensionality Reduction on CLIP embedded Texts), a model for LanZ-DML on images that exclusively uses a few text prompts for training. InDiReCT utilizes CLIP as a fixed feature extractor for images and texts and transfers the variation in text prompt embeddings to the image embedding space. Extensive experiments on five datasets and overall thirteen similarity notions show that, despite not seeing any images during training, InDiReCT performs better than strong baselines and approaches the performance of fully-supervised models. An analysis reveals that InDiReCT learns to focus on regions of the image that correlate with the desired similarity notion, which makes it a fast to train and easy to use method to create custom embedding spaces only using natural language.

Vincenzo Perri, Lisi Qarkaxhija, Albin Zehe et al.

Natural Language Processing and Machine Learning have considerably advanced Computational Literary Studies. Similarly, the construction of co-occurrence networks of literary characters, and their analysis using methods from social network analysis and network science, have provided insights into the micro- and macro-level structure of literary texts. Combining these perspectives, in this work we study character networks extracted from a text corpus of J.R.R. Tolkien's Legendarium. We show that this perspective helps us to analyse and visualise the narrative style that characterises Tolkien's works. Addressing character classification, embedding and co-occurrence prediction, we further investigate the advantages of state-of-the-art Graph Neural Networks over a popular word embedding method. Our results highlight the large potential of graph learning in Computational Literary Studies.

Christian Rack, Konstantin Kobs, Tamara Fernando et al.

Various machine learning approaches have proven to be useful for user verification and identification based on motion data in eXtended Reality (XR). However, their real-world application still faces significant challenges concerning versatility, i.e., in terms of extensibility and generalization capability. This article presents a solution that is both extensible to new users without expensive retraining, and that generalizes well across different sessions, devices, and user tasks. To this end, we developed a similarity-learning model and pretrained it on the "Who Is Alyx?" dataset. This dataset features a wide array of tasks and hence motions from users playing the VR game "Half-Life: Alyx". In contrast to previous works, we used a dedicated set of users for model validation and final evaluation. Furthermore, we extended this evaluation using an independent dataset that features completely different users, tasks, and three different XR devices. In comparison with a traditional classification-learning baseline, our model shows superior performance, especially in scenarios with limited enrollment data. The pretraining process allows immediate deployment in a diverse range of XR applications while maintaining high versatility. Looking ahead, our approach paves the way for easy integration of pretrained motion-based identification models in production XR systems.

Konstantin Kobs, Andreas Hotho

Deep Metric Learning trains a neural network to map input images to a lower-dimensional embedding space such that similar images are closer together than dissimilar images. When used for item retrieval, a query image is embedded using the trained model and the closest items from a database storing their respective embeddings are returned as the most similar items for the query. Especially in product retrieval, where a user searches for a certain product by taking a photo of it, the image background is usually not important and thus should not influence the embedding process. Ideally, the retrieval process always returns fitting items for the photographed object, regardless of the environment the photo was taken in. In this paper, we analyze the influence of the image background on Deep Metric Learning models by utilizing five common loss functions and three common datasets. We find that Deep Metric Learning networks are prone to so-called background bias, which can lead to a severe decrease in retrieval performance when changing the image background during inference. We also show that replacing the background of images during training with random background images alleviates this issue. Since we use an automatic background removal method to do this background replacement, no additional manual labeling work and model changes are required while inference time stays the same. Qualitative and quantitative analyses, for which we introduce a new evaluation metric, confirm that models trained with replaced backgrounds attend more to the main object in the image, benefitting item retrieval systems.

Padraig Davidson, Michael Steininger, André Huhn et al.

Time series are ubiquitous and therefore inherently hard to analyze and ultimately to label or cluster. With the rise of the Internet of Things (IoT) and its smart devices, data is collected in large amounts any given second. The collected data is rich in information, as one can detect accidents (e.g. cars) in real time, or assess injury/sickness over a given time span (e.g. health devices). Due to its chaotic nature and massive amounts of datapoints, timeseries are hard to label manually. Furthermore new classes within the data could emerge over time (contrary to e.g. handwritten digits), which would require relabeling the data. In this paper we present SuSL4TS, a deep generative Gaussian mixture model for semi-unsupervised learning, to classify time series data. With our approach we can alleviate manual labeling steps, since we can detect sparsely labeled classes (semi-supervised) and identify emerging classes hidden in the data (unsupervised). We demonstrate the efficacy of our approach with established time series classification datasets from different domains.

Paul Heinisch, Andrzej Dulny, Anna Krause et al.

Modeling data obtained from dynamical systems has gained attention in recent years as a challenging task for machine learning models. Previous approaches assume the measurements to be distributed on a grid. However, for real-world applications like weather prediction, the observations are taken from arbitrary locations within the spatial domain. In this paper, we propose TaylorPDENet - a novel machine learning method that is designed to overcome this challenge. Our algorithm uses the multidimensional Taylor expansion of a dynamical system at each observation point to estimate the spatial derivatives to perform predictions. TaylorPDENet is able to accomplish two objectives simultaneously: accurately forecast the evolution of a complex dynamical system and explicitly reconstruct the underlying differential equation describing the system. We evaluate our model on a variety of advection-diffusion equations with different parameters and show that it performs similarly to equivalent approaches on grid-structured data while being able to process unstructured data as well.

Maximilian Wolf, Dieter Landes, Andreas Hotho et al.

The detection of cyber-attacks in computer networks is a crucial and ongoing research challenge. Machine learning-based attack classification offers a promising solution, as these models can be continuously updated with new data, enhancing the effectiveness of network intrusion detection systems (NIDS). Unlike binary classification models that simply indicate the presence of an attack, multi-class models can identify specific types of attacks, allowing for more targeted and effective incident responses. However, a significant drawback of these classification models is their sensitivity to imbalanced training data. Recent advances suggest that generative models can assist in data augmentation, claiming to offer superior solutions for imbalanced datasets. Classical balancing methods, although less novel, also provide potential remedies for this issue. Despite these claims, a comprehensive comparison of these methods within the NIDS domain is lacking. Most existing studies focus narrowly on individual methods, making it difficult to compare results due to varying experimental setups. To close this gap, we designed a systematic framework to compare classical and generative resampling methods for class balancing across multiple popular classification models in the NIDS domain, evaluated on several NIDS benchmark datasets. Our experiments indicate that resampling methods for balancing training data do not reliably improve classification performance. Although some instances show performance improvements, the majority of results indicate decreased performance, with no consistent trend in favor of a specific resampling technique enhancing a particular classifier.

Elisabeth Fischer, Albin Zehe, Andreas Hotho et al.

Analyzing sequences of interactions between users and items, sequential recommendation models can learn user intent and make predictions about the next item. Next to item interactions, most systems also have interactions with what we call non-item pages: these pages are not related to specific items but still can provide insights into the user's interests, as, for example, navigation pages. We therefore propose a general way to include these non-item pages in sequential recommendation models to enhance next-item prediction. First, we demonstrate the influence of non-item pages on following interactions using the hypotheses testing framework HypTrails and propose methods for representing non-item pages in sequential recommendation models. Subsequently, we adapt popular sequential recommender models to integrate non-item pages and investigate their performance with different item representation strategies as well as their ability to handle noisy data. To show the general capabilities of the models to integrate non-item pages, we create a synthetic dataset for a controlled setting and then evaluate the improvements from including non-item pages on two real-world datasets. Our results show that non-item pages are a valuable source of information, and incorporating them in sequential recommendation models increases the performance of next-item prediction across all analyzed model architectures.

Tobias Koopmann, Jan Pfister, André Markus et al.

Analyzing, understanding, and describing human behavior is advantageous in different settings, such as web browsing or traffic navigation. Understanding human behavior naturally helps to improve and optimize the underlying infrastructure or user interfaces. Typically, human navigation is represented by sequences of transitions between states. Previous work suggests to use hypotheses, representing different intuitions about the navigation to analyze these transitions. To mathematically grasp this setting, first-order Markov chains are used to capture the behavior, consequently allowing to apply different kinds of graph comparisons, but comes with the inherent drawback of losing information about higher-order dependencies within the sequences. To this end, we propose to analyze entire sequences using autoregressive language models, as they are traditionally used to model higher-order dependencies in sequences. We show that our approach can be easily adapted to model different settings introduced in previous work, namely HypTrails, MixedTrails and even SubTrails, while at the same time bringing unique advantages: 1. Modeling higher-order dependencies between state transitions, while 2. being able to identify short comings in proposed hypotheses, and 3. naturally introducing a unified approach to model all settings. To show the expressiveness of our approach, we evaluate our approach on different synthetic datasets and conclude with an exemplary analysis of a real-world dataset, examining the behavior of users who interact with voice assistants.

Konstantin Kobs, Michael Steininger, Albin Zehe et al.

One common loss function in neural network classification tasks is Categorical Cross Entropy (CCE), which punishes all misclassifications equally. However, classes often have an inherent structure. For instance, classifying an image of a rose as "violet" is better than as "truck". We introduce SimLoss, a drop-in replacement for CCE that incorporates class similarities along with two techniques to construct such matrices from task-specific knowledge. We test SimLoss on Age Estimation and Image Classification and find that it brings significant improvements over CCE on several metrics. SimLoss therefore allows for explicit modeling of background knowledge by simply exchanging the loss function, while keeping the neural network architecture the same. Code and additional resources can be found at https://github.com/konstantinkobs/SimLoss.

Muhammad Usman Butt, Andreas Hotho, Daniel Schlör

Classic Network Intrusion Detection Systems (NIDS) often rely on manual feature engineering to extract meaningful patterns from network traffic data. However, this approach requires domain expertise and runs counter to the widely adopted principle of modern machine learning and neural networks: that models themselves should learn meaningful representations directly from data. We investigate whether tabular representation learning techniques can improve intrusion detection performance by automatically learning robust feature representations for NetFlow data. This paper presents a systematic evaluation of state-of-the-art representation learning methods on benchmark NetFlow datasets, comparing against traditional autoencoders and end-to-end transformer baselines. We evaluate learned representations using both supervised classifiers and unsupervised anomaly detectors, with comprehensive hyperparameter exploration for each combination. Our results reveal strong dataset-model dependency, with no single approach consistently dominating across all scenarios. For supervised classification, TabICL achieves the best performance on CIDDS, while autoencoders follow closely and tie with end-to-end transformer models for the best average rank across datasets. Supervised approaches substantially outperform unsupervised anomaly detection methods, where no single combination consistently dominates as optimal choices depend on the dataset. Cross-dataset transfer experiments demonstrate that learned representations can generalize across network environments with appropriate method and classifier selection. However, transfer performance varies substantially depending on the source-target dataset combination, indicating sensitivity to distributional differences between network environments.

Julia Wunderle, Anton Ehrmanntraut, Jan Pfister et al.

Encoders remain essential for efficient German NLP and NLU scenarios despite the rise of decoder-only LLMs. This work studies two routes to high-quality German encoders under identical data and training constraints: 1) training from scratch and 2) converting decoders via LLM2Vec. We introduce two resources: ModernGBERT (134M, 1B), fully transparent German encoders in the ModernBERT style, and LLäMmleinVec (120M, 1B, 7B), decoder-to-encoder conversions trained with masked next-token prediction, both undergoing a context extension to 8.192 tokens. Across SuperGLEBer, ModernGBERT 1B sets a new state of the art (avg 0.808), surpassing GBERT Large (+4%) and the seven-times larger converted 7B model (0.787). On German MTEB after supervised fine-tuning, ModernGBERT 1B (0.551) approaches the converted 7B model (0.557). We release all models, checkpoints, datasets, and full training records, and introduce an encoder-adapted QA-NIAH evaluation. All in all, our results provide actionable guidance: when parameter efficiency and latency matter, from-scratch encoders dominate. When a pre-trained decoder exists and compute is a limited, conversion offers an effective alternative. ModernGBERT and LLäMmleinVec, including all code, data and intermediary checkpoints are published under a research-only RAIL license.

Jan Pfister, Julia Wunderle, Andreas Hotho · meta-ai

We create two German-only decoder models, LLäMmlein 120M and 1B, transparently from scratch and publish them, along with the training data, for the German NLP research community to use. The model training involved several key steps, including extensive data preprocessing, the creation of a custom German tokenizer, the training itself, as well as the evaluation of the final models on various benchmarks. Throughout the training process, multiple checkpoints were saved and analyzed using the SuperGLEBer benchmark to monitor the models' learning dynamics. Compared to state-of-the-art models on the SuperGLEBer benchmark, both LLäMmlein models performed competitively, consistently matching or surpassing models with similar parameter sizes. The results show that the models' quality scales with size as expected, but performance improvements on some tasks plateaued early, offering valuable insights into resource allocation for future model development.

Andrzej Dulny, Paul Heinisch, Andreas Hotho et al.

Predicting the evolution of spatiotemporal physical systems from sparse and scattered observational data poses a significant challenge in various scientific domains. Traditional methods rely on dense grid-structured data, limiting their applicability in scenarios with sparse observations. To address this challenge, we introduce GrINd (Grid Interpolation Network for Scattered Observations), a novel network architecture that leverages the high-performance of grid-based models by mapping scattered observations onto a high-resolution grid using a Fourier Interpolation Layer. In the high-resolution space, a NeuralPDE-class model predicts the system's state at future timepoints using differentiable ODE solvers and fully convolutional neural networks parametrizing the system's dynamics. We empirically evaluate GrINd on the DynaBench benchmark dataset, comprising six different physical systems observed at scattered locations, demonstrating its state-of-the-art performance compared to existing models. GrINd offers a promising approach for forecasting physical systems from sparse, scattered observational data, extending the applicability of deep learning methods to real-world scenarios with limited data availability.

Florian Gallusser, Simon Hentschel, Anna Krause et al.

Deep Learning models have achieved state-of-the-art performance in medium-range weather prediction but often fail to maintain physically consistent rollouts beyond 14 days. In contrast, a few atmospheric models demonstrate stability over decades, though the key design choices enabling this remain unclear. This study quantitatively compares the long-term stability of three prominent DL-MWP architectures - FourCastNet, SFNO, and ClimaX - trained on ERA5 reanalysis data at 5.625° resolution. We systematically assess the impact of autoregressive training steps, model capacity, and choice of prognostic variables, identifying configurations that enable stable 10-year rollouts while preserving the statistical properties of the reference dataset. Notably, rollouts with SFNO exhibit the greatest robustness to hyperparameter choices, yet all models can experience instability depending on the random seed and the set of prognostic variables

Andrzej Dulny, Farzad Jabbarigargari, Andreas Hotho et al.

We propose a 3D U-Net model to predict the spatial distribution of electromagnetic fields inside a radio-frequency (RF) coil with a subject present, using the phase, amplitude, and position of the coils, along with the density, permittivity, and conductivity of the surrounding medium as inputs. To improve accuracy, we introduce a physics-augmented variant, U-Net Phys, which incorporates Gauss's law of magnetism into the loss function using finite differences. We train our models on electromagnetic field simulations from CST Studio Suite for an eight-channel dipole array RF coil at 7T MRI. Experimental results show that U-Net Phys significantly outperforms the standard U-Net, particularly in predicting fields within the subject, demonstrating the advantage of integrating physical constraints into deep learning-based field prediction.

Melanie Schaller, Daniel Schlör, Andreas Hotho

External influences such as traffic and environmental factors induce vibrations in structures, leading to material degradation over time. These vibrations result in cracks due to the material's lack of plasticity compromising structural integrity. Detecting such damage requires the installation of vibration sensors to capture the internal dynamics. However, distinguishing relevant eigenmodes from external noise necessitates the use of Deep Learning models. The detection of changes in eigenmodes can be used to anticipate these shifts in material properties and to discern between normal and anomalous structural behavior. Eigenmodes, representing characteristic vibration patterns, provide insights into structural dynamics and deviations from expected states. Thus, we propose ModeConv to automatically capture and analyze changes in eigenmodes, facilitating effective anomaly detection in structures and material properties. In the conducted experiments, ModeConv demonstrates computational efficiency improvements, resulting in reduced runtime for model calculations. The novel ModeConv neural network layer is tailored for temporal graph neural networks, in which every node represents one sensor. ModeConv employs a singular value decomposition based convolutional filter design for complex numbers and leverages modal transformation in lieu of Fourier or Laplace transformations in spectral graph convolutions. We include a mathematical complexity analysis illustrating the runtime reduction.

Pascal Janetzky, Florian Gallusser, Simon Hentschel et al.

Accurate vegetation models can produce further insights into the complex interaction between vegetation activity and ecosystem processes. Previous research has established that long-term trends and short-term variability of temperature and precipitation affect vegetation activity. Motivated by the recent success of Transformer-based Deep Learning models for medium-range weather forecasting, we adapt the publicly available pre-trained FourCastNet to model vegetation activity while accounting for the short-term dynamics of climate variability. We investigate how the learned global representation of the atmosphere's state can be transferred to model the normalized difference vegetation index (NDVI). Our model globally estimates vegetation activity at a resolution of \SI{0.25}{\degree} while relying only on meteorological data. We demonstrate that leveraging pre-trained weather models improves the NDVI estimates compared to learning an NDVI model from scratch. Additionally, we compare our results to other recent data-driven NDVI modeling approaches from machine learning and ecology literature. We further provide experimental evidence on how much data and training time is necessary to turn FourCastNet into an effective vegetation model. Code and models will be made available upon publication.

Albin Zehe, Julian Schröter, Andreas Hotho

Suspense is an important tool in storytelling to keep readers engaged and wanting to read more. However, it has so far not been studied extensively in Computational Literary Studies. In this paper, we focus on one of the elements authors can use to build up suspense: dangerous situations. We introduce a corpus of texts annotated with dangerous situations, distinguishing between 7 types of danger. Additionally, we annotate parts of the text that describe fear experienced by a character, regardless of the actual presence of danger. We present experiments towards the automatic detection of these situations, finding that unsupervised baseline methods can provide valuable signals for the detection, but more complex methods are necessary for further analysis. Not unexpectedly, the description of danger and fear often relies heavily on the context, both local (e.g., situations where danger is only mentioned, but not actually present) and global (e.g., "storm" being used in a literal sense in an adventure novel, but metaphorically in a romance novel).

Christian Schell, Andreas Hotho, Marc Erich Latoschik

Reliable and robust user identification and authentication are important and often necessary requirements for many digital services. It becomes paramount in social virtual reality (VR) to ensure trust, specifically in digital encounters with lifelike realistic-looking avatars as faithful replications of real persons. Recent research has shown that the movements of users in extended reality (XR) systems carry user-specific information and can thus be used to verify their identities. This article compares three different potential encodings of the motion data from head and hands (scene-relative, body-relative, and body-relative velocities), and the performances of five different machine learning architectures (random forest, multi-layer perceptron, fully recurrent neural network, long-short term memory, gated recurrent unit). We use the publicly available dataset "Talking with Hands" and publish all code to allow reproducibility and to provide baselines for future work. After hyperparameter optimization, the combination of a long-short term memory architecture and body-relative data outperformed competing combinations: the model correctly identifies any of the 34 subjects with an accuracy of 100% within 150 seconds. Altogether, our approach provides an effective foundation for behaviometric-based identification and authentication to guide researchers and practitioners. Data and code are published under https://go.uniwue.de/58w1r.

Alexander Dallmann, Johannes Kohlmann, Daniel Zoller et al.

Multiplayer Online Battle Arena (MOBA) games such as Dota 2 attract hundreds of thousands of players every year. Despite the large player base, it is still important to attract new players to prevent the community of a game from becoming inactive. Entering MOBA games is, however, often demanding, requiring the player to learn numerous skills at once. An important factor of success is buying the correct items which forms a complex task depending on various in-game factors such as already purchased items, the team composition, or available resources. A recommendation system can support players by reducing the mental effort required to choose a suitable item, helping, e.g., newer players or players returning to the game after a longer break, to focus on other aspects of the game. Since Sequential Item Recommendation (SIR) has proven to be effective in various domains (e.g. e-commerce, movie recommendation or playlist continuation), we explore the applicability of well-known SIR models in the context of purchase recommendations in Dota 2. To facilitate this research, we collect, analyze and publish Dota-350k, a new large dataset based on recent Dota 2 matches. We find that SIR models can be employed effectively for item recommendation in Dota 2. Our results show that models that consider the order of purchases are the most effective. In contrast to other domains, we find RNN-based models to outperform the more recent Transformer-based architectures on Dota-350k.

Andrzej Dulny, Andreas Hotho, Anna Krause

Many physical processes such as weather phenomena or fluid mechanics are governed by partial differential equations (PDEs). Modelling such dynamical systems using Neural Networks is an active research field. However, current methods are still very limited, as they do not exploit the knowledge about the dynamical nature of the system, require extensive prior knowledge about the governing equations or are limited to linear or first-order equations. In this work we make the observation that the Method of Lines used to solve PDEs can be represented using convolutions which makes convolutional neural networks (CNNs) the natural choice to parametrize arbitrary PDE dynamics. We combine this parametrization with differentiable ODE solvers to form the NeuralPDE Model, which explicitly takes into account the fact that the data is governed by differential equations. We show in several experiments on toy and real-world data that our model consistently outperforms state-of-the-art models used to learn dynamical systems.

Padraig Davidson, Michael Steininger, Florian Lautenschlager et al.

Sensor-equipped beehives allow monitoring the living conditions of bees. Machine learning models can use the data of such hives to learn behavioral patterns and find anomalous events. One type of event that is of particular interest to apiarists for economical reasons is bee swarming. Other events of interest are behavioral anomalies from illness and technical anomalies, e.g. sensor failure. Beekeepers can be supported by suitable machine learning models which can detect these events. In this paper we compare multiple machine learning models for anomaly detection and evaluate them for their applicability in the context of beehives. Namely we employed Deep Recurrent Autoencoder, Elliptic Envelope, Isolation Forest, Local Outlier Factor and One-Class SVM. Through evaluation with real world datasets of different hives and with different sensor setups we find that the autoencoder is the best multi-purpose anomaly detector in comparison.

Alexander Dallmann, Daniel Zoller, Andreas Hotho

At the present time, sequential item recommendation models are compared by calculating metrics on a small item subset (target set) to speed up computation. The target set contains the relevant item and a set of negative items that are sampled from the full item set. Two well-known strategies to sample negative items are uniform random sampling and sampling by popularity to better approximate the item frequency distribution in the dataset. Most recently published papers on sequential item recommendation rely on sampling by popularity to compare the evaluated models. However, recent work has already shown that an evaluation with uniform random sampling may not be consistent with the full ranking, that is, the model ranking obtained by evaluating a metric using the full item set as target set, which raises the question whether the ranking obtained by sampling by popularity is equal to the full ranking. In this work, we re-evaluate current state-of-the-art sequential recommender models from the point of view, whether these sampling strategies have an impact on the final ranking of the models. We therefore train four recently proposed sequential recommendation models on five widely known datasets. For each dataset and model, we employ three evaluation strategies. First, we compute the full model ranking. Then we evaluate all models on a target set sampled by the two different sampling strategies, uniform random sampling and sampling by popularity with the commonly used target set size of 100, compute the model ranking for each strategy and compare them with each other. Additionally, we vary the size of the sampled target set. Overall, we find that both sampling strategies can produce inconsistent rankings compared with the full ranking of the models. Furthermore, both sampling by popularity and uniform random sampling do not consistently produce the same ranking ...

Michael Steininger, Daniel Abel, Katrin Ziegler et al.

Climate models are an important tool for the assessment of prospective climate change effects but they suffer from systematic and representation errors, especially for precipitation. Model output statistics (MOS) reduce these errors by fitting the model output to observational data with machine learning. In this work, we explore the feasibility and potential of deep learning with convolutional neural networks (CNNs) for MOS. We propose the CNN architecture ConvMOS specifically designed for reducing errors in climate model outputs and apply it to the climate model REMO. Our results show a considerable reduction of errors and mostly improved performance compared to three commonly used MOS approaches.

Michael Fischer, Konstantin Kobs, Andreas Hotho

Fully- or semi-automatic image enhancement software helps users to increase the visual appeal of photos and does not require in-depth knowledge of manual image editing. However, fully-automatic approaches usually enhance the image in a black-box manner that does not give the user any control over the optimization process, possibly leading to edited images that do not subjectively appeal to the user. Semi-automatic methods mostly allow for controlling which pre-defined editing step is taken, which restricts the users in their creativity and ability to make detailed adjustments, such as brightness or contrast. We argue that incorporating user preferences by guiding an automated enhancement method simplifies image editing and increases the enhancement's focus on the user. This work thus proposes the Neural Image Correction & Enhancement Routine (NICER), a neural network based approach to no-reference image enhancement in a fully-, semi-automatic or fully manual process that is interactive and user-centered. NICER iteratively adjusts image editing parameters in order to maximize an aesthetic score based on image style and content. Users can modify these parameters at any time and guide the optimization process towards a desired direction. This interactive workflow is a novelty in the field of human-computer interaction for image enhancement tasks. In a user study, we show that NICER can improve image aesthetics without user interaction and that allowing user interaction leads to diverse enhancement outcomes that are strongly preferred over the unedited image. We make our code publicly available to facilitate further research in this direction.

Daniel Schlör, Markus Ring, Andreas Hotho

Neural networks have to capture mathematical relationships in order to learn various tasks. They approximate these relations implicitly and therefore often do not generalize well. The recently proposed Neural Arithmetic Logic Unit (NALU) is a novel neural architecture which is able to explicitly represent the mathematical relationships by the units of the network to learn operations such as summation, subtraction or multiplication. Although NALUs have been shown to perform well on various downstream tasks, an in-depth analysis reveals practical shortcomings by design, such as the inability to multiply or divide negative input values or training stability issues for deeper networks. We address these issues and propose an improved model architecture. We evaluate our model empirically in various settings from learning basic arithmetic operations to more complex functions. Our experiments indicate that our model solves stability issues and outperforms the original NALU model in means of arithmetic precision and convergence.

Padraig Davidson, Michael Steininger, Florian Lautenschlager et al.

Precision beekeeping allows to monitor bees' living conditions by equipping beehives with sensors. The data recorded by these hives can be analyzed by machine learning models to learn behavioral patterns of or search for unusual events in bee colonies. One typical target is the early detection of bee swarming as apiarists want to avoid this due to economical reasons. Advanced methods should be able to detect any other unusual or abnormal behavior arising from illness of bees or from technical reasons, e.g. sensor failure. In this position paper we present an autoencoder, a deep learning model, which detects any type of anomaly in data independent of its origin. Our model is able to reveal the same swarms as a simple rule-based swarm detection algorithm but is also triggered by any other anomaly. We evaluated our model on real world data sets that were collected on different hives and with different sensor setups.

Michael Steininger, Konstantin Kobs, Albin Zehe et al.

Land-use regression (LUR) models are important for the assessment of air pollution concentrations in areas without measurement stations. While many such models exist, they often use manually constructed features based on restricted, locally available data. Thus, they are typically hard to reproduce and challenging to adapt to areas beyond those they have been developed for. In this paper, we advocate a paradigm shift for LUR models: We propose the Data-driven, Open, Global (DOG) paradigm that entails models based on purely data-driven approaches using only openly and globally available data. Progress within this paradigm will alleviate the need for experts to adapt models to the local characteristics of the available data sources and thus facilitate the generalizability of air pollution models to new areas on a global scale. In order to illustrate the feasibility of the DOG paradigm for LUR, we introduce a deep learning model called MapLUR. It is based on a convolutional neural network architecture and is trained exclusively on globally and openly available map data without requiring manual feature engineering. We compare our model to state-of-the-art baselines like linear regression, random forests and multi-layer perceptrons using a large data set of modeled $\text{NO}_2$ concentrations in Central London. Our results show that MapLUR significantly outperforms these approaches even though they are provided with manually tailored features. Furthermore, we illustrate that the automatic feature extraction inherent to models based on the DOG paradigm can learn features that are readily interpretable and closely resemble those commonly used in traditional LUR approaches.

Sarah Wunderlich, Markus Ring, Dieter Landes et al.

Over the years, artificial neural networks have been applied successfully in many areas including IT security. Yet, neural networks can only process continuous input data. This is particularly challenging for security-related non-continuous data like system calls. This work focuses on four different options to preprocess sequences of system calls so that they can be processed by neural networks. These input options are based on one-hot encoding and learning word2vec or GloVe representations of system calls. As an additional option, we analyze if the mapping of system calls to their respective kernel modules is an adequate generalization step for (a) replacing system calls or (b) enhancing system call data with additional information regarding their context. However, when performing such preprocessing steps it is important to ensure that no relevant information is lost during the process. The overall objective of system call based intrusion detection is to categorize sequences of system calls as benign or malicious behavior. Therefore, this scenario is used to evaluate the different input options as a classification task. The results show, that each of the four different methods is a valid option when preprocessing input data, but the use of kernel modules only is not recommended because too much information is being lost during the mapping process.

Markus Ring, Sarah Wunderlich, Deniz Scheuring et al.

Labeled data sets are necessary to train and evaluate anomaly-based network intrusion detection systems. This work provides a focused literature survey of data sets for network-based intrusion detection and describes the underlying packet- and flow-based network data in detail. The paper identifies 15 different properties to assess the suitability of individual data sets for specific evaluation scenarios. These properties cover a wide range of criteria and are grouped into five categories such as data volume or recording environment for offering a structured search. Based on these properties, a comprehensive overview of existing data sets is given. This overview also highlights the peculiarities of each data set. Furthermore, this work briefly touches upon other sources for network-based data such as traffic generators and traffic repositories. Finally, we discuss our observations and provide some recommendations for the use and creation of network-based data sets.

Markus Ring, Daniel Schlör, Dieter Landes et al.

Flow-based data sets are necessary for evaluating network-based intrusion detection systems (NIDS). In this work, we propose a novel methodology for generating realistic flow-based network traffic. Our approach is based on Generative Adversarial Networks (GANs) which achieve good results for image generation. A major challenge lies in the fact that GANs can only process continuous attributes. However, flow-based data inevitably contain categorical attributes such as IP addresses or port numbers. Therefore, we propose three different preprocessing approaches for flow-based data in order to transform them into continuous values. Further, we present a new method for evaluating the generated flow-based network traffic which uses domain knowledge to define quality tests. We use the three approaches for generating flow-based network traffic based on the CIDDS-001 data set. Experiments indicate that two of the three approaches are able to generate high quality data.

Lena Hettinger, Alexander Dallmann, Albin Zehe et al.

In this paper we describe our post-evaluation results for SemEval-2018 Task 7 on clas- sification of semantic relations in scientific literature for clean (subtask 1.1) and noisy data (subtask 1.2). This is an extended ver- sion of our workshop paper (Hettinger et al., 2018) including further technical details (Sec- tions 3.2 and 4.3) and changes made to the preprocessing step in the post-evaluation phase (Section 2.1). Due to these changes Classification of Relations using Embeddings (ClaiRE) achieved an improved F1 score of 75.11% for the first subtask and 81.44% for the second.

Mark Kibanov, Martin Becker, Juergen Mueller et al.

The k-Nearest Neighbor (kNN) classification approach is conceptually simple - yet widely applied since it often performs well in practical applications. However, using a global constant k does not always provide an optimal solution, e.g., for datasets with an irregular density distribution of data points. This paper proposes an adaptive kNN classifier where k is chosen dynamically for each instance (point) to be classified, such that the expected accuracy of classification is maximized. We define the expected accuracy as the accuracy of a set of structurally similar observations. An arbitrary similarity function can be used to find these observations. We introduce and evaluate different similarity functions. For the evaluation, we use five different classification tasks based on geo-spatial data. Each classification task consists of (tens of) thousands of items. We demonstrate, that the presented expected accuracy measures can be a good estimator for kNN performance, and the proposed adaptive kNN classifier outperforms common kNN and previously introduced adaptive kNN algorithms. Also, we show that the range of considered k can be significantly reduced to speed up the algorithm without negative influence on classification accuracy.

Alexander Dallmann, Alexander Grimm, Christian Pölitz et al.

Recently, Recurrent Neural Networks (RNNs) have been applied to the task of session-based recommendation. These approaches use RNNs to predict the next item in a user session based on the previ- ously visited items. While some approaches consider additional item properties, we argue that item dwell time can be used as an implicit measure of user interest to improve session-based item recommen- dations. We propose an extension to existing RNN approaches that captures user dwell time in addition to the visited items and show that recommendation performance can be improved. Additionally, we investigate the usefulness of a single validation split for model selection in the case of minor improvements and find that in our case the best model is not selected and a fold-like study with different validation sets is necessary to ensure the selection of the best model.

Thomas Niebler, Martin Becker, Christian Pölitz et al.

Assessing the degree of semantic relatedness between words is an important task with a variety of semantic applications, such as ontology learning for the Semantic Web, semantic search or query expansion. To accomplish this in an automated fashion, many relatedness measures have been proposed. However, most of these metrics only encode information contained in the underlying corpus and thus do not directly model human intuition. To solve this, we propose to utilize a metric learning approach to improve existing semantic relatedness measures by learning from additional information, such as explicit human feedback. For this, we argue to use word embeddings instead of traditional high-dimensional vector representations in order to leverage their semantic density and to reduce computational cost. We rigorously test our approach on several domains including tagging data as well as publicly available embeddings based on Wikipedia texts and navigation. Human feedback about semantic relatedness for learning and evaluation is extracted from publicly available datasets such as MEN or WS-353. We find that our method can significantly improve semantic relatedness measures by learning from additional information, such as explicit human feedback. For tagging data, we are the first to generate and study embeddings. Our results are of special interest for ontology and recommendation engineers, but also for any other researchers and practitioners of Semantic Web techniques.

Fotis Jannidis, Isabella Reger, Albin Zehe et al.

With regard to a computational representation of literary plot, this paper looks at the use of sentiment analysis for happy ending detection in German novels. Its focus lies on the investigation of previously proposed sentiment features in order to gain insight about the relevance of specific features on the one hand and the implications of their performance on the other hand. Therefore, we study various partitionings of novels, considering the highly variable concept of "ending". We also show that our approach, even though still rather simple, can potentially lead to substantial findings relevant to literary studies.

Stephan Doerfel, Daniel Zoller, Philipp Singer et al.

Social tagging systems have established themselves as an important part in today's web and have attracted the interest from our research community in a variety of investigations. The overall vision of our community is that simply through interactions with the system, i.e., through tagging and sharing of resources, users would contribute to building useful semantic structures as well as resource indexes using uncontrolled vocabulary not only due to the easy-to-use mechanics. Henceforth, a variety of assumptions about social tagging systems have emerged, yet testing them has been difficult due to the absence of suitable data. In this work we thoroughly investigate three available assumptions - e.g., is a tagging system really social? - by examining live log data gathered from the real-world public social tagging system BibSonomy. Our empirical results indicate that while some of these assumptions hold to a certain extent, other assumptions need to be reflected and viewed in a very critical light. Our observations have implications for the design of future search and other algorithms to better reflect the actual user behavior.

Nikolas Landia, Stephan Doerfel, Robert Jäschke et al.

The information contained in social tagging systems is often modelled as a graph of connections between users, items and tags. Recommendation algorithms such as FolkRank, have the potential to leverage complex relationships in the data, corresponding to multiple hops in the graph. We present an in-depth analysis and evaluation of graph models for social tagging data and propose novel adaptations and extensions of FolkRank to improve tag recommendations. We highlight implicit assumptions made by the widely used folksonomy model, and propose an alternative and more accurate graph-representation of the data. Our extensions of FolkRank address the new item problem by incorporating content data into the algorithm, and significantly improve prediction results on unpruned datasets. Our adaptations address issues in the iterative weight spreading calculation that potentially hinder FolkRank's ability to leverage the deep graph as an information source. Moreover, we evaluate the benefit of considering each deeper level of the graph, and present important insights regarding the characteristics of social tagging data in general. Our results suggest that the base assumption made by conventional weight propagation methods, that closeness in the graph always implies a positive relationship, does not hold for the social tagging domain.