ZDIN-Zahlen im Detail

Active Learning (AL) befasst sich mit der Ermittlung der informativsten Stichproben für die Beschriftung, um die Kosten für die Datenannotation bei überwachten Lernaufgaben zu senken. Die AL-Forschung leidet unter der Tatsache, dass Aufzüge aus der Literatur schlecht verallgemeinert werden können und dass nur eine kleine Anzahl von Wiederholungen von Experimenten durchgeführt wird. Um diese Hindernisse zu überwinden, schlagen wir CDALBench vor, den ersten Benchmark für aktives Lernen, der Aufgaben aus den Bereichen Computer Vision, Verarbeitung natürlicher Sprache und tabellarisches Lernen umfasst. Durch die Bereitstellung eines effizienten, gierigen Orakels kann CDALBench mit 50 Durchläufen für jedes Experiment ausgewertet werden. Wir zeigen, dass sowohl der domänenübergreifende Charakter als auch eine große Anzahl von Wiederholungen für eine anspruchsvolle Bewertung der AL-Forschung entscheidend sind. Konkret zeigen wir, dass die Überlegenheit bestimmter Methoden in den verschiedenen Domänen variiert, was es wichtig macht, Active Learning mit einem domänenübergreifenden Benchmark zu evaluieren. Außerdem zeigen wir, dass eine große Anzahl von Durchläufen entscheidend ist. Wenn nur drei Durchläufe durchgeführt werden, wie es in der Literatur häufig der Fall ist, kann die Überlegenheit bestimmter Methoden stark von den einzelnen Durchläufen abhängen. Dieser Effekt ist so stark, dass, abhängig vom Seed, sogar die Leistung einer gut etablierten Methode für denselben Datensatz signifikant besser und signifikant schlechter als zufällig sein kann.

In this paper, we study the spatial implications of digital entrepreneurship. Leveraging detailed micro-data on the universe of new venture formations in Germany between 2011 and 2018, we illustrate regional determinants of digital entrepreneurship. Unlike conventional entrepreneurship, digital entrepreneurship demonstrates sustained growth rates throughout this time period, highlighting the policy importance of understanding the drivers of digital ventures’ location choices. The key insight of our study is that digital entrepreneurship requires both digital infrastructure and highly-skilled human capital. If both are present, digital entrepreneurship can flourish in rural areas, even if digital venture formations generally concentrate in urban centers.

The COVID-19 pandemic has caused economic crises and increased inter-regional economic disparities. However, the catalyst effect that the pandemic has had on digitalization may change the traditional pattern of internal migration, in favour of rural areas. Using time-sensitive register data we explain changes in net migration rates of German districts from an economic geography perspective. We show that, since the beginning of the pandemic, rural regions benefited more from migration. In particular, younger and highly skilled individuals increasingly move to rural areas. This confirms that rural regions can benefit from increasing remote work, but these effects differ across region types and individual skill levels.

In industrial environments, efficient management of task takeovers is critical, especially when tasks are complex and workers face frequent interruptions such as shift changes. These takeovers often increase the cognitive load and the potential for errors. We propose the integration of Augmented Reality (AR) to assist users in handling takeovers of complex, and spatial distributed tasks. We developed three visualization techniques with increasing levels of spatial registration – Diagram View as Baseline, Map View, and Location-bound View – ranging from a simple overview to spatially anchored to enhance spatial and temporal awareness during task takeovers. A comparative user study (n=24) was conducted to evaluate the effectiveness of these techniques in improving task takeovers efficiency. Our results showed that the Location-bound View, in particular, significantly reduced cognitive load and improved task performance by integrating contextual information directly into the user’s field of view (FOV). This work provides insights for AR system designers and suggests that AR overviews with increasing levels of spatial registration can effectively support the takeover of complex distributed tasks in industrial settings.

This research proposes an effective way to address the issues faced by pick-up and delivery services. The real-world variables that affect delivery routes are frequently overlooked by traditional routing technologies, resulting in differences between intended and actual trajectories. Similarly, the issue of forecasting the Estimated Time of Arrival involves unique challenges due to its high dimensionality. We suggest an integrated predictive modeling methodology that tackles routing prediction in a dynamic environment and ETA prediction at the same time to overcome these difficulties. Our method, Graph2RETA, uses a dynamic spatial-temporal graph-based model to forecast delivery workers’ future routing behaviors while integrating route inference into ETA prediction. Graph2RETA leverages rich decision context and spatial-temporal information to improve the prediction accuracy of the concurrent state-of-the-art while capturing dynamic interactions between workers and timesteps by incorporating the underlying graph structure and features.