Abstracts Track 2023

Area 1 - Software Systems and Applications

Nr:	124
Title:	Deep Learning Methods for Employee Turnover Prediction: The Role of Turnover Contagion
Authors:	Xin Wei, Xi Zhang, Carol J. Ou, Emiel Caron and Hengshu Zhu
Abstract:	Past studies have demonstrated that turnover contagion in the workplace leads to higher costs than individual turnover, such as declines in firm performance and productive capacity. Consequently, turnover contagion has been regarded as a paramount concern for human resource management scholars and practitioners. While turnover contagion has become increasingly problematic in companies, related research on this topic is still under explored due to the lack of conceptual and empirical development, let alone an effective quantitative method to capture turnover contagion in workplace. To better describe the turnover contagion process and predict subsequent employee’s turnover, this paper proposes the concept of the employee infection index, which refers to each individual incumbent employee’s vulnerability of being influenced by the turnover status of her/his co-workers connected in the same work network, including both resigned employees and other incumbent employees. We argue such influence by coworkers accumulates over time and determinates individual employee’s stay or resignation in each stage. To capture three critical features of turnover contagion - the time variant, cumulative, and variating natures from person to person - we propose that employees’ relations contribute to the development of turnover contagion, as largely captured by the communication pattern of co-workers at the enterprise social network (ESN). In addition, we argue that by following the susceptible-infectious epidemic model (SIEM), the turnover contagion process can be described as resigned employees that have had contact with incumbent employees during a certain time period and have generated negative influences on incumbent employees, to determine the incumbent employee’s time-lagged resignation decision. Thus, SIEM can be used to calculate the infection index and capture the dynamics of the turnover contagion. However, we meanwhile argue the SIEM has yet to be able to capture the positive impacts of incumbent employees because SIEM does not consider the impacts of healthy people on contagion. To address these research gaps, our key research question is: how to conceptualize and quantify turnover contagion through the infection index to better predict employee turnover? Specifically, we develop a deep turnover contagion (DTC) model by 1) embedding SIEM in the deep learning method to enhance the interpretability of turnover prediction and 2) exploring the role of the infection index in the turnover contagion process. We use real-life but pseudonymized ESN and turnover record data from a world-leading high-tech company, including more than 20,000 employees and 3,000,000 social connections, from 2019 to April 2021, for a total of 28 months. Based on it, our DTC model demonstrates its effectiveness in predicting employee turnover based on infection index with a higher interpretability than other baseline models. The results show that the infection index is the leading predictive indicator for turnover prediction. Furthermore, in the short term, the contagion effect can accelerate the turnover rate, reaching a peak in the 6th and 7th months. This study lays a theoretical foundation to leverage social network data to examine turnover contagion and contributes to the managerial practice in designing turnover prediction systems that are effective, efficient and interpretable.

Nr:	145
Title:	Cross-Domain Applications of Data Science in Autonomous Robotics: A Study of Methods and Insights
Authors:	Mark C. Carson, Timo Lange, Philipp Meyer, Matthis Keppner, Juri Zach and Tim Tiedemann
Abstract:	This work elaborates on the application of data science methodologies for improvements and gain of new insights in three distinct robotic projects, each with a unique goal and environment: SmartRecycling-UP, an autonomous construction waste sorting system; SHIVAA, an autonomous strawberry harvesting robot; and AuTag BeoFisch, an Autonomous Underwater Vehicle (AUV) for habitat monitoring. SmartRecycling-UP involves an autonomous robot designed for the sorting of construction and demolition waste. The system leverages reinforcement learning to train an artificial neural network that controls a hydraulic crane, complemented by advanced sensory systems for precise object and material identification. Planned methods to improve the classification accuracy for this project include applying various model interpretability techniques. The RoLand project, SHIVAA, describes an autonomous strawberry harvesting robot designed for open-field operations. The system employs machine learning techniques, including a Multi-Layer Perceptron and Convolutional Neural Networks, for effective fruit classification using MSI data. To deepen our understanding of the model's decision-making process, we plan to conduct feature importance analysis and visualize CNN filters, revealing potential areas for improvement. AuTag BeoFisch introduces a novel approach for underwater monitoring using Autonomous Underwater Vehicles (AUVs). This system employs self-supervised learning for object detection in challenging and often turbulent underwater environments. To gauge the effectiveness of this self-supervised learning approach, we propose employing visualization techniques for high-dimensional data and comparing the results with other pre-training methods using conventional classification metrics. Across all projects, we will consider the unique challenges associated with each data source. Techniques like t-SNE and UMAP will be employed to reduce the dimensionality of the data, facilitating visualization and analysis. This will necessitate tailored data preprocessing methods, such as normalization, outlier detection, and handling of missing data. Deliberatively applying these methods to different project data from varied domains allows us to gain new, application-independent insights. Our goal is to evaluate whether novel approaches in one domain can inspire innovative strategies in another. This comprehensive evaluation aims to provide valuable insights into each project's performance, offering potential avenues for enhancement. J. Zach, C. Busse, S. Funk, C. Möllmann, B. -C. Renner and T. Tiedemann, “Towards Non-invasive Fish Monitoring in Hard-to-Access Habitats Using Autonomous Underwater Vehicles and Machine Learning,” OCEANS 2021: San Diego – Porto, San Diego, CA, USA, 2021, pp. 1-8, doi: 10.23919/OCEANS44145.2021.9705867. T. Tiedemann, F. Cordes, M. Keppner, and H. Peters, "Challenges of Autonomous In-Field Fruit Harvesting and Concept of a Robotic Solution," in Proceedings of the 19th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2022), vol. 1, 2022, pp. 508-515. doi: 10.5220/0011321300003271. T. Tiedemann, M. Keppner, T. Runge, T. Vögele, M. Wittmaier, and S. Wolff, "Concept of a Robotic System for Autonomous Coarse Waste Recycling," in Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics (ICINCO), 2021, pp. 493-500. ISBN: 978-989-758-522-7.