Abstracts Track 2024

Nr: 176
Title:

Data-Based Value Creation for an Expert Service and Research
Authors:

Heli Väätäjä, Katriina Tiira and Nina Helander
Abstract: This research aims to address how data created in a company providing expert services creates value for multiple stakeholders, including customers, company providing the expert services, as well as in research. The case highlights how value is co-created from data jointly and reciprocally by multiple stakeholder groups (Vargo et al., 2008, p.146, Xie et al. 2016). Stakeholders include two types of customers, the company itself, researchers working in collaboration with the company, as well as the wider public and academia. Research is carried out with a case study approach (Yin, 2003). First, the data creation and utilization was analyzed jointly with the company owner to understand how the data is created, who is involved in the data creation and utilization at different phases, how data-based value is created for each of the identified stakeholder groups, and at what phases of the data creation and utilization process (Lim et al. 2018). Furthermore, a feedback loop was identified that creates value from research to the provided expert services and to the customers. This part of the value creation will be analyzed more in depth later. The case company, smartDOG Oy, was founded by a researcher, who ideated to offer science-based dog cognition testing for dog owners, dog breeders, and professional dog trainer organizations. In addition, the only employee of the company, the founder and owner, lectures on dog behavior and cognition related topics based on scientific research results. Company has currently 13 license testers who participate the testing of the dogs. They have been selected by the company owner to a paid education, and pay a license fee for each of the tested dogs. There are four types of tests from 4 month old puppies to adult dogs. Tests last from 60-90 minutes on average and are paid by the customer. They get a 5-6 page report on their dog’s behavior and results for each of the tested areas. In addition, for the ¾ tests, they get access to a database, where their dog’s result is saved with their permission. Owners can then compare their dog’s results to the breed average and to average of all tested dogs. Today, there are about 7000 dogs tested. The data created in the tests is used for research purposes with the permission of the dog owners. Practically all customers give permission. The company founder acts as a docent in University of Helsinki. She supervises theses from master’s level to doctoral level that utilize the collected data. In addition, she cooperates with other research groups in several universities on analyzing the commercially collected data as well as on further research with complementing data collection with, e.g., questionnaires aimed at customers, and experimental laboratory testing. Research is published in international peer-reviewed journals. Press releases and university blog posts are distributed worldwide to reach the wider audience. Research feeds back to testing situations and reports, and to expert lectures given by the company owner, as well as to other services, such as training services, provided by the license testers in their own companies. We discuss in the presentation the data-based value creation and how value is co-created reciprocally in the case. The motivation of stakeholders and willingness of the customers to participate in co-creation is essential for the success of the services, as well as for the data-based value creation.
Nr: 196
Title:

Use of Generative AI for Knowledge Management in the Nuclear Industry
Authors:

Victor Richet, Robert Plana, Lies Benmiloud-Bechet and Frédéric Godest
Abstract: On one hand, due to high level of technical expertise and intricacies of scientific and engineering domains, as well as stringent requirements in terms of traceability, transparency and safety, nuclear industry is regarded as complex. On the other hand, due to world situation and global warming, energy domain became more and more critical the last years, and is expected to be under increasing tension. Given its low carbon footprint and reliability, nuclear power is to play a key role. Most nuclear capabilities and facilities (especially in Europe & USA) have also been built in the 70’s and 80’s and are to be decommissioned in the coming decades. Thus, the demand in terms of manpower and expertise within the nuclear industry is expected to rise drastically. Combined, those two items result in the need to integrate into a complex industry a large number of non-familiar individuals, whether they are coming from other industries, or newcomers. Indeed, those newcomers in the industry would need to integrate to a certain extent knowledge and return of experience resulting from decades of reactor operations and legacy data. Hence, knowledge management and efficient transfer is a key challenge to ensure availability of trained professionals to deliver projects. A wide range of approaches have already been implemented to address this, some being document-centric, other being people-based and training-centric. Though successful to some extent, those approaches have also raised major drawbacks, either too local to address large-scale needs, or too much relying on individual willingness to be noticeably efficient. However, the emergence of AI and Large Language Model (LLM) as a usable tool in an industrial context has enabled the exploration of new approaches of Knowledge Management. CurieLM, a Generative AI powered by Mistral technology, has been developed by Assystem and integrated into a wider approach called ‘Digital Learning from Experience’ aiming at (i) capture, formalize & digitize existing knowledge whether implicit or explicit, (ii) store and structure the acquired knowledge and (iii) ensure optimal availability of that knowledge where and when it matters. This LLM has been fine-tuned using Assystem data and knowledge with several steps of knowledge ingestion, and evaluation of results. It is currently at version 3. Selected use-case is question answering on the basis of a large document set, with automatically generated answers in natural language, on the basis of ‘elementary’ questions (i.e. not safety-related, due to intrinsic non-deterministic character of AI). This use-case has been selected due to its criticality in onboarding new people in projects, as well as its ability to be inserted into production process with as little as possible disturbances with efficient levels of knowledge acquisition. Annexed paper demonstrates usability of Generative AI in such a context, using a set of nuclear data from International Atomic Energy Agency and auto-generated/expert-validated questions, and provide explanation about potential setup for such an approach, basis for scalability and capitalization. Possible extension would involve implementation of a LLM selection processes, called Mixture Of Expert architecture (MOE), would greatly enhance flexibility and scalability of the tool. Mixture Of Expert architecture would enable to have domain-specific LLM increasing the ability to address accurately multiple domains and questions.
Nr: 198
Title:

A Generative Model Based Recommender System Against Shilling Attack
Authors:

Thi-Hanh Le, Padipat Sitkrongwong, Panagiotis Andriotis and Atsuhiro Takasu
Abstract: Recommender systems (RSs) are widely used in e-commerce systems where the user-item interactions are used to create the vector representations that represent user profiles and item features. These systems, however, are vulnerable to manipulations by malicious entities, such as unsolicited users or vendors, through various shilling attacks. Such attacks intentionally distort recommendations by introducing biased data to promote or undermine specific products or services. Our research introduces a generative model designed to mitigate the impact of shilling attacks within an adversarial learning framework. We employ a diffusion model as the generator to process the inherently noisy and sparse historical user-item interactions. The discriminator is a multi-layer perceptron that utilizes Wasserstein distance as its loss function. We conducted preliminary experiments using three well-known evaluation datasets: MovieLens 100K, MovieLens 1M, and Yelp. By simulating various attack scenarios and injecting fake interactions into the datasets, we demonstrate that our model outperforms DiffRec, a diffusion model-based recommender system, across nearly all datasets and attack types, showing enhanced robustness against deterioration of the recommendation performance.
Nr: 199
Title:

End-User-Centric Business Process Analysis: Identifying Differences Between Intended and Real Processes
Authors:

Szabina Fodor and Bernadett Sarró-Oláh
Abstract: The extensive digitalisation of businesses in various sectors has resulted in a significant increase in the volume of data available for these businesses to extract essential insights into their performance and find areas for improvement. Examining internal processes is crucial for companies to prevent redundant work and discover areas that can be automated. Over the last ten years, there has been a significant increase in the adoption of process mining (PM), a reassuring sign of its relevance, as demonstrated by numerous reported use cases in industry and academia. Nevertheless, PM research has predominantly focused on technical aspects, prioritizing the creation of algorithms rather than providing direction for individual users. To tackle this issue, we suggest a complex framework and a technology ecosystem concept that allows users to understand real-world processes based on event log files. Furthermore, users can understand the differences between these real processes and their intended processes, which can be used to improve or automate them. Our solution performs a comparative analysis of mined and designed business processes by transforming the processes into process ontologies using Java-based, publicly available, open-source software and comparing these process ontologies. The results of the comparison are reported to the end-user who is not familiar with process modelling. To demonstrate the practical application of our proposed framework, we implemented a proof-of-concept in an educational environment. In this case, we used the event log files of Moodle, a widely used Learning Management System (LMS), to show how our solution could benefit the end-user, in this instance, the teacher, in gaining a more accurate understanding of the student’s learning process. We specifically applied our solution to the STEM course 'Introduction to Artificial Intelligence' at Corvinus University of Budapest. The rapid pace of technological change underscores the significance of our educational focus. Today's professionals must update their skills every 12 to 18 months and continually acquire new skills throughout their careers, as jobs increasingly require diverse skills, and their nature evolves over time. One effective way to assist employees in adapting to these new paradigms is through e-learning initiatives, which promote self-paced and accessible learning.