Master Theses

Μόνιμο URI για αυτήν τη συλλογήhttps://repository.ihu.gr/handle/11544/7

Περιηγούμαι

Τώρα δείχνει 1 - 20 από 2959

Demand Forecasting and Inventory Optimization in Businesses
Maropoulos, Alkinoos
This thesis was written as part of the MSc in Data Science program at the International Hellenic University and focuses on demand forecasting and inventory management in businesses. Accurate estimation of product demand is a critical prerequisite for the effective operation of the supply chain and for making informed business decisions, as it directly impacts service levels, inventory costs, and overall operational efficiency. This study examines contemporary data analysis methods and forecasting models, with the aim of creating an integrated system that links demand forecasts with inventory replenishment strategies. The empirical analysis utilized the real-world dataset "Corporación Favorita Grocery Sales Forecasting" (Kaggle), which comes from a large retail chain and includes daily sales by store and product, records of promotional activities, holidays, seasonal events, store characteristics, and transactions. Data preprocessing and exploratory analysis were performed, along with feature engineering, to enhance the predictive power of the models. Subsequently, different categories of models were applied and compared, such as statistical time series models (SARIMA), machine learning methods (Random Forest, XGBoost, LightGBM), and neural networks (LSTM). The accuracy of the forecasts was evaluated using appropriate metrics, while the results were utilized to calculate critical inventory management parameters, such as safety stock and the reorder point. This study highlights the importance of incorporating real-world data (promotions, holidays, store-specific features) into forecasting models, as well as the operational benefits of adopting a forecast-based inventory management system. The findings in dicate that advanced machine learning methods achieve significantly higher accuracy compared to classical statistical approaches, while the use of these forecasts in inventory models can reduce both stockouts and excess inventory. This thesis was completed under the supervision of Dr. Paraskevas Koukaras, whom I would like to thank warmly for his valuable guidance and insightful comments.
Deep learning-based identification of brain tumors in MRI scans
Louverdis, Ilias
Biomedical object detection must balance localization performance, computational efficiency, reproducibility, and robustness to dataset shift. This dissertation evaluates these trade-offs across three Ultralytics YOLO11 architectures: YOLO11n, YOLO11m, and YOLO11l. The primary experiments used a four-class T1-weighted contrast-enhanced brain MRI development cohort of 5,249 images and an independent 200-image hold-out test set, checked for exact file duplicates using MD5 hashing. A second benchmark used the BCCD blood-cell dataset, while a third evaluation used 417 FLAIR slices from TCIA-LGG for zero-shot external testing under sequence and dataset shift. A structured hyperparameter sweep examined learning rate, optimizer, training duration, image resolution, augmentation, batch size, loss weighting, and regularization. Best-performing configurations were subsequently evaluated across three random seeds. On the primary MRI validation set, mAP@50-95 reached 0.7851, 0.8134, and 0.8173 for YOLO11n, YOLO11m, and YOLO11l, respectively. On the independent MRI hold-out set, corresponding mean mAP@50-95 values were 0.4752, 0.4611, and 0.4669. AdamW based configurations achieved the highest validation scores on BCCD, while external TCIA-LGG evaluation produced mAP@50 values of 0.6111–0.6306 and mAP@50-95 values of 0.3050–0.3130. The findings indicate that optimal training configurations vary with model scale and dataset characteristics, while external dataset shift substantially reduces localization performance despite low run-to-run variability. Overall, the study provides a reproducible technical benchmark for hyperparameter optimization and external-dataset evaluation in biomedical object detection. This dissertation was completed as part of the MSc in Mobile and Web Computing: Internet of Things Applications at the International Hellenic University. I sincerely thank my supervisor, Dr. Leonidas Akritidis, for his invaluable guidance, constructive feedback, and continuous support throughout the completion of this dissertation.
Generative Artificial Intelligence in Higher Education: Ethical Challenges and Institutional Governance Responses: A Systematic Literature Review and Comparative Policy Analysis
Paraschou, Asimina-Maria
The rapid adoption of Generative Artificial Intelligence (GenAI) in higher education has created significant opportunities for teaching, learning, assessment, and institutional innovation, while simultaneously raising complex ethical and governance challenges. This dissertation examines these challenges and investigates how higher education institutions have responded to the issues through emerging governance frameworks, policies, and guidelines. The study addresses two research questions: the ethical challenges and dilemmas associated with GenAI adoption in higher education, and the ways in which higher education institutions respond to these challenges through institutional governance. A Systematic Literature Review (SLR) of 29 studies was conducted to identify and synthesize the principal ethical concerns associated with the wide use of GenAI in higher education. This was complemented by a Comparative Policy Analysis (CPA) of publicly available AI-related policy documents from ten selected universities. The literature review identified five interconnected ethical dimensions: academic integrity, fairness and algorithmic bias, transparency, accountability and AI literacy. The policy analysis subsequently examined how these dimensions are reflected within the documented institutional governance responses. The findings demonstrate that academic integrity and transparency represent prominent concerns within both the literature and the published policies. Universities have developed guidance concerning acceptable AI use, disclosure, assessment, and human responsibility, while AI literacy is also emerging as an important component of responsible institutional governance. However, the analysis reveals uneven attention across the ethical dimensions with other concerns receiving less explicit treatment and therefore indicating a partial, rather than complete, alignment between the challenges and the governance responses. The dissertation concludes that responsible GenAI adoption requires an integrated and adaptive governance approach, rather than treating GenAI solely as a threat to higher education institutions. Therefore, governance frameworks are necessary to support responsible technological innovation while preserving academic standards, educational values, and meaningful human judgement.
Development of a Machine Learning Framework for Photovoltaic Power Forecasting Using Meteorological Time-Series Data
Zachariadou, Eirini
This dissertation presents the development of a machine learning framework for forecasting photovoltaic energy production using real operational and meteorological data collected from a 500 kWp photovoltaic installation located in Chalkidiki, Greece. The study investigates the relationship between environmental conditions and photovoltaic energy generation to improve forecasting accuracy. Initially, the dissertation reviews the fundamental principles of photovoltaic systems, machine learning techniques, and the PVsyst software as a widely used photovoltaic simulation tool. A comprehensive data preprocessing procedure was then applied to prepare the dataset for predictive analysis. Several machine learning models, including Linear Regression, K-Nearest Neighbors, Decision Tree, Random Forest, and Gradient Boosting, were developed and evaluated using chronological training and testing datasets. The results showed that ensemble learning methods, particularly Random Forest, achieved the highest forecasting accuracy. Finally, a scenario-based analysis for 2026 was performed using historical meteorological conditions to estimate future photovoltaic energy production under repeated environmental patterns. Overall, the proposed framework demonstrates that machine learning techniques can provide accurate and reliable photovoltaic energy forecasts, supporting more efficient planning and operation of photovoltaic energy systems.
The impact of Artificial Intelligence personalization in e-commerce on user engagement and satisfaction
Goksin, Isik
Artificial intelligence (AI) personalization in e-commerce is an essential part of online retail product recommendations as well as personal homepage triggered contents. The thesis focuses on how AI personalization in e-commerce impacts users' engagement and satisfaction with respect to the trust and privacy issues. The data was collected from 38 online shoppers. The shoppers were from Turkey (42.1 percent) and Greece (39.5 percent), as well as the United States, Germany and Bulgaria. The main aim of the thesis was to propose an integrated conceptual model. This model incorporates the three components of the S-O-R model, TAM, and ECT. The results were analyzed descriptively with cross tabulations for demographic and behavioural sub-groups. In addition, the results were analyzed for a structured online questionnaire on awareness, scenario-based reactions, attitudinal measures, trust and privacy and behavioural engagement outcomes. The results reveal high levels of awareness of personalization. Also, it reveals a moderate positive overall assessment of the impact of personalization on engagement and satisfaction. There is a significant difference between reactions based on features: content-based and style-based. Meanwhile, there is low trust in platforms with personal information (just 15.8 percent say they trust platforms overall), and great concern about the collection of data (94.7 percent). There is a high level of personalization fatigue: 94.7 percent of the respondents have rejected a recommendation that they found wrong, and 78.9 percent say that they would give up on a platform if the personalization were repetitive or annoying.
Dynamic load prioritization in smart environments using IoT technologies
Tekidis, Pavlos
Penetration of various types of high-powered appliances, electric cars, smart-home devices, and renewable energy sources significantly increased peak load requirements, resulting in new challenges for smart-home energy management systems. Standard load management algorithms based on fixed load priorities cannot efficiently adjust to constantly changing house requirements, resulting in poor load distribution, significant device delays, and lower user satisfaction. This thesis introduces a Dynamic Load Prioritization Controller for the energy management of the smart home which allows real-time load scheduling for household devices in case of limited power availability while maintaining the balance between technical efficiency and user comfort. The proposed methodology utilizes a novel Dynamic Priority Score (DPS) which combines six decision-making criteria including device importance, user preferences, urgency, accumulated device delay, context, and normalized power consumption. These criteria are used in the weighted mathematical model, which is updated every five minutes, thus, changing priorities of household appliances. The load scheduling algorithm performs sequential device shedding when total load exceeds contractual power limitation of the house and reconnection of shed devices according to their updated priority scores. Some additional techniques such as hysteresis control, minimum device ON/OFF time constraints, and delay accumulation were introduced to improve robustness and reduce unnecessary switching operations. The simulation framework has been developed in Python programming language to analyze the proposed controller operation in the smart house with ten household appliances during twenty-four hours. Three scenarios of the controller operation have been tested, including an uncontrolled scenario, static-priority approach, and the proposed dynamic controller. It turned out that both controllers successfully managed to limit the peak load of the household from 13.7 kW to the contractual limit of 6.0 kW resulting in 56.2% peak reduction and the elimination of the overload cases. However, the proposed dynamic controller reduces average device delay by 60% compared to the static-priority controller thanks to the more effective redistribution of the load among household devices. Despite the increased number of switching operations and occasional lack of served energy, the user comfort index of the proposed algorithm exceeded 0.97 for the entire simulated period.
Music recommendation system on a social media platform based on user following data
Pierratou, Eleftheria
Recommendation Systems have become more and more essential for digital platforms. Their goal is to help users navigate the rapidly increasing catalogs of products, media and content. Nowadays, social media platforms have also included these systems to their designs. One example is Instagram’s Notes feature, which allows users to share a thirty-second snippet of the preferred music. However, the feature’s personalized For You list faces a common limitation in recommendation systems. This problem is called the cold-start problem and reduces the personalized results for users with little or no uploading history. Thus, less music is discovered on the platform and the users are unsatisfied. With this dissertation a new music recommendation system approach is proposed, addressing this limitation. The user’s social graph is used rather than activity history and four signals are combined to generate personalized results. These signals are related to a following relationship between artists and the target user, artists and the users with similar musical preferences to the target, artists and users connected to the target and finally song popularity. Then, a ranked list is created with the recommended songs. By including social information, the proposed system aims to provide meaningful suggestions from the user’s very first interaction with this feature. To evaluate the model, a small sample of twenty-two participants was used. The metrics used to examine the model’s performance were Precision, Recall, Hit Rate and NDCG. The results showed that the proposed RS performs strongly for personalized recommendations, although discovery suggestions got lower scores, probably because of the small dataset size.
Developing a chatbot using the Mitos repository
Zafeiraki, Styliani
This dissertation was written as part of the MSc in Data Science at the International Hellenic University. It explores the challenges citizens face when accessing administrative information through the Greek National Registry of Administrative Public Services (Mitos), where the gap between everyday and official language, along with the volume and complexity of the registered procedures, often makes the relevant information hard to locate and understand. The thesis examines whether a conversational system using large language models can make this information easier to access through a natural-language interface. Instead of focusing on one design, it compares three approaches for integrating administrative knowledge: a no-tool baseline that uses only the model's parametric memory, a keyword-based search architecture that queries the live Mitos registry, and a semantic retrieval-augmented generation (RAG) architecture using a local vector index, with both retrieval approaches exposed to the model through the Model Context Protocol. These are combined with five language models including a commercial cloud model, a Greek-specialized open model and a general-purpose multilingual open model, with the two open models additionally fine-tuned on synthetic Mitos interaction data. In total, fifteen configurations are evaluated over a stratified dataset of one hundred Greek citizen queries. The configurations are evaluated along multiple dimensions, such as retrieval accuracy, human-rated answer quality, tool-use behavior, cost, latency and privacy. The results show that retrieval is indispensable since no configuration can identify the correct procedures from parametric knowledge alone. The cloud model combined with keyword search achieves the strongest overall retrieval accuracy and answer quality, while semantic retrieval benefits local models most and a fine-tuned Greek-specialized model is the strongest local alternative. Fine- tuning also improves tool-use behavior and does not just store registry knowledge. No single architecture proves best across all criteria, therefore the main contribution of the thesis is not one optimal system but an open and reproducible framework that characterizes these trade-offs, and translates them into deployment-oriented guidance for conversational systems in Greek public administration. I would like to express my sincere gratitude to my supervisor, Dr. Dimitrios Karapiperis, for his valuable guidance and feedback throughout the development of this thesis, as well as to my family and friends for their continuous support during my studies.
Disease prediction and flare-ups in rheumatoid arthritis using machine learning
Tzeikou, Konstantina
Background: Rheumatoid arthritis flares are episodic exacerbations that accelerate joint damage and frequently occur between visits, making clinical management reactive. Existing predictive models rely heavily on narrow clinical trial data, with none validated for Greek and Southern European populations. This study evaluates an interpretable machine learning (ML) framework to predict next-visit flare occurrences using real-world, multi-modal longitudinal outpatient data. Additionally, this study investigates direct comparison between task-specific supervised classifiers and large language model (LLM) zero-shot and few-shot inference, examining whether pretrained medical knowledge can substitute patient-specific learning. Methods: A retrospective longitudinal dataset of 106 patients was analyzed from a dual-setting rheumatology cohort in Thessaloniki, Greece. Four tuned classifiers, Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), were trained within a patient-aware stratified group nested cross-validation scheme to prevent leakage. Models were evaluated using average precision (AP) and Area Under the Receiver Operating Characteristic Curve (ROC-AUC). Concurrently, Claude Haiku was evaluated across cross-validation partitions using zero- and few-shot prompting paradigms across two data serialization formats, establishing a benchmark against the ML models. Results: LR performed best (AP: 0.508, ROC-AUC: 0.695, Brier score: 0.201), outperforming baselines (AP: 0.331, ROC-AUC: 0.500). Comparative AP values were 0.470, 0.450, and 0.414 for SVM, XGBoost, and RF, respectively. At its F1-optimal threshold (0.295), LR achieved 83% sensitivity (108/130 flares) and ~48% specificity. Elastic Net regularization produced a sparse five-predictor equation dominated by Disease Activity Score 28 (odds ratio: 1.667 per standard deviation). LLM configurations consistently underperformed ML baselines, yielding APs of 0.40–0.42 and ROC-AUCs of 0.61–0.64. Notably, the impact of context exemplars was layout-dependent; few-shot prompting marginally improved performance for the structured List template (ROC-AUC from 0.63 to 0.64) but degraded performance for the narrative Text template (ROC-AUC from 0.62 to 0.61). Conclusions: Real-world outpatient data demonstrates the feasibility of interpretable flare prediction, where a linear model outperformed ensemble techniques. Although LLMs did not match the discriminative performance of patient-specific classifiers in this evaluation, it achieved higher overall accuracy due to a conservative prediction strategy, offering potential utility as a training-free baseline in data-scarce environments. Constraints include sample size and label subjectivity, but learning-curve trends justify multi-center expansion and prospective validation.
Microplastic Surface Properties as Drivers of Microbial Biofilm Assembly in Agricultural Soils
Koutalos, Ioannis
Abstract This dissertation was written as part of the MSc in Sustainable Agriculture and Business at the International Hellenic University. It presents a systematic literature review examining how microplastic (MP) surface properties drive microbial biofilm assembly in agricultural soils, and under which conditions the resulting plastisphere communities represent measurable environmental risks. Through a PRISMA 2020-compliant search of Scopus, Web of Science, PubMed, and Google Scholar, 30 peer-reviewed studies published between 2015 and 2025 were identified and synthesized. The review demonstrates that surface roughness and crystallinity function as early-stage assembly filters with the strongest documented associations with community composition on agricultural polyethylene (PE) plastispheres, particularly driving the enrichment of Actinobacteria and the genus Nocardia, which emerged as the most consistent polymer-specific taxonomic signature of an agricultural PE plastisphere across the included studies. Plastisphere communities were compositionally distinct from bulk soil across all studies and exhibited reduced alpha diversity and evidence of microbial homogenization at a continental scale. Environmental risks, including pathogen enrichment, accumulation of antibiotic resistance genes (ARGs), and biogeochemical disruptions, increased with MP polymer diversity, weathering state, and co-occurring chemical stressors. Critically, functional perturbation was documented in the absence of measurable differences in community composition, indicating that risk assessment of MP-contaminated agricultural soils based solely on 16S rRNA profiling is insufficient.
Effects of Treated Wastewater Irrigation and Plant Identity on Soil Nematode Community Structure and Function
Konstantinidou, Kyriaki
The increasing use of reclaimed water for agricultural irrigation provides an important strategy for reducing pressure on freshwater resources, but its consequences for below-ground ecological communities remain insufficiently understood. Soil nematodes occupy multiple trophic positions and respond to changes in resource availability and environmental conditions, making them informative indicators of soil food-web responses to irrigation management. This study examined whether reclaimed-water irrigation alters the abundance, diversity, taxonomic composition and functional organization of soil nematode communities, and whether these responses are mediated by vegetation and sampling period. A field experiment conducted in Lesvos, Greece, compared freshwater and reclaimed-water irrigation across maize (Zea mays), lavender (Lavandula angustifolia) and unplanted soils. Nematode communities were assessed during two sampling periods using measures of total abundance, genus richness and Shannon diversity, genus-level community composition, trophic-group structure and the Maturity, Enrichment, Structure and Channel Indices. Reclaimed-water irrigation produced a pronounced increase in total nematode abundance, although this numerical response was not accompanied by a corresponding change in genus richness. Community composition nevertheless differed between irrigation regimes and was strongly structured by vegetation and sampling period, with significant interactions demonstrating that the effect of irrigation varied across plant treatments and over time. Functional responses were also selective rather than uniform. Irrigation influenced the relative representation of major microbial-feeding trophic groups in a vegetation-dependent manner and significantly affected the Maturity and Structure Indices, whereas the Enrichment and Channel Indices showed no significant response. These findings demonstrate that changes in nematode abundance under reclaimed-water irrigation cannot be interpreted independently of changes in community composition and functional organization. During the experimental period, reclaimed water did not cause a generalized reduction in nematode taxonomic diversity; instead, it produced a context dependent reorganization of the nematode community, shaped by vegetation and temporal variation. The study therefore highlights the value of integrating taxonomic and functional nematode indicators when evaluating the ecological consequences of reclaimed-water irrigation and contributes to a more comprehensive assessment of its implications for agricultural soil food webs.
AI-Generated Works and Copyright Law: Reassessing Authorship and Originality in the Age of Generative Artificial Intelligence
Papafylaki, Magdalini
Generative artificial intelligence complicates the application of copyright principles that have traditionally been connected to human intellectual creation. This dissertation examines the copyright status of works produced with generative AI, focusing particularly on the degree of human contribution required for their protection. It further considers human authorship, originality, and ownership and whether the existing legal framework remains adequate as the use of generative AI continues to expand. The research follows a doctrinal legal analysis and a comparison of the approaches adopted in the European Union, the United Kingdom, and the United States. Primary and secondary legal sources are examined to assess how established copyright principles apply when generative AI participates in the creative process. The comparative analysis further highlights differences between the selected jurisdictions and considers whether greater legal clarification is required. The analysis indicates that the use of generative AI in the production of a work does not in itself exclude copyright protection. The more significant question is whether identifiable human creative choices are reflected in the expression of the final work. Where generative AI functions as a tool and sufficient human contribution can be established, the relevant elements of the work may qualify for copyright protection. By contrast, where sufficient human authorship is absent, fully AI-generated outputs may remain outside the scope of copyright protection, while the analysis finds insufficient justification at present for introducing a separate sui generis right. On this basis, the dissertation argues that generative AI does not currently require the replacement of the human-centred foundations of copyright law. The more immediate need is to clarify how those principles apply in practice, especially in relation to the threshold of human contribution and the distinction between AI-assisted and fully AI-generated works. Greater legal certainty, international coordination, and clearer transparency requirements could further support the adaptation of copyright law to continuing technological change.
AI-Based credit risk modelling for financially excluded m-pesa users in Kenya using alternative data.
Githenya, Abednego Mwangi
Access to credit is one of the biggest challenges facing informal borrowers in Kenya and Africa. This is because they lack the formal credit history that traditional lenders like banks use to give out loans. Despite this, this unbanked population in Kenya transacts large sums of money every day using Kenya’s largest mobile money wallet, M-Pesa. This transaction data contains valuable information that can be used to build a model for understanding the user’s earnings, spending habits, sentiments, and ability to repay loans. For instance, M-Pesa’s overdraft feature, Fuliza, can give us a good indication of the customer’s ability to service future loans. We will use 6 months of anonymised M-Pesa transaction records, capturing patterns in deposits, withdrawals, transfers, bill payments, and Fuliza loans, and then apply machine learning techniques using Orange 3 to classify borrowers by their risk profile. This will help us to create a model that can predict with more accuracy the ability of a customer to pay loans, which will reduce one of the barriers they face in accessing credit.
Post-Frozen Geopolitics: The Evolution of Status Quo from the Roman-Pontic Wars to Modern Regional Security Complexes
Hazal, Çakmak
This study examines the chronic crises in the Black Sea and Eastern Mediterranean basins not merely as modern political problems, but through the lens of recurrent strategic patterns and logic that have persisted across different historical epochs. The thesis' central starting point is the question of why the status quo, characterized as frozen conflict in the post-Cold War era, has evolved into a liquid phase resolved by military force today. The research utilizes the historical rivalry between the Roman Republic and the Kingdom of Pontus as a structural laboratory to understand today's unfreezing processes. This bridge, established between ancient buffer zone strategies and the military upheavals in current crisis regions such as Karabakh, Ukraine, and Gaza, aims to reveal the historical recurrent strategic logic of regional geopolitics. The study analyzes how crises can escalate into total war overnight when the international system and diplomacy are paralyzed, within the framework of the concept of "Post-Frozen Geopolitics." Adopting a qualitative comparative case analysis method, this thesis defines the Black Sea and the Eastern Mediterranean as an integrated security complex directly influencing each other. The changing doctrines of actors such as Turkey, Russia, and Israel, energy geopolitics, and proxy wars are examined as the fundamental dynamics of new solution models in this integrated basin. Consequently, this study aims to introduce the concept of "Post-Frozen Geopolitics" to the literature by emphasizing the new reality created by technological military superiority and regional alliances in a conjuncture where international institutions have become ineffective.
Exploitation of Music in Video Games: Copyright Protection, Composer Rights, and Licensing Models in the Digital Age
Zampeti Kitsou, Maria
This thesis examines the legal framework governing the exploitation of music in video games, with particular focus on copyright protection, composer rights, and licensing models in the digital age. Drawing on a comparative analysis of European Union, Greek, and United States law, and engaging critically with both case law and academic scholarship, the thesis argues that existing copyright regimes are structurally ill suited to the interactive and procedurally generative characteristics of video game music. The analysis confronts unresolved doctrinal questions surrounding fixation and originality in adaptive music systems, the attribution of authorship in technically mediated collaborative production, and the persistent failure of collective management mechanisms to capture royalties in the streaming and gameplay broadcasting context. The contractual position of composers is examined in depth, with particular attention to the work-for-hire doctrine, the practical limits of moral rights enforcement, and the equitable remuneration obligations introduced by Directive 2019/790/EU on Copyright in the Digital Single Market. The thesis concludes with policy recommendations directed at legislators, collective management organisations, and industry participants.
Natural molecules for biocides' formulation.
Panagopoulos, Drosos
This dissertation was written as part of the MSc Sustainable Agriculture and Business at the International Hellenic University. The extensive use of synthetic biocides, including quaternary ammonium compounds (QACs) and triclosan, has triggered significant environmental and public health emergencies, mostly considering aquatic toxicity and the increase of antimicrobial resistance (AMR). To address these compounding challenges, this dissertation examines the viability of natural molecules and specifically plant derived secondary metabolites, essential oils and antimicrobial peptides as sustainable alternatives to conventional synthetic agents. Unlike single-target synthetic agents, botanical biocides act simultaneously on multiple targets by disrupting pathogens membranes, inhibiting energy production and interfering with quorum sensing, mechanisms that minimize the evolutionary selective pressure that leads to antimicrobial resistance. Moreover, this study investigates the multisectoral applications of these bio-based agents, in healthcare, food preservation, material protection and industrial facility management. It highlights how advanced delivery systems like biopolymer encapsulation and nano-emulsions can overcome specific physicochemical limitations like volatility and hydrophobicity. However, to successfully commercialize these natural compounds they must comply with the European Union’s Biocidal Product Regulation (BPR) which is designed for single-molecule synthetic chemicals rather than complex botanical mixtures. Ultimately, aligning botanical biocides with the “One Health” approach and the Safe and Sustainable by Design (SSbD) framework provides a viable pathway to maintain rigorous modern hygiene standards while achieving a zero-pollution environment.
The new Era of Pharmaceutical Waste Management in EU: Exploring the gaps in EU Regulation compared to US.
Alexiou, Chrysoula
This dissertation was written as part of the MSc in Bioeconomy and Law at the International Hellenic University. This dissertation wants to explore the legal and regulatory frameworks governing pharmaceutical waste management in the European Union (EU) while exploring the gaps in regulation and compare them with those of the United States (US). It includes a comparative review of how these jurisdictions address the emerging problem of pharmaceuticals, and more specifically unused and expired medicines, their collection and take-back schemes as well as their producer responsibility. The dissertation shows that the EU follows a national level approach regarding pharmaceutical waste management but an already harmonized approach in environmental protection and extended producer responsibility principles. The US has currently regulated pharmaceutical waste ‘under the roof’ of hazardous waste management. The study highlights the unregulated parts in EU as well as the similarities and differences, strengths and weaknesses among the two systems, noting common problems, such as limited public participation, and gaps in legislation and enforcement. It concludes by recommending public and individual measures to improve pharmaceutical waste governance, including stronger legal clarity, and international exchange of best practices and transnational cooperation.
Transparency as CSR: A Study of Pharmaceutical Industry Transparency Reporting within the EFPIA Framework
Papadopoulou, Athina
This study looks at how the top ten pharmaceutical companies show transparency and compliance in their corporate social responsibility reports. It uses the EFPIA Disclosure Code as a guide and creates a special index, called the Transparency-as-CSR Index, to see how well these companies communicate transparency in their reports, including documents about transferring value, sustainability, and ethics. The index has six parts to assess how transparency is shown, explained, and included in these reports. By doing this, the study aims to understand how these companies share information about their activities and how they follow the rules.
Building Intelligent Farming Assistants via Agentic LLMs: A Market-Aware System Design
Grammatiki, Ioanna
This dissertation was written as part of the MSc in Sustainable Agriculture and Business at the International Hellenic University. Large Language Models (LLMs) have transformed the traditional chatbots into intelligence agents. However, their application to agriculture remains limited. It is undeniable fact that the already existing advisory tools in agriculture are stateless. What that means is that the existing chatbots forget all the past interactions with the farmers as well as their farming practices history. Another gap is their lack of reactiveness, meaning they cannot answer single questions without autonomously planning or gathering information. Also, they have no clear image of the market for agricultural products, and they offer advice that is disconnected from the prices, costs and regulatory constraints that determine whether a recommendation is economically worthwhile. This thesis addresses the three gaps that the existing advisory systems have. The architecture proposed is evaluated through a survey of 45 Greek farmers who attended the Agrotica 2026 international exhibition in Thessaloniki. The survey includes demographic and digital profiles, trust in agricultural information sources, technology adoption and attitudes towards AI, market access and regulatory challenges, and the perceived attractiveness of the proposed system. This thesis through the proposed agent makes three contributions. The first is a contribution in the architectural design, a contribution in providing a farming dataset specifically referring to Greek farmers and a methodological contribution in framing agricultural AI architecture research within the Design Science Research paradigm. The findings suggest that an agentic, memory and market-aware advisory system address a real need in Greek agriculture, while acknowledging that an implemented pilot deployment, ideally through cooperatives or agronomist-mediated channels, is required before claims of practical impact can be made.
Evaluating the attractiveness of sustainable practices to retail customers
Koutsoumaraki, Christina
This thesis examines the sustainability of the retail sector. More specifically, the parts that are taken into account are the European regulation, the current theoretical framework of sustainability, several business approaches to sustainability, consumers’ behavior, and developments made in the bioeconomy. In doing so, several academic articles, policy texts, and company reports about sustainability in retail are examined. The study is organized into five sections: Chapter one analyzes the European regulatory context. More specifically, it explores how major EU legislative instruments, including the Corporate Sustainability Reporting Directive, the EU Taxonomy Regulation, and the Corporate Sustainability Due Diligence Directive, regulate retailers’ obligations related to transparency, accountability, and responsible supply-chain practices. Chapter two presents the theoretical framework for sustainable practices in retail, including the Triple Bottom Line, Corporate Social Responsibility, and circular economy thinking. The main purpose of this chapter is to outline a practical guide for the environmental, social, and economic aspects of retail activity, with special emphasis on green supply chains, sustainable packaging, energy efficiency, fair-trade sourcing, and cooperation with local suppliers. Chapter three examines how consumers respond to and engage with sustainability practices in the retail sector. More specifically, it analyses the gap between what consumers claim to value in terms of sustainability and how they actually behave. Chapters four and five outline how biobased products and/or digital technologies can contribute to sustainability, gradually transforming retail supply chains and creating new possibilities and challenges for more sustainable consumption habits. The overall structure of the study is designed to address specific research questions that emerge from the review of the relevant literature, supported by additional analysis. The findings contribute to both academic understanding and practical decision-making for those involved in the design and implementation of sustainability strategies in the retail sector. Keywords: Retail sustainability; ESG regulation; Triple Bottom Line; Circular economy; Bioeconomy Student Name: Koutsoumaraki Christina Date: 01/05/2026

Περιηγούμαι

Πρόσφατες Υποβολές