Improved data science methods for real world business applications
Φόρτωση...
Ημερομηνία
2025-10-08
Συγγραφείς
Papageorgiou, Georgios
Τίτλος Εφημερίδας
Περιοδικό ISSN
Τίτλος τόμου
Εκδότης
Παραπομπή
Παραπομπή
Άδεια Creative Commons
Εκτός εάν σημειώνεται διαφορετικά, η άδεια αυτού του αντικειμένου περιγράφεται ως Attribution-NonCommercial CC BY-NC
Περίληψη
Περίληψη
Data Science (DS) serves as a fundamental pillar across all industries, with businesses relying more than ever on data-driven insights for decision-making. Various methodologies are applied to industry data, often with multiple variations, aiming to extract valuable insights that can be translated by the corresponding stakeholders into actionable decisions. However, technological advancements and research innovations necessitate the exploration of new methodologies to drive progress in DS, Data Mining (DM), and Machine Learning (ML) research applied to industry data. This thesis sets out to explore three main pillars and introduce methods designed to enhance DS and DM applications within business contexts. To investigate and demonstrate these methods, we applied them to real-world business data from diverse domains, specifically basketball data, financial stock data, and energy-related data. Our primary research questions are focused on exploring, introducing, and presenting advanced methods for applying DS and DM techniques to real-world business data, including strategies for data retrieval. This thesis introduces innovative ML modeling approaches, new business-oriented Key Performance Indicators (KPIs) for method evaluation, novel ML evaluation methods rooted in DS fundamentals, advanced techniques for anomaly detection, and the application of Association Rule Mining (ARM) to clean datasets. Additionally, we present a new innovative normalization method for time-series data. This thesis first addressed advancements in individual modeling for highly volatile sports data, examining how optimal results can be achieved relative to standard ML methodologies. This effort aims to benchmark the forecasting performance of 14 ML models based on 18 advanced basketball statistics and KPIs. Additionally, business-oriented target variables in basketball are forecasted, focusing on minimizing forecasting errors for volatile data. A new business-oriented evaluation metric called Weighted Average Percentage Error (WAPE) is introduced. Additionally, an ensemble of anomaly detection methods was applied, including the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. The Apriori algorithm for ARM is also implemented to identify associations among different types of data. Lastly, we introduced a novel normalization method for time-series data called Adaptive Sliding Window Normalization (ASWN). This method dynamically adjusts normalization window sizes, based on anomalies detected through multiple methods. For finalizing these anomalies, DBSCAN is employed, and the Akaike Information Criterion (AIC), in conjunction with AutoRegressive Integrated Moving Average (ARIMA) models, are used to determine optimal window sizes in the absence of anomalies. By presenting all these advances in DS, spanning data processing, forecasting, and pattern and association recognition, this work aims to contribute to the fields of DS, DM, and ML as applied to real-world data and scenarios. All methods introduced are evaluated and discussed in terms of their applications, highlighting their contributions to the DS field, as well as their applicability in real-world business contexts. The primary focus remains on advancing techniques for extracting valuable insights from diverse types of raw data. Additionally, further research objectives are outlined to explore these methods and processes further, aiming to achieve optimal knowledge extraction and continued improvements in the field.