International Hellenic University
Εξόρυξη και ανάλυση δεδομένων για τη βελτίωση των διαδικασιών του ψηφιακού μάρκετινγκ και του μάρκετινγκ κοινωνικών δικτύων
Abstract
dc:descriptionThis doctoral dissertation investigates how recommender systems can be used more effectively in modern digital-content platforms through two complementary strategies: improving existing methods and selecting the method or architecture best suited to a particular application context. Recommender systems are approached not merely as prediction algorithms, but as personalization mechanisms that transform behavioral data into content choices and therefore play a direct role in digital marketing and user experience. A central premise of the dissertation is that increasing model complexity does not necessarily translate into greater value; the suitability of a recommender depends on the available data, the service objective, the required performance profile, and the operational constraints of the platform.The research follows an experimental, quantitative, and sequential design and develops through three interconnected scientific contributions. The first investigates whether a classical collaborative-filtering approach can be improved through a different representation of behavioral information. Movie rating data are used as the experimental dataset. Sparse user–movie ratings are transformed into compact and semantically interpretable genre-preference profiles, users are clustered on the basis of these profiles, and collaborative filtering is then applied within the resulting groups. Under the evaluated MovieLens protocol, this approach achieves lower RMSE and MAE values and shorter execution time than the corresponding non-clustered KNN approach. The findings show that improving data representation and organization can enhance an existing recommender method without necessarily moving to a more complex model family.The second contribution shifts the focus from improving a specific configuration to selecting the most appropriate method within the same model. Eight clustering algorithms are evaluated within the same collaborative-filtering framework under two alternative representations of user preferences: genre-based and movie-based. The results show that no single clustering method is consistently preferable across all settings. Relative performance changes with the representation of the data and with the trade-off sought between prediction error and computational efficiency. Clustering-method selection therefore emerges as a design problem in its own right rather than as a fixed preprocessing choice.The third contribution generalizes the selection problem to the level of the overall personalization architecture and examines it using real-world digital-learning data from Coursera. Population-level neighborhood models, behaviorally segmented neighborhood models, and Neural Collaborative Filtering variants with different metadata representations are compared under a common evaluation framework. The evaluation extends beyond RMSE and MAE to include full-catalog ranking, Precision@K, Recall@K, personalized-neighborhood coverage, and distinct dimensions of computational cost. The final study also adopts a guarded and reproducible model-selection procedure in which the selection criterion, hyperparameters, feature representation, and final configuration are determined using validation data and frozen before confirmatory testing.The results reveal clear trade-offs among the alternative architectures. No rank-capable recommender dominates across all evaluated dimensions. KNN achieves the lowest RMSE among rank-capable models and maintains high personalized-neighborhood coverage; ClusteredKNN provides the strongest observed Top-K retrieval but with lower usable-neighbor support; and NCF-all-safe achieves the lowest MAE among rank-capable models at substantially greater fitting cost. The UserMean baseline attains even lower numerical prediction error but assigns the same score to unseen courses within each user and therefore cannot meaningfully rank them. This result provides a particularly clear demonstration that rating-prediction accuracy and recommendation quality are distinct objectives.Taken together, the three contributions show that the effective use of recommender systems is both a method-improvement problem and a model-selection problem. The most appropriate solution is not necessarily the most complex or the most granular form of personalization, but the one whose performance profile best matches the service objective, the available evidence, and the operational environment. The dissertation therefore contributes not only individual algorithmic comparisons, but also a coherent decision-oriented framework for recommender-system selection based on multiple criteria, reproducibility, and explicit governance of the model-selection process. Personalized user or customer experience and satisfaction remain the ultimate strategic objectives of this approach; however, they are not directly measured by the dissertation's offline experiments and require subsequent validation under real-world usage conditions.Scientific Area: Data Mining, Recommender Systems, Artificial Intelligence, and Digital Marketing, Data Analytics.Keywords: Personalization; Recommender Systems; Collaborative Filtering; Clustering; Data Mining; Neural Collaborative Filtering; Artificial Intelligence; Digital Marketing; Model Selection Governance.
Degree
thesis:*- Grantor dc:publisher
- International Hellenic University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sevaslidou, Nerantzoula
- Σεβασλίδου, Νεραντζούλα
Subjects
dc:subject × 30- Συστήματα προτάσεων
- Εξατομίκευση συστάσεων
- Συνεργατικό φιλτράρισμα
- Ομαδοποίηση
- Εξόρυξη δεδομένων
- Νευρωνικό Συνεργατικό Φιλτράρισμα
- Τεχνητή νοημοσύνη
- Ψηφιακό μάρκετινγκ
- Διακυβέρνηση Επιλογής Μοντέλου
- Recommender systems
- Personalization
- Collaborative filtering
- Clustering
- Data mining
- Neural Collaborative Filtering
- Artificial intelligence
- Digital marketing
- Model Selection Governance
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Διεπιστημονικές εφαρμογές επιστήμης ηλεκτρονικών υπολογιστών
- Οικονομικά και Επιχειρήσεις
- Κοινωνικές Επιστήμες
- Μάρκετινγκ
- Computer and Information Sciences
- Natural Sciences
- Computer Science Interdisciplinary Applications
- Economics and Business
- Social Sciences
- Marketing
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62938