This thesis presents an applied project for the development of a personalized recommendation system aimed at the dynamic generation of the digital promotional flyer. The project was carried out during an internship at Technology Reply S.r.l., in collaboration with the marketing department of Aspiag Service S.r.l., the licensee of the Despar brand. The objective of the project is to optimize engagement and improve conversions through the dynamic personalization of the flyer’s content ranking for APP (mobile application) customers, based on their purchasing behavior. The proposed solution is based on a Two Towers architecture, widely adopted in modern large-scale recommender systems, with the additional option of integrating a second re-ranking stage using LightGBM or CatBoost, in view of a potential hybrid pipeline. The thesis describes in detail the data preparation and feature engineering phases—starting from customer records, purchase history, and product metadata—explaining the extraction and preprocessing decisions, as well as implementation and tuning steps. The training was performed using one year of transaction history from 20,000 customers, supported by the complete product catalog consisting of over 500,000 items. The model’s effectiveness was evaluated through ranking metrics and simulated flyer generation scenarios, with particular attention to scalability and applicability in real-world settings. The results confirm the validity of the adopted approach and open up future directions for system extension.
Questa tesi presenta un progetto applicato di sviluppo di un sistema di raccomandazione personalizzato per la generazione dinamica del volantino digitale promozionale, realizzato nel contesto di un tirocinio presso Technology Reply S.r.l. e in collaborazione con il reparto marketing di Aspiag Service S.r.l., concessionaria del marchio Despar. L’obiettivo del progetto è ottimizzare l’engagement e migliorare le conversioni attraverso la personalizzazione dinamica dell'ordinamento dei contenuti del volantino degli sconti per i clienti APP (applicazione mobile) in base al comportamento d’acquisto. La soluzione proposta si basa su un’architettura Two Towers, ampiamente adottata nei moderni sistemi di raccomandazione su larga scala, ed è stata inoltre prevista la possibilità di integrare un secondo stadio di re-ranking basato su LightGBM o CatBoost, in un’ottica di una possibile pipeline ibrida. L'elaborato descrive nel dettaglio la fase di data preparation e feature engineering, a partire da anagrafiche clienti, storico di acquisti e metadati dei prodotti – illustrando le scelte effettuate in termini di estrazione e preprocessing – così come gli step implementativi e di tuning. Per l’allenamento è stato utilizzato lo storico di transazioni di 1 anno di 20000 clienti, con il supporto dell’intero catalogo di prodotti, composto da oltre 500000 elementi. L’efficacia del modello è stata valutata attraverso metriche di ranking e scenari simulati di generazione del volantino, con un’attenzione particolare alla scalabilità e all’applicabilità in contesti reali. I risultati ottenuti confermano la validità dell’approccio adottato e aprono a scenari futuri di estensione del sistema.
Personalizzazione Dinamica del Volantino Digitale nel Retail: Un Approccio Data-Driven con Architettura Two Towers
FAVRETTO, MATTIA
2024/2025
Abstract
This thesis presents an applied project for the development of a personalized recommendation system aimed at the dynamic generation of the digital promotional flyer. The project was carried out during an internship at Technology Reply S.r.l., in collaboration with the marketing department of Aspiag Service S.r.l., the licensee of the Despar brand. The objective of the project is to optimize engagement and improve conversions through the dynamic personalization of the flyer’s content ranking for APP (mobile application) customers, based on their purchasing behavior. The proposed solution is based on a Two Towers architecture, widely adopted in modern large-scale recommender systems, with the additional option of integrating a second re-ranking stage using LightGBM or CatBoost, in view of a potential hybrid pipeline. The thesis describes in detail the data preparation and feature engineering phases—starting from customer records, purchase history, and product metadata—explaining the extraction and preprocessing decisions, as well as implementation and tuning steps. The training was performed using one year of transaction history from 20,000 customers, supported by the complete product catalog consisting of over 500,000 items. The model’s effectiveness was evaluated through ranking metrics and simulated flyer generation scenarios, with particular attention to scalability and applicability in real-world settings. The results confirm the validity of the adopted approach and open up future directions for system extension.| File | Dimensione | Formato | |
|---|---|---|---|
|
Favretto_Mattia.pdf
Accesso riservato
Dimensione
18.2 MB
Formato
Adobe PDF
|
18.2 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/92190