Πρόβλεψη αποχώρησης πελατών από πολυκαναλικά δεδομένα CRM με τη χρήση μοντέλων βαθιάς μάθησης και αρχιτεκτονικών προσοχής

Ψύμαρνος, Βασίλειος

Customer churn prediction from omnichannel CRM data using deep learning models and attention architectures

Bachelor Dissertation

Author

Ψύμαρνος, Βασίλειος

Date

2026-07

Abstract

The subject of this thesis is the study, development, and comparison of advanced Deep Learning models aimed at predicting customer churn. In today’s business world, retaining existing customers is vital, as acquiring new ones costs significantly more. For this reason, companies seek intelligent ways to identify early which consumers show signs of diminished interest, with the goal of offering them incentives to stay. For the purposes of this work, real and complex business data from an Omnichannel CRM system were used (customer data, online orders, interactions with customer service, device history, coupon usage, and email campaigns). Instead of a static analysis of this data, this approach focuses on the temporal evolution of each customer’s behavior. Within this framework, a comprehensive data preprocessing procedure was designed and implemented, which transformed the scattered history of user actions into structured fixed-length temporal sequences. Three different neural network architectures were trained and compared based on these time sequences: 1. LSTM (Long Short-Term Memory) networks, which specialize in memorizing long-term dependencies in data sequences. 2. GRU (Gated Recurrent Unit) networks, which are a simpler and computationally lighter variant of LSTMs. 3. Transformer models, which use the Multi-Head Attention mechanism and represent cutting-edge technology in modern Artificial Intelligence. The models were evaluated and compared using established Data Science metrics such as Accuracy, Precision, Recall, F1-Score, and the AUC-ROC curve. The research results showed that using time sequences allows neural networks to detect patterns of gradual customer disengagement. The final comparison highlights the advantages and disadvantages of each architecture, offering valuable insights both scientifically and practically for the application of such systems in the market.

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Πληροφορικής

Number of pages

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19633

Collections

Τμήμα Πληροφορικής

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα