Hybrid Deep Networks Based On Periodshift Cosine Annealing For Customer Retention Prediction In Telecom Industry
In the dynamic landscape of Customer Retention Prediction (CRP), the imperative to strategically direct marketing and promotion efforts towards targeted customers has never been more crucial. Identifying potential churn indicators and continually exploring innovative retention methods becomes par...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/62044/ http://eprints.usm.my/62044/1/JOHNSON%20OLANREWAJU%20VICTOR%20-%20TESIS24.pdf |
| _version_ | 1848884874296950784 |
|---|---|
| author | Victor, Johnson Olanrewaju |
| author_facet | Victor, Johnson Olanrewaju |
| author_sort | Victor, Johnson Olanrewaju |
| building | USM Institutional Repository |
| collection | Online Access |
| description | In the dynamic landscape of Customer Retention Prediction (CRP), the
imperative to strategically direct marketing and promotion efforts towards targeted
customers has never been more crucial. Identifying potential churn indicators and
continually exploring innovative retention methods becomes paramount. However, a
major challenge is customers terminating their services are rarely known among the
loyal ones leading to an imbalance problem. Conventional Machine Learning (ML),
with its prevalent reliance on feature extraction and data sampling methods, including
cost-sensitive techniques, grapples with issues such as overfitting, computational
complexity, and an undue emphasis on rare cases. Deep Learning (DL) techniques
applied to CRP is promising for automatic feature extraction compared to the
handcrafted method used in ML. However, non-cost-sensitive nature, appropriately
chosen Learning Rate (LR) for better convergence, and quality feature learning in DL
models still pose challenges. This thesis introduces a Class Imbalance Ratio Weight
(CIRW) designed to tackle the imbalance problem in DL classifiers without incurring
additional computational costs or loss of data symmetry. Additionally, it proposes a
novel Period-Shift Cosine Annealing Learning Rate (ps-CALR) method to address LR
dynamics during DL model training, thereby enhancing generalization. Finally, a
hybrid DL model, combining an improved multilayer perceptron and a onedimensional
convolutional neural network, is developed to learn improved features for
customer retention analysis. |
| first_indexed | 2025-11-15T19:13:38Z |
| format | Thesis |
| id | usm-62044 |
| institution | Universiti Sains Malaysia |
| institution_category | Local University |
| language | English |
| last_indexed | 2025-11-15T19:13:38Z |
| publishDate | 2024 |
| recordtype | eprints |
| repository_type | Digital Repository |
| spelling | usm-620442025-03-24T02:51:22Z http://eprints.usm.my/62044/ Hybrid Deep Networks Based On Periodshift Cosine Annealing For Customer Retention Prediction In Telecom Industry Victor, Johnson Olanrewaju QA75.5-76.95 Electronic computers. Computer science In the dynamic landscape of Customer Retention Prediction (CRP), the imperative to strategically direct marketing and promotion efforts towards targeted customers has never been more crucial. Identifying potential churn indicators and continually exploring innovative retention methods becomes paramount. However, a major challenge is customers terminating their services are rarely known among the loyal ones leading to an imbalance problem. Conventional Machine Learning (ML), with its prevalent reliance on feature extraction and data sampling methods, including cost-sensitive techniques, grapples with issues such as overfitting, computational complexity, and an undue emphasis on rare cases. Deep Learning (DL) techniques applied to CRP is promising for automatic feature extraction compared to the handcrafted method used in ML. However, non-cost-sensitive nature, appropriately chosen Learning Rate (LR) for better convergence, and quality feature learning in DL models still pose challenges. This thesis introduces a Class Imbalance Ratio Weight (CIRW) designed to tackle the imbalance problem in DL classifiers without incurring additional computational costs or loss of data symmetry. Additionally, it proposes a novel Period-Shift Cosine Annealing Learning Rate (ps-CALR) method to address LR dynamics during DL model training, thereby enhancing generalization. Finally, a hybrid DL model, combining an improved multilayer perceptron and a onedimensional convolutional neural network, is developed to learn improved features for customer retention analysis. 2024-08 Thesis NonPeerReviewed application/pdf en http://eprints.usm.my/62044/1/JOHNSON%20OLANREWAJU%20VICTOR%20-%20TESIS24.pdf Victor, Johnson Olanrewaju (2024) Hybrid Deep Networks Based On Periodshift Cosine Annealing For Customer Retention Prediction In Telecom Industry. PhD thesis, Universiti Sains Malaysia. |
| spellingShingle | QA75.5-76.95 Electronic computers. Computer science Victor, Johnson Olanrewaju Hybrid Deep Networks Based On Periodshift Cosine Annealing For Customer Retention Prediction In Telecom Industry |
| title | Hybrid Deep Networks Based On Periodshift
Cosine Annealing For Customer
Retention Prediction In Telecom
Industry |
| title_full | Hybrid Deep Networks Based On Periodshift
Cosine Annealing For Customer
Retention Prediction In Telecom
Industry |
| title_fullStr | Hybrid Deep Networks Based On Periodshift
Cosine Annealing For Customer
Retention Prediction In Telecom
Industry |
| title_full_unstemmed | Hybrid Deep Networks Based On Periodshift
Cosine Annealing For Customer
Retention Prediction In Telecom
Industry |
| title_short | Hybrid Deep Networks Based On Periodshift
Cosine Annealing For Customer
Retention Prediction In Telecom
Industry |
| title_sort | hybrid deep networks based on periodshift
cosine annealing for customer
retention prediction in telecom
industry |
| topic | QA75.5-76.95 Electronic computers. Computer science |
| url | http://eprints.usm.my/62044/ http://eprints.usm.my/62044/1/JOHNSON%20OLANREWAJU%20VICTOR%20-%20TESIS24.pdf |