Thesis Defense - Demet Ayvaz (PHDCS)

Demet Ayvaz - Ph.D. Computer Science

Asst. Prof. Murat Şensoy- Advisor

ENHANCING DEEP LEARNING MODELS FOR
CAMPAIGN PARTICIPATION PREDICTION

Date: 31.07.2019

Time: 10:30

Location: AB1 245

Thesis Committee: 

Asst. Prof. Murat Şensoy,  Özyeğin University

Asst. Prof. Toga Akçura,  Özyeğin University

Asst. Prof. Furkan Kıraç, Özyeğin University

Assoc. Prof. Ali Fuat Alkaya, Marmara University

Asst. Prof. Boray Tek, Işık University

Abstract

With the growing volume, complexity, and dynamicity of online data, the need to have a better understanding of data and underlying relations gained even more attention. Those relations play a crucial role in understanding implicit interactions within data and in return improves the performance of learning models. In this thesis, we explore methods to improve the performance of the Wide & Deep models by making better use of data and underlying cross relations. Using a wide shallow neural network, these models combine manually crafted features from experts with the high-level features learned through a deep neural network to increase the accuracy of neural networks for tasks such as product recommendations. To eliminate the need for expert knowledge, we proposed a decision tree-based method that automatically learns higher-level features, which replaces manually crafted ones. To compare our approach with existing approaches including DeepFM, we conducted a set of comprehensive experiments on a real data for GSM customers campaign participation from Turkcell, census income dataset from UCI, and a click-through rate (CTR) prediction dataset from Criteo. The results have shown that the Wide & Deep models can be enhanced significantly with the proposed approach. In order to compare models on previously unseen oers, we run another set of experiments on previously unseen data and observe that the proposed method performs significantly better on all datasets. Besides accuracy, we also compared models in terms of training time and evaluation time and verified that the proposed model can be used in real-time. We also extended our research towards classification uncertainty and enhanced Tensor ow1 estimator API to measure prediction uncertainty using a recent approach Then, we run experiments to evaluate different models in terms of prediction uncertainty. The results have shown that the proposed model is more condent for its correct predictions while it is more uncertain for its wrong prediction.

Bio

Demet Ayvaz received her B.Sc. degree in Computer Engineering from Marmara University in 2004. Later in 2006, she received her M.Sc. degree in Computer Engineering from Boğaziçi University. She started her professional career in the area of Telecommunications in 2006 and joined Turkcell Technology in June 2011. Prior to her current position, she worked as Postpaid Solutions Specialist at Avea İletişim Hizmetleri A.Ş. Currently, she is pursuing Ph.D. degree with the Department of Computer Engineering, Özyeğin University and working as Expert Data Analytics Developer at Artificial Intelligence & Analytic Solutions department of Turkcell Technology.