Industrial Engineering MS Thesis Defense by Kaan Telciler



 

*******************************************************************

KOÇ UNIVERSITY

GRADUATE SCHOOL OF SCIENCE & ENGINEERING

INDUSTRIAL ENGINEERING

MS THESIS DEFENSE BY KAAN TELCİLER

******************************************************************

Title: Spatio-Temporal Expenditure Forecasting for Bank Customers using Transactional Data

 

Speaker: Kaan Telciler

 

Time: September 28, 2017, 14:00

 

Place: ENG 208

Koç University

Rumeli Feneri Yolu

Sariyer, Istanbul

Thesis Committee Members:

Assoc. Prof. Sibel Salman (Advisor, Koc University)

Prof. Burçin Bozkaya (Sabancı University)

Assoc. Prof. Özden Gür Ali (Koc University)

 

Abstract:

Making an accurate “next place and time” prediction for an individual bank customer’s credit card expenditure opens up profit opportunities for a bank. Working together with retailers, a bank can offer targeted campaigns to customers, potentially resulting in higher profits and better customer satisfaction. We propose a data mining approach to predict whether a bank customer will make a credit card expenditure in a given geographical area and time interval. We analyze a one-year dataset of a commercial bank. Besides features related to demographic, financial and product usage information of the customer, we include behavioral features with respect to location, time and purchase category.  In addition, we introduce proximity features that measure the distance between an input parameter and the past transactions of the customer in terms of location and time. By testing six data mining algorithms with respect to five performance measures, we predict the expenditures with an accuracy of 85.4% in datasets generated with different locations and time intervals using a sample of 10,000 customers. We present the effects of the features in the prediction performance and observe that spatial features play the most critical role in the prediction, followed by temporal features such as time between transactions. We also conduct a sensitivity analysis on prediction radius and time interval and observe significant changes in prediction performance and feature effectiveness. Another important observation is that training the model with different radius and time intervals provides the best prediction performance.