An analysis of wide and deep classification techniques for human behavioral data

Date: 14 December 2017

Venue: University of Antwerp, Stadsampus - Hof van Liere - Willem Elsschotzaal, Prinsstraat 13 - 2000 Antwerp (route: UAntwerpen, Stadscampus)

Time: 5:00 PM

PhD candidate: Sofie De Cnudde

Principal investigator: Prof David Martens

Short description: PhD defence Sofie De Cnudde - Faculty of Applied Economics


The explosion of data collection and storage possibilities has proliferated the concept of big data. One such class of data is behavioral big data. As more and more aspects of people’s lives are recorded and quantified, the opportunities for data collection regarding human instances have largely increased. These massive and rich trails of both active and passive footprints harbor major potential for predictive analysis, on the one hand resulting in improvements to peoples’ lives and on the other hand resulting in profits for organizations and governments. One such application entails the use of webpage visiting behavior to predict whether a user is interested in an advertisement. Human behavioral data is characterized by being very high-dimensional and highly sparse. As many state-of-the-art data mining algorithms were designed in a pre-big-data era, the question arises as to how these techniques cope with the challenges of this ubiquitous type of big data. This work attempts to contribute to this gap in two ways.

First, a benchmarking study is performed of commonly-used, wide classification techniques. Secondly, an investigation is made into the performance of the emerging class of deep learning techniques. These structured, comparative analyses are enriched with three case studies each exploiting human behavioral data in an innovative fashion. In collaboration with the city of Antwerp, location visiting behavior is employed to increase the use of a government loyalty card, enhancing citizens' cultural participation. A second case study is performed in cooperation with microfinance company Lenddo to automate the credit scoring process for ‘underbanked’ loan applicants using social network data from Facebook. The use of behavioral data in this setting could help contribute to the economic growth of developing countries. Lastly, a case study in an e-commerce setting explores the value in sequences of human behavior by learning time-dependent user embeddings for efficient customer targeting.