Graduation Year

2021

Document Type

Dissertation

Degree

D.B.A.

Degree Granting Department

Information Systems and Decision Sciences

Major Professor

Kaushik Dutta, Ph.D.

Committee Member

Balaji Padmanabhan, Ph.D.

Committee Member

Manish Agrawal, Ph.D.

Committee Member

He Zhang, Ph.D.

Committee Member

Nasir Ghani, Ph.D.

Keywords

feature selection, game theory, machine learning, mobile advertisements, optimization

Abstract

With the proliferation of smart, handheld devices, there has been a multifold increase in the ability of firms to target and engage with customers through mobile advertising. Therefore, not surprisingly, mobile advertising campaigns have become an integral aspect of firms’ brand building activities, such as improving the awareness and overall visibility of firms' brands. In addition, retailers are increasingly using mobile advertising for targeted promotional activities that increase in-store visits and eventual sales conversions. However, in recent years, mobile or in general online advertising campaigns have been facing one major challenge and one major threat that can negatively impact the effectiveness of advertising campaigns. The challenge is the curse of high dimensionality while the threat is the curse of identification. We refer to the former as a challenge as it currently exists as a major problem, but we use the term threat for the latter as it is becoming a major problem.

The main focus of this dissertation is on resolving the curse of high dimensionality due to it "current" status. However, this dissertation also includes some solutions for addressing the curse of identification too. The curse of high dimensionality is an issue in online advertising domain that refers to the fact that personalized advertisement is obtained with the help of learning predicative models trained on a large number of users' specific features to predict a user outcome (e.g., will the user buy our product or not after seeing the ad) for each available decision (e.g., should the ad agency expose the user or not). The curse of high dimensionality is not limited to online advertising datasets, the same issue exists in many other Information Systems (IS) related datasets such as online marketing, healthcare, finance (such as bank marketing), social media as well. A main solution to resolve the curse of high dimensionality is selecting the best subset of features/attributes. This is because identifying the most relevant features are crucial for knowledge discovery and building generalizable and accurate models.

While selecting the best subset of features sounds like a classical problem, it comes with two unique characteristics in the online advertising domain compared to other IS domains. In other domains, firms do not usually need to buy data because it usually can be collected automatically. However, in online advertising domain, firms have a limited budget and the features are either costly to be purchased or costly to be computed. Therefore, if a firm invests in collecting data to run a campaign but does not gain any benefits, then that investment was unnecessary. With this mind, we propose two solution approaches for selecting the best subset of features in this dissertation. One is suitable for cases where the cost of all features are the same, and therefore the focus is on minimizing the number of futures while keeping the accuracy at the highest possible. The novelty of our first approach is that it employs the concept of Nash bargaining optimization in the field of cooperative game theory to solve the best subset selection problem in advertising domain. Our second approach is suitable for cases where the cost of features are not the same and there are limited budgets. We show that for such cases, it is significantly more efficient if users are first geographically partitioned into groups and then for each group different subsets of attributes to be selected. In order to select the best subset of features in all groups at the same time considering the budget constraint, we propose a novel optimization model and algorithm.

Regarding the curse of identification, it is important to first note that in online advertising ecosystems, collecting raw data about users are dependent on the ability to track them. The main tool for so doing is the so-called Device Identifier (device ID), assigned by Apple and Android to identify every individual smartphone in the world. So, the curse of identification means that the access to this valuable tool of obtaining data is become more and more restricted over time. Therefore, identifying appropriate way to create accurate profiling of users who are the recipients of targeted advertisement seems integral for ad agencies. Without finding appropriate way to profile users, advertisers cannot send relevant ads to users and consequently it can damage their brand reputation and waste their financial resources. Therefore, due to less accessibility to users device IDs, we propose an optimization framework for identifying unique devices in online advertising ecosystems. The hope is that by applying the framework helps ad agencies to target right customers and help mobile advertising market to continue generate revenues.

Download

Share

COinS