Access Type

Open Access Dissertation

Date of Award

January 2021

Degree Type


Degree Name



Industrial and Manufacturing Engineering

First Advisor

Ratna B. Chinnam






August 2021

Advisor: Dr. Ratna Babu Chinnam Major: Industrial & Systems Engineering Degree: Doctor of Philosophy

User engagement has emerged as the engine driving online business growth. Many firms have pay incentives tied to engagement and growth metrics. These corporations are turning to recommender systems as the tool of choice in the business of maximizing engagement. LinkedIn reported a 40% higher email response with the introduction of a new recommender system. At Amazon 35% of sales originate from recommendations, while Netflix reports that ‘75% of what people watch is from some sort of recommendation,’ with an estimated business value of $1 billion per year. While the leading companies have been quite successful at harnessing the power of recommenders to boost user engagement across the digital ecosystem, small and medium businesses (SMB) are struggling with declining engagement across many channels as competition for user attention intensifies. The SMBs often lack the technical expertise and big data infrastructure necessary to operationalize recommender systems.

The purpose of this study is to explore the methods of building a learning agent that can be used to personalize a persuasive request to maximize user engagement in a data-efficient setting. We frame the task as a sequential decision-making problem, modelled as MDP, and solved using a generalized reinforcement learning (RL) algorithm. We leverage an approach that eliminates or at least greatly reduces the need for massive amounts of training data, thus moving away from a purely data-driven approach. By incorporating domain knowledge from the literature on persuasion into the message composition, we are able to train the RL agent in a sample efficient and operant manner.

In our methodology, the RL agent nominates a candidate from a catalog of persuasion principles to drive higher user response and engagement. To enable the effective use of RL in our specific setting, we first build a reduced state space representation by compressing the data using an exponential moving average scheme. A regularized DQN agent is deployed to learn an optimal policy, which is then applied in recommending one (or a combination) of six universal principles most likely to trigger responses from users during the next message cycle. In this study, email messaging is used as the vehicle to deliver persuasion principles to the user. At a time of declining click-through rates with marketing emails, business executives continue to show heightened interest in the email channel owing to higher-than-usual return on investment of $42 for every dollar spent when compared to other marketing channels such as social media.

Coupled with the state space transformation, our novel regularized Deep Q-learning (DQN) agent was able to train and perform well based on a few observed users’ responses. First, we explored the average positive effect of using persuasion-based messages in a live email marketing campaign, without deploying a learning algorithm to recommend the influence principles. The selection of persuasion tactics was done heuristically, using only domain knowledge. Our results suggest that embedding certain principles of persuasion in campaign emails can significantly increase user engagement for an online business (and have a positive impact on revenues) without putting pressure on marketing or advertising budgets. During the study, the store had a customer retention rate of 76% and sales grew by a half-million dollars from the three field trials combined. The key assumption was that users are predisposed to respond to certain persuasion principles and learning the right principles to incorporate in the message header or body copy would lead to higher response and engagement.

With the hypothesis validated, we set forth to build a DQN agent to recommend candidate actions from a catalog of persuasion principles most likely to drive higher engagement in the next messaging cycle. A simulation and a real live campaign are implemented to verify the proposed methodology. The results demonstrate the agent’s superior performance compared to a human expert and a control baseline by a significant margin (~ up to 300%). As the quest for effective methods and tools to maximize user engagement intensifies, our methodology could help to boost user engagement for struggling SMBs without prohibitive increase in costs, by enabling the targeting of messages (with the right persuasion principle) to the right user.