Augmented Machine Learning and Optimization for Marketing

Abstract

This dissertation consists of three essays exploring how to augment machine learning and optimization methods for marketing management.

The first essay considers an augmentation of deep-learning-based recommender system for sales force management. Helping new salespeople succeed is critical for many organizations. We develop a deep-learning-based recommender system to help new salespeople recognize suitable customers, leveraging historical sales records of experienced salespeople. One challenge is how to learn from experienced salespeople’s own failures, which are prevalent but often do not show up in sales records. We develop a parsimonious model to capture these “missing by choice” sales records and incorporate the model into a neural network to form an augmented, deep-learning-based recommender system. We validate our method using sales force transaction data from a large insurance company. Our method outperforms common benchmarks in prediction accuracy and recommendation quality, while being simple, interpretable, and flexible. We demonstrate the value of our method in improving sales force productivity.

The second essay explores an augmentation of large-scale linear programming optimization method for targeting with constraints. Personalization, which aims to target different marketing actions to different customers, has attracted broad attention in both academia and industry. While most research has focused on training personalization policies without constraints, in practice, many firms face constraints when implementing these policies. For example, firms may face volume constraints on the maximum or minimum number of actions they can take, or on the minimum acceptable outcomes for different customer segments. They may also face fairness constraints that require similar actions with different groups of customers. These constraints can introduce difficult optimization challenges, particularly when the firm intends to implement personalization policies at scale. Traditional optimization methods face challenges solving large-scale problems that contain either many customers or many constraints. We show how recent advances in linear programming can be adapted to the personalization of marketing actions. We provide a new theoretical guarantee comparing how the proposed method scales compared to state-of-the-art benchmarks (primal simplex, dual simplex and barrier methods). We also extend existing guarantees on optimality and computation speed, by adapting them to accommodate the characteristics of personalization problems. We implement the proposed method, and compare it with these benchmark methods on feasibility, computation speed, and profit. We conclude that, volume and similarity (fairness) constraints should not prevent firms from optimizing and implementing personalization policies at scale.

The third essay studies collective search in an organization. In this paper, we build a two-member two-period model to show that when a group of people with different preferences conduct search and make a decision together, they can benefit from making a commitment on the number of products to search ex ante when the search cost is very small or relatively large. The underlying mechanism is that, because of the preference divergence between group members, they tend to search fewer products and thus have lower expected utility in group search than in single-agent search, and making a commitment on the number of products to search can help mitigate the preference divergence problem in group search. If consumers can observe product prices before search and the firm sets product prices endogenously, the firm can benefit from letting consumers commit to the number of products to search ex ante if consumers search as a group and their search cost is small. We also consider several extensions to show the robustness and boundary conditions of our findings.