Utilizing Embeddings to Learn a Universal Customer Behavior Representation in E-Commerce

Inhaltsverzeichnis

1. Frontmatter

2. 1. Introduction

   Miguel Alves Gomes

   Abstract

   The current century has been defined by the digitalisation of society, facilitated by the proliferation of mobile devices and the connectivity and utilisation of the Internet. This development has led to a notable simplification in the use of online services and online shopping for the customer. For instance, the advent of one-click purchasing has made it significantly easier for customers to find and purchase products online without navigating through multiple pages or performing extensive searches.

3. 2. Fundamentals and Research Scope

   Miguel Alves Gomes

   Abstract

   In this chapter, the fundamental concepts and methodologies relevant for answering the research questions are established. This includes the core principles of e-commerce analytics such as the customer journey and customer targeting, and the ML techniques necessary to create a customer representation as well as predicting future customer behavior in e-commerce.

4. 3. State-of-the-Art

   Miguel Alves Gomes

   Abstract

   As introduced in Chap. 1, the primary objective of this thesis is to explore methodologies for constructing a UCR to enable personalized customer targeting. Consequently, this chapter provides an overview of state-of-the-art approaches for representing customer behavior. Sect. 3.1 presents general methodologies for developing and utilizing UCR, as outlined in Chap. 2. Sect. 3.2 examines approaches that focus on task-specific customer representation, highlighting techniques tailored to distinct applications.

5. 4. Use Cases and Data

   Miguel Alves Gomes

   Abstract

   As highlighted in Chap. 1, modern e-commerce platforms must offer personalized experiences to retain customers. This necessity demands the implementation of diverse and adaptable marketing strategies tailored to individual customers. However, achieving personalization at scale requires knowledge about the customers and their intentions, which can be challenging due to the dynamic nature of customer interactions.

6. 5. Learning Universal Customer Behavior Representation in E-Commerce with Embeddings

   Miguel Alves Gomes

   Abstract

   Following the discussion of the theoretical foundations, the limitations of existing approaches, and the goal for applying UCR in a real-world e-commerce scenario, this chapter addresses the first research question.

7. 6. Enhancing Customer Behavior Embeddings with Additional Information

   Miguel Alves Gomes

   Abstract

   Building upon the findings from Research Question 1, it has been demonstrated that self-supervised embeddings derived solely from customer interaction data serve as effective UCR.

8. 7. Lifelong Learning Embeddings for Adaptive Customer Behavior Modeling

   Miguel Alves Gomes

   Abstract

   Building upon the previous chapter, which demonstrated how enriching the UCR approach with temporal information enhances its expressiveness and predictive power, this chapter extends the discussion to another critical challenge, which is relevant for academia and industry alike: the need for continuous adaptability in dynamic e-commerce environments. While TEE addresses the question of how to encode additional session-based behavioral signals, they do not fully resolve the problem of how to maintain and evolve customer representations over time as new products, interactions, and behavioral patterns emerge. Accordingly, this chapter tackles the challenge of adaptability by addressing the third research question of this thesis.

9. 8. Beyond E-Commerce: Generalizing Self-Supervised Behavior Embedding Representation

   Miguel Alves Gomes

   Abstract

   The empirical findings presented in the preceding chapters demonstrated that customer behavior in e-commerce can be effectively captured using a self-supervised learning approach that constructs a UCR from interaction data alone. The proposed UCR embeddings have been shown to be effective across diverse prediction tasks, including purchase prediction, churn estimation, and CTR forecasting, when tested on different datasets. Furthermore, the approach satisfied critical operational requirements, such as low inference latency, adaptability to heterogeneous data sources, and robustness under dynamic conditions.

10. 9. Critical Reflection and Outlook

    Miguel Alves Gomes

    Abstract

    Building upon the preceding chapters, in which the design, implementation, and evaluation of the UCR, TEE, and LLE approaches were presented in detail, this chapter provides a critical reflection of their limitations and unresolved challenges, including constraints arising from the experimental design, methodological choices, and aspects of the implementation that could have been strengthened. While the proposed approaches demonstrate that the integration of privacy-awareness into customer representation learning is feasible, it must be acknowledged that alternative architectures and methods offer complementary capabilities in terms of efficiency, robustness, or adaptability. Furthermore, by discussing theoretical, architectural, and experimental shortcomings a more balanced perspective on the generalizability, scalability, and practical feasibility of the contributions is provided.

11. 10. Summary

    Miguel Alves Gomes

    Abstract

    This thesis contributes to the development and realization of a Universal Customer Representation (UCR) capable of accurately modeling and predicting customer behavior in e-commerce contexts. The e-commerce sector is highly dynamic and competitive, where customer satisfaction is a key driver of commercial success. To achieve this, effective personalization strategies are essential, as they enable tailored experiences that respond to individual customer needs. Given the diversity of customer behavior, personalization requires multiple marketing strategies and thus a detailed understanding of customer preferences.

12. Backmatter