Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building a Scalable, Effective, and Steerable Search and Ranking Platform

Marjan Celikik, Jacek Wasilewski, Ana Peleteiro Ramallo, Alexey Kurennoy, Evgeny Labzin, Danilo Ascione, Tural Gurbanov, Géraud Le Falher, Andrii Dzhoha, Ian Harris

TL;DR

This paper presents a personalized, adaptable near real-time ranking platform that is reusable across various use cases, such as browsing and search, and that is able to cater to millions of items and customers under heavy load (thousands of requests per second).

Abstract

Modern e-commerce platforms offer vast product selections, making it difficult for customers to find items that they like and that are relevant to their current session intent. This is why it is key for e-commerce platforms to have near real-time scalable and adaptable personalized ranking and search systems. While numerous methods exist in the scientific literature for building such systems, many are unsuitable for large-scale industrial use due to complexity and performance limitations. Consequently, industrial ranking systems often resort to computationally efficient yet simplistic retrieval or candidate generation approaches, which overlook near real-time and heterogeneous customer signals, which results in a less personalized and relevant experience. Moreover, related customer experiences are served by completely different systems, which increases complexity, maintenance, and inconsistent experiences. In this paper, we present a personalized, adaptable near real-time ranking platform that is reusable across various use cases, such as browsing and search, and that is able to cater to millions of items and customers under heavy load (thousands of requests per second). We employ transformer-based models through different ranking layers which can learn complex behavior patterns directly from customer action sequences while being able to incorporate temporal (e.g. in-session) and contextual information. We validate our system through a series of comprehensive offline and online real-world experiments at a large online e-commerce platform, and we demonstrate its superiority when compared to existing systems, both in terms of customer experience as well as in net revenue. Finally, we share the lessons learned from building a comprehensive, modern ranking platform for use in a large-scale e-commerce environment.

Building a Scalable, Effective, and Steerable Search and Ranking Platform

TL;DR

This paper presents a personalized, adaptable near real-time ranking platform that is reusable across various use cases, such as browsing and search, and that is able to cater to millions of items and customers under heavy load (thousands of requests per second).

Abstract

Modern e-commerce platforms offer vast product selections, making it difficult for customers to find items that they like and that are relevant to their current session intent. This is why it is key for e-commerce platforms to have near real-time scalable and adaptable personalized ranking and search systems. While numerous methods exist in the scientific literature for building such systems, many are unsuitable for large-scale industrial use due to complexity and performance limitations. Consequently, industrial ranking systems often resort to computationally efficient yet simplistic retrieval or candidate generation approaches, which overlook near real-time and heterogeneous customer signals, which results in a less personalized and relevant experience. Moreover, related customer experiences are served by completely different systems, which increases complexity, maintenance, and inconsistent experiences. In this paper, we present a personalized, adaptable near real-time ranking platform that is reusable across various use cases, such as browsing and search, and that is able to cater to millions of items and customers under heavy load (thousands of requests per second). We employ transformer-based models through different ranking layers which can learn complex behavior patterns directly from customer action sequences while being able to incorporate temporal (e.g. in-session) and contextual information. We validate our system through a series of comprehensive offline and online real-world experiments at a large online e-commerce platform, and we demonstrate its superiority when compared to existing systems, both in terms of customer experience as well as in net revenue. Finally, we share the lessons learned from building a comprehensive, modern ranking platform for use in a large-scale e-commerce environment.
Paper Structure (23 sections, 1 theorem, 4 equations, 5 figures, 7 tables, 1 algorithm)

This paper contains 23 sections, 1 theorem, 4 equations, 5 figures, 7 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. SYSTEM DESIGN
  4. CANDIDATE GENERATION LAYER
  5. RANKING LAYER
  6. Model Architecture
  7. Encoding of the Customer Journey Data
  8. Position Debiasing
  9. POLICY LAYER
  10. Model Productionization
  11. EXPERIMENTS
  12. Offline Experiments
  13. Dataset
  14. Metrics and evaluation protocol
  15. Candidate Generation
  16. ...and 8 more sections

Key Result

theorem 1

Let the range of customer and article features be bounded: $C^\ell_i \le c_i \le C^u_i$, $i=1,\,\ldots,\,k_c$, and $A^\ell_j \le a_j \le A^u_j$, $j=1,\,\ldots,\,k_a$, and let the target function $f$ be continuous on the feature domain Then for any $\varepsilon > 0$, there exist $n>0$ and transformations $\varphi\colon \mathrm{R}^{k_c}\mapsto\mathrm{R}^n$ and $\psi\colon \mathrm{R}^{k_a}\mapsto\ma

Figures (5)

  • Figure 1: Item catalog browse and search page. On the left is the item category tree, and in the top-right the search query box.
  • Figure 2: Ranking Platform: Overview
  • Figure 3: A two-tower model in candidate generation layer used to learn (customer, context) and item embeddings.
  • Figure 4: The ranking layer model architecture. The model consists of an embedding layer, transformer encoders, ranking heads (per positive action), a candidate branch, and a position branch used for position de-biasing.
  • Figure 5: Histogram of the number of occurrences of interaction types used in the user sequences (cc for purchases, a2w for add-to-wishlist, a2c for add-to-cart, and pc for product click). Occurrences are on the logarithm scale.

Theorems & Definitions (1)

  • theorem 1