January 13, 2023
Consumers today demand personalized experiences wherever they shop. At the same time, they’re more and more concerned about privacy. So how can marketplaces provide both personalization and privacy across their platforms? We’ll discuss how retail media innovators can use machine learning to solve this challenge.
To preserve user privacy, marketplaces need to build data management systems that follow privacy-by-design industry best practices like NIST Privacy Framework or the cloud vendors’ architectural guidance. For instance, a fundamental way to minimize data exposure is by sharing only necessary data with people or machines on the principle of least privilege.
On the flip side, merchants want to promote their products to shoppers who are most likely to buy and reach shoppers to expand their audience pool. Knowing what products to recommend to whom requires knowing the shopper's history and behaviors, as this is the essence of personalization.
To what extent do merchants need access to marketplace data? Merchants want to promote their products to their audience, so they want to know the behavioral patterns of individual shoppers to target those most likely to convert effectively. But at the same time, shoppers want to protect their privacy more than ever. 137 out of 194 countries have implemented some form of regulation in an attempt to secure data and privacy protections. Many marketplaces choose not to share shopper behavioral data with merchants because they don’t want to miss the mark on safeguarding shoppers’ information.
Such a limit makes it difficult for merchants to choose which items to recommend or promote. And even if they can access behavioral data, it’d be almost impossible to parse and derive valuable insights from the massive datasets — precisely and constantly at scale.
That’s why marketplaces need to provide intelligent, automated personalization mechanisms to protect user privacy and enable merchants to reach and convert their shoppers. Here’s where machine learning (ML) comes in.
For example, oHouse, a leading home and living marketplace in South Korea, sells furniture every seven seconds during the busy season with a monthly revenue of 180 billion Korean Won ($138 million). The company grew sales based in part on a balance of data privacy and personalization by implementing Moloco Retail Media Platform (RMP), an ML-based ad tech solution.
The power of ML comes from serving suitable ads to the right shoppers — by profoundly understanding and predicting shopper preferences based on behavioral patterns. For marketplaces, that means enabling merchants to effectively reach their audiences without direct access to private shopper data. Let’s walk through the two techniques under the hood: product understanding and navigation understanding.
The product catalog contains information about what is for sale, like the manufacturer, dimensions, or product images. By understanding the product catalog data, you can determine the items that shoppers will like based on similarities. For instance, there’s a good chance that shoppers who look at 65-inch Samsung TVs are likely also interested in similar TVs by other manufacturers like LG.
In the late ‘90s, Amazon listed 4.7 million titles for sale. Today, the e-commerce giant sells billions of products. To scale its business, it evolved from relying on data mining technologies with on-premise data warehouses to ML-based recommendation systems on its cloud infrastructure. Amazon’s personalized experiences are made possible by ML technology that predicts the best items to maximize basket size. At Moloco, we’re focused on enabling marketplaces big and small with the technology for personalized experiences.
We have developed deep learning-based product understanding techniques that understand metadata like item names, categories, prices, and even images.
Training ML technology on shopper behavioral data, or user events, enables it to predict future behavior, which is critical to item recommendation. For instance, a shopper enters search queries on a travel site, “2 bedroom hotel in Hawaii” and “vacation home in Hawaii” within a short period. By understanding user navigation, ML predicts the shopper is looking for a place to stay in Hawaii.
Navigation understanding, or sequential event understanding, is the cornerstone of many daily phenomena, like weather forecasts, stock market predictions, language translation, and item recommendations in e-commerce. ML algorithms vary widely, each with its characteristics and tradeoffs. For example, the Long Short-Term Memory (LTSM) ML algorithm is good at prediction but takes a long time to train. On the other hand, Convolutional Neural Networks (CNN) takes quicker training time, but the prediction accuracy is often lower than LSTM.
In recent years, industry and ML communities have adopted new models like the Transformer into their applications. The Transformer provides a good balance of training, prediction accuracy, and low latency responses on prediction requests, making it beneficial to operational efficiency.
Still, it takes abundant effort to properly incorporate cutting-edge technology into item recommendation scenarios since engineers often need to account for multi-faceted domains like natural language processing, computer vision, and audio recognition. At Moloco, we built our Transformer ML architecture to be customized for navigation understanding.
As we can see, building a personalized shopping experience takes tremendous resources, time, and talent. But innovative marketplaces are implementing privacy-first personalization at scale — leveraging ML-based solutions atop secure data management systems. Though many marketplaces don’t have the infrastructure or expertise to create an ML solution from scratch, they’re still able to quickly stand up and scale a retail media business by partnering with a solution provider.
Moloco built our ad tech solution on 10 years of machine learning expertise. Our open APIs enable our marketplace customers to launch their ad business in as little as eight weeks. By giving marketplaces the control of data inputs for ML training, we minimize the sharing of user personal information in an effort to strengthen user privacy while helping the marketplaces’ merchants drive sales and preserving the core shopping experience.
As I already touched on here, ML is critical to how we serve the diverse needs of marketplaces, merchants, and end users in a retail media context. In the next blog post, we will walk you through how we build deep learning models for high-performance ad prediction systems at scale.
Discover how operational machine learning, like that implemented by Amazon, can deliver limitless budgets, high ROAS, and a competitive edge for retailers and marketplaces, unlocking the potential for Amazon-like monetization.
Fueled by machine learning, retail media can drive innate e-commerce growth, on top of a new revenue stream.
As retail media networks begin to mature, how should e-commerce marketplaces approach monetization in 2023? We’ll explore the biggest trends in the space, highlight key tactics marketplaces can use to better capture ad spend, and best practices to increase profitability in the year ahead.