Improving E-Commerce with Personalized Feeds

Personalized feeds are becoming a must-have for e-commerce platforms. They help customers find items they want quickly. One popular option is using Embeddings, which are ways to represent actions and items in a format that machines can understand. By using embeddings, systems can figure out how similar items are and suggest them to users. This paper discusses a new method that brings speed and efficiency to personalized feeds while improving user engagement.

#The Challenge with Current Methods

Personalized recommender systems have some issues. Two main problems are often encountered:

1. Limited Diversity: When users are represented by only one embedding, their wide range of interests may not be fully captured. This means that the items suggested may not reflect the user's varied tastes.

2. Cost of Real-Time Updates: Keeping user profiles current is important for giving users the best recommendations. However, maintaining a complex infrastructure for real-time updates can be expensive.

The aim of this work is to overcome both of these hurdles in a practical way.

#Proposed Solution

The proposed method updates Customer Profiles dynamically and creates personalized feeds at regular intervals, every two minutes. To do this, it uses embeddings that are calculated beforehand, along with their similarities, to generate suggestions quickly and efficiently. The solution has already been implemented at Bol, a leading e-commerce platform in the Netherlands and Belgium, where it improved customer engagement and increased conversion rates.

#Understanding Personalized Feeds

Personalized feeds can be thought of like search engines where customers are the queries, and items in the catalog are the results. To ensure these feeds work well, there are four major challenges that need to be addressed:

#1. Customer Representation Challenge

When customers shop online, their behaviors can be complex. They might search for items, view them, read reviews, and compare prices. This makes it tough to summarize this behavior into a clear representation. In addition to recent interactions, it can be beneficial to include basic customer details, such as age or preferences.

#2. Item Representation Challenge

Items in an online store come with lots of information. This includes IDs, titles, descriptions, and even feedback from previous customers. The challenge lies in figuring out which data is the most relevant for each item. Two significant factors complicate this process:

• The variety of item features: Different categories of items may require different types of information to be effectively represented.
• New products: Items that have just been released may not have enough historical data available.

#3. Candidate Retrieval Challenge

This challenge focuses on picking items that align closely with a customer’s tastes. It dives into two main aspects: ensuring the model can effectively map customer and item representations to the same space and finding the best matches from a vast catalog.

#4. Ranking Challenge

After potential items are retrieved, they need to be ranked based on how well they fit the customer's interests. A more complex model can help with this by taking into account various features of both the items and the customer queries.

#Focus Areas

The focus of this method is on the first three challenges: how to represent customers and items effectively, and how to retrieve the best candidate items.

#Contributions of the New Method

Traditional methods for creating personalized feeds rely on a user-item framework. Two neural networks, known as dual encoders, create embeddings for users and items. However, this has some notable drawbacks:

• Single Vector Representation Bottleneck: Representing users with just one vector limits the ability to capture the complexity of their interests.
• High Infrastructure and Maintenance Costs: Keeping user embeddings updated in real-time can be costly.

The new method addresses these issues by focusing on item-to-item relationships. Instead of just looking at how a user interacts with items, this approach also considers how items relate to each other.

#Methodology

The method utilizes a transformer-based architecture to generate several embeddings in one go. Each item can be seen in three different roles:

1. View Query: Items users have clicked on during their browsing sessions.
2. Buy Query: Items frequently bought in conjunction with others.
3. Target Item: The item that is ultimately purchased.

By creating these representations in one run, the method enhances efficiency and reduces the number of processes required.

#Training the Model

The model is trained using two sets of relationships: view-buy and buy-buy. In this context:

• A view-buy pair consists of an item viewed by a user that led to a purchase.
• A buy-buy pair consists of items bought together or in close succession.

By using dual encoders during training, the process focuses on making the embeddings for matching pairs similar while keeping them distinct for non-matching pairs.

#Generating Personalized Feeds

The process to create these feeds involves several steps:

1. Precomputing Embeddings: After training, embeddings are generated for all items in the catalog.
2. Precomputing Similarities: Each item's target embeddings are indexed, allowing the system to search for the most relevant match efficiently.
3. Generating Feeds: Recommendations come from selecting queries for each customer, matching potential purchases, and ranking these items based on various criteria.

#Personalized Item Feeds at Bol

The new method has been successfully implemented at Bol, generating several feeds with titles like "Top Deals for You," "Top Picks for You," and "New for You." Each of these feeds focuses on different factors for personalization:

• Top Deals for You: This feed highlights discounted items. It looks at the most recent customer interactions to pull in the latest promotions.

• New for You: This feed focuses on recently released items. The method cleverly works around the cold-start problem, where new products lack sufficient interaction data, by relying on textual metadata for embedding.

By leveraging these strategies, Bol has seen significant improvements in customer engagement and conversion rates.

#Evaluation of the Method

Both offline and online experiments were conducted to assess the effectiveness of the new method.

#Offline Testing

The offline tests used recall metrics to compare different design choices. The dataset was divided into training, validation, and test sets. The goal was to capture how well the method performed by looking at how often the recommended item appeared among the top selections.

#Online A/B Testing

An online experiment compared a group receiving personalized recommendations against a control group with standard ones. Results indicated a 4.9% rise in conversion rates and a significant increase in items added to users' wish lists, showcasing the method's success.

#Conclusion

The new approach to creating personalized feeds offers a faster and more efficient way of delivering tailored recommendations to users. By focusing on item-to-item relationships, the method not only simplifies the real-time deployment but also enhances interpretability and efficiency.

#Future Work

There is room for improvement in the method's design. Two main areas are being considered for future enhancements:

1. Handling Many-to-Many Relationships: The current method's ability to address complex relationships can be enhanced. Exploring multi-vector representations may allow for broader item selections, accommodating varied user needs better.

2. Incorporating Memorization and Generalization: Balancing item popularity with the system's ability to suggest fresh recommendations could be improved. Future designs can incorporate features that adaptively use both memorization and generalization to enhance predictive performance.

Through these advancements, the goal is to further refine the approach to personalized feeds, ensuring it meets the growing demands of e-commerce customers.