Imagine a Lebron James fan who buys a pair of sneakers from the player on eBay, and if the eBay recommendation system is a seasoned salesman, what merchandise should be recommended to him?
The answer will most likely not be another pair of sneakers, but rather a look at other Lebron James collections, such as player cards, collector's prints, or other merchandise that customers don't even know exist, but might be interested in. This is because, rather than buying a bunch of sneakers in a short period of time, it is more in line with the consumption habits of ordinary people to see every interesting little thing to buy.
70% of Gen Z believe product discovery is the best part of online shopping. eBay AI Chief Nitzan Mekel-Bobrov laughs: "I agree with young people on at least this point. However, eBay doesn't have a Lebron James product category, and the product similarity recommendation system has no way of knowing the correlation between these products, so the fan won't have the opportunity to discover the idol's various merchandise, but will instead see a whole row of sneakers that he has already bought.
In order for the recommendation system to be as observant as a seasoned salesman and to pique the interest of customers, eBay had to move to a new model of customer-focused recommendations. To do this, Nitzan Mekel-Bobrov believes that it is important to understand the interests of customers and to be familiar with their own products. This also means that the paradigm transfer of eBay recommendation system does not need to abandon the technical construction of the previous era of similarity recommendation, but based on the accumulated product understanding in the past, add customer observation, and comprehensively judge the recommended products.
The key difference between the old and new models is the perspective from which the relevance of the product is judged. Historically, the practice has been to reduce customer behavior to static products or product category data assuming that product similarity equals relevance, and then matching products in neighboring categories.
The new approach is to start from customer behavior and infer how they perceive product relevance. This correlation is diverse, not only focusing on similarities, but also any other concept that connects the goods.
In order to do this, eBay outlines the customer's journey, guesses the purchase interest behind seemingly unrelated behaviors, and then dynamically recommends related products. This also means that as the system learns more about the customer, the same customer may see different recommendations for the same main product at different stages of their journey, and those recommendations will not be limited by eBay's default classification method.
To create this ultra-personalized product discovery experience, near.
In a year or two, they began to vigorously study how to further mine more useful information from customer data to strengthen the product recommendation mechanism.
Extrapolate recommendation performance from customer behavior data, rather than similarity
Under the new paradigm, eBay began to analyze customer behavior data more deeply and used it to enhance the effectiveness of recommendations. One example is adding a customer behavior data dimension to improve the recommendation performance of similar product ad fields.
This is a ** advertising page based on the transaction amount to take the advertising cost, when the customer browses product A, through this advertising recommendation, and actually buys product B, product A and product B are regarded as related, as an important reference for subsequent similarity recommendations. When other people look at product A, they will also recommend product B to him. If the purchase is not completed, it will be considered irrelevant.
This highly focused design on purchase behavior can effectively bring about purchase conversion, but it also gives rise to three problems.
The first is the issue of false negatives. Typically, a customer will only purchase at most one product from the recommendation field, and other products will be marked as unrelated, even if they are highly relevant. In addition, the number of purchase behaviors is sparse compared to other customer behaviors, so there is relatively little data that can be used to strengthen the recommendation system. Third, this mechanism does not take into account other types of behavior, which may also represent different levels of purchasing interest.
eBay tries to add more past customer behaviors to the recommendation model to address the above three problems, but there are still two challenges for customer service, first, to determine which behavior tags are influential enough to be considered. Then, give these behaviors the right weight, otherwise the recommendation will be less effective than originally designed.
In order to solve these two challenges, eBay trained more than 2,000 models according to different combinations of behavior tags and tag hyperparameters, and tested the web version and mobile app separately before deciding on the behavior tags and weights such as clicking, adding to cart, offering, clicking the buy now button, adding to the watch list, and buying, and putting the new model into the world**.
Regardless of how similar the recommended product is to the main product in the browsing, only ask whether the customer is interested in deciding the recommendation order, which is the spirit of the product from the customer's point of view in the new model.
Start with the customer's journey as the starting point to create a hyper-personalized experience that is more immediate and free from the limitations of established perceptions
It's not just a series of similar products that can be used to further infer customer interest based on their historical behavior. eBay also analyzes the customer's journey throughout eBay**, finds out their interests in real time, speculates how the customer perceives the relevance of the product, and uses this perspective to classify the product and make recommendations.
eBay's old recommendation method does not collect much customer behavior, mainly customer clicks, purchases, and search data. In order to analyze customer interests more deeply, they organized the behavior of customers into 63 events, which belong to 8 stages of exploration, product listing, browsing, search, decision-making, payment, delivery, and revisit, so as to outline the customer journey.
eBay not only records the occurrence time of these 63 events, and then converts them into sequence embeddings, but also further records the duration of each event, which is converted into time embedding vectors to represent the importance of this event.
Then, they combined the two items into a click stream embedding vector, and then compared them with the knowledge graph of their own products to dynamically form an interest graph to infer what the possible relevance of the product to customers.
Nitzan Mekel-Bobrov said that event sequences and temporal embedding vectors are computed using bi-directional long short-term memory (bilstm) networks, which are computationally intensive and therefore not typically done in the industry, but eBay believes that this is essential for in-depth and immediate understanding of customer interests.
Based on these dynamic inferences of customer interest and relevance, the system generates product categories as the basis for recommendations. In this way, you can jump out of the product categories created by eBay with similarity, and each customer will have their own classification method.
The new approach doesn't just allow eBay to get rid of its own categorical views. Nitzan Mekel-Bobrov adds that customer data analysis is so granular that it is possible to understand the interests of customers in the so-called "multi-directional".
Personalization is a one-way approach, and the more data accumulates, the more likely it is that wrong or outdated information will accumulate, and the system is not likely to forget this inapplicable information in time. He added that the "customer vector" like the ** vector is a kind of static data, which is difficult to update after acquisition, if you can understand the customer's interests more immediately, it will be easier to respond quickly according to the customer's journey, and add, remove or correct the system's understanding of the customer's interests.
For example, although the system initially determines that the customer likes the toy camera, when the customer clicks on more retro wooden toy cameras, the system will update the understanding of interest and change the understanding of the other party to the retro wooden toy camera.
However, when a customer starts clicking on more non-camera vintage wooden toys, the camera tag will be removed and the customer interest tag will be removed instead to determine that the customer's interest is a vintage wooden toy, which is the opposite of the previous optimization direction, and improves personalization by removing noise. If a customer clicks on a toy made of a different material, or shows an interest in a specific type of vintage wooden toy, the system can also "horizontally" move to the tab of other types of interests based on the customer's understanding of the customer's interest. "These moves are based on dynamically generated customer interest graphs, not predefined product assortments. Nitzan Mekel-Bobrov emphasized.
Focusing on every behavior of the customer's discovery journey and updating the customer's interest through a variety of optimization practices is a key step in eBay's development of hyper-personalization.