eBay uses machine learning to improve sale listings

​Translator | Bugatti

Reviewer | Sun Shujuan

The online marketplace eBay has added additional buying signals to its machine learning model, such as “add to watchlist”, “ Bid" and "Add to Cart" to increase the relevance of recommended ad listings based on the initial product being searched. Chen Xue gave a very detailed introduction in this recent article​​.

# eBay’s Promotional Listing Standard (PLS) is a paid option for sellers. Using the PLSIM option, eBay's recommendation engine will recommend sponsored products similar to the one the potential buyer just clicked on. PLSIM pays on a CPA model (sellers only pay eBay when a sale is made), so this is a great incentive to create the most efficient model to promote the best listings. This often works out for sellers, buyers, and eBay.

The PLSIM journey is as follows:

1. User searches for products.

2. The user clicks on the results from the search -> Log in to View Items (VI) page to view the listed items (eBay calls them seed items).

3. Users scroll down the VI page and can see recommended products in PLSIM.

4. The user clicks on the product from PLSIM, performs an action (view, add to shopping cart, buy now, etc.), or view another new set of recommended products.

From a machine learning perspective, the PLSIM journey is as follows:

Retrieve the subset candidate promotion list criteria that are most closely related to the seed item ("Check the complete collection").
1. Use a trained machine learning sorter to sort the product list in the search set according to the likelihood of purchase.
2. Reorder the product list based on advertising rates to balance seller sales speed achieved through promotions with recommendation relevance.

Ranking model

The ranking model is based on the following historical data:

Data of recommended products
• Recommended products similar to the seed product
• Context (Country and Product Category)
• User Personalization Features

eBay uses gradient boosting trees that, for a specific seed item, Sort items based on their relative purchase probability.

From binary feedback to multiple correlation feedback

In the past, purchase probability relied on binary purchase data. It is "relevant" if it is purchased with the seed item, otherwise it is "irrelevant". This is a failed approach, but there are several major areas where it can be optimized:

False Negatives: Since users typically only purchase one item from the recommended list, the purchase does not occur when the purchase is not made. In some cases, good recommendations may be viewed as bad recommendations, leading to false positives.
• Few purchases: Compared to other user events, it is becoming challenging to train a model with sufficient number and diversity of purchases to predict the forward class.
• Missing data: From clicks to add to cart, numerous user actions reveal a wealth of user information and reveal possible outcomes.

To summarize, eBay engineers consider the following user actions, in addition to initial clicks and how to add them to the ranking model:

Buy Now (only applicable At Buy-It-Now i.e. BIN Listing)
• Add to Cart (BIN Listing Only)
• Bid (Best Bid Listing Only)
• Ask for Bid (Applies to Auction Listings only)
• Add to Watchlist (Applies to BIN, Best Bid, or Auction Listings)

User Interface Example

Relevance Levels of Multiple Relevance Feedback

eBay now knows that purchases are extremely relevant, so it needs to add additional actions, but the new question is: where do these actions fall within the relevance hierarchy?

The image below illustrates how eBay sorts the remaining possible actions - "Bid," "Buy Now," "Add to Watchlist," and "Add to Cart."

In the historical training data for seed items, each potential item is labeled with a relevance level by the following scale.

Marked result is that during training, the sorter penalizes incorrectly ordered purchases more severely than incorrectly ordered "Buy Now" , and so on.

Sample weights for multiple correlation feedback

Gradient boosted trees support multiple labels to capture a range of correlations, but there is no direct way to achieve the magnitude of the correlation.

eBay had to run the tests iteratively until it came up with numbers that made the model work. The researchers added additional weights (called "sample weights") that were fed into the pairwise loss function. They optimized the hyperparameter tuning and ran it for 25 iterations before arriving at the best sample weights - "Add to Watchlist" (6), "Add to Cart" (15), "Bid" (38 ), "Buy Now" (8) and "Buy" (15). Without sample weights, the new model will perform worse. With sample weights, the new model outperforms the binary model.

They tried adding only clicks as additional relevant feedback and applied the tuned hyperparameter "Purchase" sample weight of 150. Offline results are also shown below, where "BOWC" stands for Buy Now, Make a Bid, Add to Watchlist, and Add to Cart actions. Purchase ranking reflects the average ranking of items purchased. The smaller the better.

Conclusion

The trained model has a total of more than 2000 instances. A/B testing is conducted in two stages. The first phase, which only included additional select tags and showed a 2.97% increase in purchase volume on the eBay mobile app and a 2.66% increase in ad revenue, was deemed successful enough to move the model into global production.

The second phase added more actions to the model, such as "Add to Watchlist", "Add to Cart", "Bid" and "Buy Now", and A/B testing showed better customer engagement (e.g. more clicks and BWC).

Original title: ​​EBay Uses Machine Learning to Refine Promoted Listings​​​, Author: Jessica Wachtel​

The above is the detailed content of eBay uses machine learning to improve sale listings. For more information, please follow other related articles on the PHP Chinese website!