DATA607: Week11 Discussion/Homework

Ebay Recommender system

Ebay is an online market place that brings sellers and buys together to buy and sell virtually anything. It has about 1.2 billion items and 179 million buyers. As other similar platforms such as Amazon, ebay has also setup intelligent systems to help buyers to carry out their intended shopping activities without being under the burden of information overload, helping them achieve rich shopping experience. From one of their blog posts, they acknowledge that in addition to the challenge of their large scale user base and products catalog, there is limited structured data attributes, such as ISBN, available for their items, making it difficult to use traditional collaborative filtering approaches for generating recommendations

Visitng the website as an anonymous user (using incognito mode in Google Chrome) first shows something like the image shown below. There are yet no clicks on any item or any shopping history.

knitr::include_graphics('https://raw.githubusercontent.com/henryvalentine/MSDS2019/master/Classes/DATA%20607/Home%20works/Week11/ebay-front-page.png')

Then I made a search for Asus Zenbook laptop, and selecting one item from the results took me to the item’s details page. iIn the section Immediately below the item’s details appears a group of items tagged “People who viewed this item also viewed…”

knitr::include_graphics('https://raw.githubusercontent.com/henryvalentine/MSDS2019/master/Classes/DATA%20607/Home%20works/Week11/ebay-rec.png')

This enumerates items that are in similar categories as the selected item but there is no where to see stuffs like addons/enhacements/accessories associated with the selected item. It then means I will need to make extra searches and go through the results so as to make a choice from among them.

Also, just as I have experienced before after purchasing an item on Ebay and few minutes later, the item no longer exists on Ebay though it was still shipped, till date, I have not seen it again on the site. It then appears Ebay trully has a volatile inventory and this can impact recommendation of similar items to customers based on their shopping history. This implies that for them to get their system right, they will need more than what collaborative filtering methods can offer. Though Ebay has a filter option to select price range when making searches, I don’t think they make recommendations based on prices but rather majorly on related items categories. Moreover

After I accessed the site with my user account, the recommender system poped up with Items related to my previous browsing activities

knitr::include_graphics('https://raw.githubusercontent.com/henryvalentine/MSDS2019/master/Classes/DATA%20607/Home%20works/Week11/rec-user.png')

Senario Design

Target Users: These are buyers and sellers who come together on Ebay to exchange items for money

Their key goals:

1. Sellers: They need an enabling platform where they can sell their merchandise whether new or used to a wider audience
2. Buyers: Need a platform where the can get whatever they need to buy from a wide price options

How to help them accomplish these goals

Through providing them a platform that is very easy to interact with, and making it as much as possible for users to be able to access the items that interest them the most

Ebay provides the platform as well as enforce policies to ensure that as much as possible, transactions are conducted between people as they should.

Reverse engineering

Category level Aggregation:

Since ebay has a volatile inventory, it makes sense that items should be aggregated at the category level. eBay has a category taxonomy and all items belong to a specific leaf category in their category tree. User purchases can be aggregated to form the user-category matrix, where the columns represent leaf categories and the entries in the matrix are either 1 if the user has purchased from that category or 0 otherwis, and then use cosine similarity, with appropriate thresholds, to find the top-K nearest categories to the input seed category.

Recall sets

After related categories have been filtered, the next step is to generate the actual candidate recommendation items. Ebay terms a set of such candidate items as a Recall Set. The input to generating the recall sets is the information about the item selected by the user. This is a very strong piece of context, so it is imperative that the recommendations shown to the user have some relevance to the seed item, just as the seed category is used to generate a set of related categories.

Under recall sets, candidate items for recommendations are generated using a variety of signals such as:

Related Products:

This recall set uses the collaborative filtering approach aggregated at the product level. From Ebay’s perspective, an item is different from a product. “An item refers to any listing posted by a seller, while a product at eBay is defined as a concrete entity in the real world. For instance, for books, think of the ISBN number. Having product information for items allows many items to be aggregated to the same product entity.”

If the seed item can be mapped to a product, A recall set of related products can be generated by taking the cosine similarity of vectors of implicit feedback in the form of product-level purchase data.

Since items are the final result that need to be shown to the user, Ebay uses a separate mapping from products to items that is stored in a cache. This product-to-item mapping is generated by aggregating the most-viewed items for a given product, which incorporates a popularity signal into the results.

Co-views:

This is a recall set that utilizes the behavioral signal of the product page. “An item purchase is an ultimate sign of user intention”. Whereas a view signal carries less intention, since a user might just be navigating the product pages, the benefit to using this signal is the sheer increase in volume/coverage of recommendations. For the fact that co-view data is dense enough, this signal is used to generate recall sets at the product level and directly at the item level. “Recall sets that use the co-view signal are high quality in terms of conversion.”

Recommendations

I think it would be nice if the recommender system shows addons/enhacements/accessories associated with certain items like laptops as this can help the customer have a complete experience with minimal efforts.

Also, I am of the opinion that it can still be relatively easy to make recommendations for similar items taken down by sellers to customers who have purchased those items. This can be done through extracting the items’ details and making recommendations for items that are in the same categories even further drilling down to the ones that share similar price tags

Ebay is a multi-seller platform where competion among sellers is likely to abound. I think making recommendations that consider product price (including discount indication) and seller’s reputation can help improve the amount of items which buyers are willing to pay for.

References:

1. Complementary Item Recommendations at eBay Scale

2. Personal price aware multi-seller recommender system: Evidence from eBay