Last Updated :
CtoC marketplace, where customers buy and sell items to each other, has a variety of challenges that are different from those in a typical e-commerce website.
C2C marketplaces face various challenges which differ from general e-commerce sites because customers buy and sell items from and to each other.
On Mercari, sellers must handle the procedures for listing items themselves, such as taking photos, entering product descriptions, and setting prices. Because Mercari is a secondary market, no two items are exactly alike, including their conditions and prices. This makes it difficult for customers to research information and appropriate prices for the products they wish to list and complete the listing process on their own.
Mercari uses AI to make it easy for anyone to list their items.
Mercari offers an “AI Listing” function that guesses a product name, brand, product category, and other information from a single photo of an item, which cuts down on the time and effort required to enter product information when listing the item. To implement this function, Mercari has built its own AI catalog that analyzes a wide variety of products listed in the marketplace.
We also provide suggestions and support related to packing, shipping methods, pricing, and more, to help any customer easily list and sell items using Mercari.
To support the entry of item information such as the product name, brand, and item category at the time of listing, we have built an estimation model based on information of items listed on Mercari to date, using the images of listed items. The model extracts features from the image data, and based on the similarity of the feature values, extracts and estimates similar items listed in the past. Because Mercari handles a diverse range of item categories, from fashion and books to household goods, appliances, and cars, being able to properly separate such diverse categories and learn features that represent distances according to similarity is a major challenge. Furthermore, instead of photos like those found in product catalogs, each customer takes photos of items by themselves, resulting in identical items having a variety of images. Item descriptions and item information entered by the customers themselves are also sometimes incorrect or contain information not consistent with the actual product. This presents challenges in terms of label information used in machine learning.
Products and product information are difficult to infer directly from feature values because new items are constantly being listed and the number of classes to be identified, feature information, and more are continually being added and updated. On the other hand, even if information on past listings is used, the current situation, where the volume of listed items continues to grow daily, makes it difficult to handle all items listed in the past. The key factor in improving the accuracy of the various inference functions is how to properly sample and continue to update the item indices for similar searches while taking into account seasonal fluctuations, recent changes in the trends of listen items, and more.
Mercari’s team of machine learning engineers uses actual data to improve these functions on a daily basis.