Personalization

Summary

The personalization project aims to use many types of data at Mercari to provide content tailored to each of our customers at the right timing. The project presently involves building and improving a platform to control content displayed on the home screen. This kind of product evolution helps the customer be able to easily find the product or information they currently want.

Purpose

Suggest items optimal for each customer from the huge number of items listed on Mercari each day should improve the Mercari experience. Each customer wants different things, but using the large amount of data we have should enable us to provide items that match each customer's needs.
While it's important to provide items that match customer needs, we believe letting users encounter new products that they haven't noticed on their own helps make Mercari be more enjoyable to use. This project aims to be able to recommend good items in the broad sense to users at suitable times and in a suitable form.

Description

Recommending items

A huge amount of items are listed every day at Mercari. The recommendation function allows customers to encounter items of interest. Beyond using search to look for a product, having customers encounter items through recommendations is important to building an effective marketplace.

However, because Mercari is a large-scale CtoC marketplace, there are many situations where methods commonly used in e-commerce don't work. We have come up against many problems with a conventional recommendation system. For example, at Mercari we have had only sold-out items which are one-of-a-kind items that are quickly bought being recommended, as well as not being able to recommend at practical times because we're trying to perform complex calculations for items with a large amount of variety. In order to solve these problems, the personalization project has used trial-and-error to look for methods that suit Mercari.

We currently combine three steps in the following manner to be able to effectively provide items customers are interested in.

By grouping items into several topics, we overcome difficulties with conventional recommendation techniques and can better recommend items to customers.

Using topic data and customer data enables us to estimate the topic the user is now interested in, and extract from a large amount of items only those that are important to the customer. After using topics to narrow down the items somewhat, we use machine learning methods to more finely recommend items suited to the customer. By recommending topics of interest to our customers by guessing what they might want to know, we can also indicate why a recommendation was made when an item is recommended. This helps us realize a recommendation function that is friendlier to users.
In this manner, we have been making various improvements in this project to recommend items suited to our customers so that we can effectively and efficiently recommend large amounts of items.

Optimizing the home screen

There is a lot of content and functionality in the Mercari app (collectively referred to as content below). The content that each customer wants differs by person and by when they use the app. Similarly to item recommendation, we are optimizing the home screen so that we can capture these differing needs and deliver matching content in the right form.

The home screen is the front door that customers open when they first visit us. What content to display on the home screen to get customers to continue to come back is important. But, similarly to items, there is an enormous amount of content at Mercari. Beyond our item listings, we have content such as items that customers have liked, fresh information collected by the Mercari content team, information on past listings by the customer, and recommended items automatically generated utilizing data.

This project aims to deliver this varied content when needed by customers and at appropriate positions. We need to place content that people always want to use at the top of the home screen, which is where it will immediately be visible. Meanwhile, we place functionality that's fun to use every now and then beyond the fold on the home screen so that customers can more effectively use the app. Controlling the home screen in this way lets us personalize the home screen.

To this end, the personalization project builds mechanisms to enable us to calculate how to match each piece of content with each user while organizing and unifying the various content we have at Mercari. We are improving our logic daily while putting our many research findings to use, such as with bandit algorithms.