Understanding the product cycle of discovery to purchase on Pinterest

Rui Huang, Song Cui | Software Engineers, Content Interest Understanding Team, Jennifer Zhao | Software Engineers, Content Core Signal Team, Sai Xiao, Felix Zhou | Software Engineers, Shopping Discovery Team

Pinners have always used Pinterest for shopping, as they often come in a planning mindset and are exploring products and styles. It’s our job to help them go from inspiration to purchase, and so over the years we’ve evolved our shopping products to make it easy to discover products and brands that match individual tastes. As a result, shopping has only grown as a company priority and use case.

Additionally, shopping on Pinterest has increasingly provided retailers with more ways to get their products and brands discovered. We recently launched visual search ads, and more new surfaces for shopping ads, as well as introduced shopping to UK Pinners.

A core piece of our shopping initiatives is Product Pins, which are dynamic and shoppable Pins that direct right to the product page on the retailer’s site. Product Pins also show information about the product right on the Pin itself including the price, availability and retailer. Today there are hundreds of millions of Product Pins on Pinterest. In order to recommend the most relevant products to Pinners, we need to understand the products and the lifecycle of those products, from discovery to purchase. We’ve developed classifiers and systems to understand various product attributes (brand, color, pattern, style, etc). Now we need to understand different types of products, too.


In addition to Search, “Related Products” (found beneath Product Pins) is a key area to discover products. However, in the past, these Pins were not always Product Pins and therefore not always directing to the most relevant content and opportunities to purchase. Imagine seeing a Pin of your dream living room. You love everything about it: the general vibe, specific elements, but most importantly, you need that yellow leather couch. You find an inspirational Pin in Related Products. You click with high expectations, and then you are directed to a product detail page of…a rug. This experience leaves you confused and disappointed.

(1) Scene Pin with yellow leather coach that redirect to the product landing page for rugs

We know one point of confusion for Pinners has been they don’t always know which Pins are shoppable as they browse Pinterest. Even when the intent is clear — you came from a search for “leather couch living room ideas” and clicked the “Shop” tab — it can still be confusing to determine how to purchase the product in that Pin you love.

(2) Search bubble for bedroom furniture ideas to include shop product type to help proactively shopping experience


It was clear we needed to understand the category of the Product Pins with high confidence. We started by building signals to understand our inventory (in stock Product Pins) and Pinners’ shopping intent (based on search queries patterns). We built two signals to serve the need: Pin2ProductCategory and Query2ProductCategory.

However before building the signals, we needed to first define the categories, and base them on the most popular shopping categories on Pinterest: fashion and home decor.

Pin2ProductCategory classifier

We then built a classifier to map our inventory of hundreds of millions of Product Pins into the product categories, and called it the Pin2ProductCategory classifier.


We leveraged both text and visual features, described in detail below.

Product title

From the offline analysis, we saw the product title had much more relative information to the product type compared to other text-based information. To fully use the product title and boost the learning weight, we used the exact match and synonym match from product title keyword to product type category as a sparse categorical feature. In the future, we will explore the FastText model as a text classification model from product title to the product categories as a sparse categorical feature for easier scaling.

Product text with word embedding

We leveraged text extracted by the annotations system. The texts are extracted from Pin title, description, landing page text, and board name and ranked through a GBDT ranker. We looked at the top five terms for each Pin and then aggregated embedding for each annotation term as text features from our in-house multitask text embedding system PinText.

Product image with visual embedding

Our model with text embedding worked decently well for the fashion category, but not for the home decor category. This is because some home decor categories are pretty similar in text, e.g., sofa table, coffee table, end tables, dining tables. However, visually they are quite different due to the shape of the table or the location of the table. This kind of difference can be captured visually, so we used the graphsage embedding as the visual features.

Model architecture

We constructed a 2-layer feed forward DNN model with aggregation of text embedding for the top five words from the Annotation model and visual embedding on the product images.

(3) Model architecture with tokenized annotations, word embedding, visual embedding and product title lexical and synonym match categorical feature 2-layer dnn model.


In order to handle product search, we needed to understand the intent of search queries.

Query2ProductCategory (Q2PC) is a query understanding engine that recommends product categories for broad queries, which can have several different product category purposes. We can use the signal to refine those broad shopping queries and help users make shopping decisions. Take the query ‘living room ideas’, for example. We can show filters like ‘table’, ‘sofa’ and ‘lamps’ which are all from the Q2PC signal. If the Pinner clicks the filter ‘table’, he/she will be directly to a search page where all the shopping Pins are tables under the query ‘living room ideas’.

To build the signal, we collected a one-year search log with all the engaged Pins. By joining the log with the P2PC signal, we can infer the yearly aggregated engagements of Product Categories (PCs) under each query. We then select top k PCs for each query in terms of the engagements. Below is the workflow of Q2PC:


Visual search

One of our visual search products links consumers from what they see in the scene of a Pin to individual products. For a scene image, we decompose it into shoppable objects. By zooming into an image or clicking a specific product via white dots or the shopping tag image, you can see visually similar products. The “Shop Similar” is powered by Pin2ProductCategory classifier.

Product filter in Related Product surface

In the “More to Shop” (Related Products) module, given a query Pin, we will recommend similar products related to it. As mentioned above, Pinners could not always identify which Pins were shoppable. We leveraged Pin2ProductCategory signal to build category filters, which significantly helped Pinners narrow down the recommended results to certain categories. We observed that the long-click per closeup rate (close up here means click to the Pin page, long-click means click from the Pin page to the merchant landing page) in category filtered feed is more than twice as high as unfiltered feed. Here is a comparison for “unfiltered feed” vs. “filtered feed”. We can clearly see that the filter better guides the shopper into the next stage of their shopping journey.

(4) Category filtered feed of Rugs

Shop the Board

Pinners often save product Pins to their boards as they are deciding what to buy. And they’ll revisit the Pins they have saved in order to choose which one to purchase. “Search my Pins” and “Board” surfaces have some of the highest long click through rates (clicks through on a result and remains on that site for more than 30 seconds) and conversion rates on Pinterest. Pinners clearly want to be able to shop the Pins they have saved. They also have extremely high intent, in that it’s clear they liked what they saw, which is why they saved it in the first place.

On top of the general need to shop what you have saved, most of the time the Pin saved isn’t directly actionable. We should be able to show recommendations in this case, whether it’s Product Pins from saved scene images or related products or an out-of-stock product shot.

In the shop tab on boards, we display the Pinner’s saved or viewed product item in a structured view. Specifically, we are introducing a category module that shows shoppable items in saved scene pins by category; and “see alternatives” for recommendations of available products if the Pin they originally saved went out of stock.

(5) Category structured shop the board

Search structured feed

Broad queries, such as ‘living room ideas’ and ‘summer outfits’, usually have a longer path to shoppability, given their exploratory nature.. Since they usually show more than 2 product category intents, we conducted an experiment using Q2PC to recommend product categories with a new bubble-like UI, helping users narrow down their requests. After the experiment, we observed a 10% increase of one of our long click through rates. The average click rate of the bubble is 40% higher than the pin placed at the same position before.

Visual search ads

Just this week we introduced ads into visual search surfaces, including camera search results and the results that appear when visually searching within a Pin. This interactive feature is a perfect match with shopping since the Pinner can crop any object inside the image to find related products.

Display shopping ads on this surface will enable the Pinner to have a better shopping experience on Pinterest when they’re searching visually similar products from a lifestyle image. Pin2product type is also used as a filter for retrieval ads candidates to check the product type match.


Product Category is widely used to power the Shopping surfaces at Pinterest with bundle signals:Pin2ProductCategory, Query2ProductCategory. We are looking forward to more adoptions!


We want to thank everyone who has contributed to this project: Alan Li (PM) & Rui Li (EM), Weiran Li (Shopping Discovery Engineer), Raymond Lee (Visual Search Team Engineer), Hanfei Ren (Shopping Product Engineer), Chen Hu(Shopping Ads Engineer).

Understanding the product cycle of discovery to purchase on Pinterest was originally published in Pinterest Engineering Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source link