性视界

Pinterest is a social network and image sharing service where people discover and save images, that in August 2020 had more than 400 million monthly active users[1] saving more than 240 billion of pins[2]. From increasing the precision of its ad targeting to recommending better pins[3] the appears on the users鈥� home pages, machine learning is crucial for Pinterest. As noted by Vijay Narayanan, Pinterest鈥檚 Head of Discovery and Content, 鈥渢here are AI projects focused on things like self-driving cars, but there鈥檚 also the everyday, accessible AI that helps people live better lives now鈥�[4].

Pinterest realized the importance of AI already in 2015, when it acquired Kosei, a machine learning start up that specialized in a technology that drives content discovery and makes highly personalized product recommendations, through recommendation algorithms and a system containing 400 million of linkages between billion of products[5].

Pinterest uses machine learning to identify content in line with items previously pinned by users and to recommend new products to its users. Its algorithms therefore inspire people by proposing them items that they might not have been initially searched.

Technology with a purpose

Pinterest motto is to develop technology with a purpose: accessible AI should be able to solve an identified problem. Indeed, the company鈥檚 mission is 鈥渢o help people discover and do what they love[6]. Therefore, Pinterest uses a learning model that, by understanding the intentions behind uses research is able to deliver highly personalized results. This personalization is the main value creation component of the company. Indeed, 80 percent of Internet users are likely to make a purchase if their experience is personalized[7].

The Discovery Problem

In order to assess the users鈥� intention, Pinterest had to disentangle the 鈥渄iscovery鈥� problem, understand what people are looking for from the three or less words they input in the search bar. While the number of Pinterest users grows and the number of items saved has crossed 100B, Pinterest need to 鈥渂uild technology to not only keep up, but make recommendations smarter鈥�[8]. Therefore, it developed PinSage, a neural network placing each image, according to a specific theme, within a graph of other images. It allows building a context for each imagine, through which Pinterest offers thematic visual recommendation to its users. In turn, thanks to PinSage, instead of a list of results, people receive a guide of personalized recommendations.

Machine learning in the home feed through Pinnability

The growing number of saved items posed another crucial challenge: How does Pinterest surface the most personalized and relevant pins? Pinterest developed Pinnability, a collection of machine learning models that support users in find the most relevant content in their home feed. Pinnability 鈥渆stimates the relevance score of how likely a Pinner will interact with a Pin鈥�[9] and it allows accurate predictions, through which the Pinterest team prioritizes Pins based on a relevance score. Therefore, Pins are not showed anymore in a chronological order as it used to be.

Engagement Abroad

Initially, Pinterest’s machine learning models were not targeted for users outside the United States, so that people with similar items saved in different countries had a similar prediction score. However, Pinterest realized that its models should be changed to expand outreach to users abroad, and designed a language detection model bases on country and language match features, including 鈥渨hether the Pinner鈥檚 language and country are the same as a Pin鈥檚 language and country, and if the Pinner鈥檚 language is among the top languages spoken by others who saved the Pin鈥�[10]. Using this method, Pinterest improved the number of items saved by international users by 10 to 20 percent[11].

The Future

In order to leveraging on its technology, Pinterest is using machine learning for a number of other applications and operations, such as for eliminating harmful and negative content from the platform, detecting spam content, executing ad performance and relevance prediction analysis. However, to win the competition with other social media which are entering the realm of targeted ad, such as Instagram, Pinterest needs to continue investing resources in research and development to design cutting-edge iterations of its model able to offer customized recommendations to its users.

 

 

[1] 聽Fiegerman, Seth. 2020. 鈥淢ore than 400 million users around the world connect to Pinterest each month (Number of the day)鈥� in CNN Blog, August 3, 2020, Available at: CNN

[2] Pinterest website. Available at:

[3] Pins are bookmarks that people use to save ideas they love on Pinterest.

[4] Wired. 201. How Pinterest uses AI to capture our imagination. Available at:

[5] Constine, Josh. 2015. 鈥淧interest Acquires Machine Learning Commerce Recommendation Engine Kosei鈥� in TechCrunch, January 21, 2015. Available at:

[6] Pinterest website. Available at:

[7] Wired, cit.

[8] Pinterest Engineering Blog. 2018. PinSage: A new graph convolutional neural network for web-scale recommender systems. Available at:

[9] Pinterest Engineering Blog. 2015. Pinnability: Machine learning in the home feed. Available at:

[10] Pinterest Engineering Blog. 2017. How machine learning significantly improves engagement abroad. Available at:

[11] Ibidem.