Understanding Pinterest’s Pin Classification Scheme
Any time you’re trying to maximize performance on a specific digital platform, you’ll need to keep your presentation in sync with what users are actually searching for at that time. This is especially true on Pinterest, which is currently positioning itself as a platform for product discovery, and which is constantly striving to create better connections with user interests and product matches.
This in turn, will theoretically increase users’ likelihood to make purchases, thereby increasing the credentials for Pinterest as an e-commerce site. This being the case, if you really want to maximize performance on Pinterest, you’ll have to understand how a Pin search actually works. You’ll also need to know how to create Pins so they can be synchronized with search and discovery within the app.
In the past, Pinterest has provided several insights into this process, and very recently the engineering team issued a further clarification of its classification processes. This may help to improve your understanding, as well as to steer you toward a better Pin creation strategy.
The classification process
For any particular Pin, data is extracted at a taxonomy level, a label level, and a score. The taxonomy level refers to popular concepts which are prevalent throughout a wide number of Pins, with each one of them containing 10 levels of granularity as well as 24 top level concepts. As an example, a top level concept such as Home Decor would include a number of specific subtopics such as Furniture, Flooring, Ceiling, Home Accessories, and Home Decor Style.
The label part of the classification refers to specific subjects or topics which are identified within an image, and which come primarily from the text which accompanies the Pin. The label can also be inferred by a visual classifier which would replace the Pin text. The score part of the classification refers to the degree of matching with systems within Pinterest, that any given Pin might have for an identified label, based on the attached information.
Other matching characteristics
It’s also true that some of the topics and matches are assumed, based on an understanding gained by Pinterest on historical elements and current trends. This means that while potential matches may not be specifically referred to in the Pin description, they can provide a contextual match based on other similarities. This is a strategy used by Pinterest to match up Pin elements, so as to ensure that all relevant factors are incorporated, and the best match is provided to a user.
What exactly is P2I?
According to the Pinterest engineering team, P2I is an artificial intelligence learning system which is made up of two primary modules, candidate generation and ranking. Candidate generation uses inexpensive, high recall methods to produce interests which may be relevant for each Pin submitted by the user. At this stage of the process, somewhere between 70 and 200 candidates are generated by using methods such as lexical expansion on the text of a Pin.
Lexical expansion may include matching terms by reordering annotations, and adding in all those with a low edit distance, for instance by simply making plurals of the original text. The ranking part of the process involves assigning an actual score to each Pin and interest pair which emerge from the candidate generation phase.
Those with the highest relevant scores are then retained for the final classification output. The ranking process itself extracts features from each candidate pair and assigns a rank to them using a binary classifier. Some of the features which are evaluated include the type of Pin, the country, popularity, gender, and several other factors as well.
Tips to business users
One of the important things for a business user to remember when trying to achieve matches on Pinterest is to use the Guided Search when you’re preparing Pins, and you’re trying to plan for topics and searches that relate to your industry. Just like with ordinary search engine optimization, you don’t want to stuff keywords, and you don’t want to go overboard trying to squeeze in all the key terms in your Pin descriptions.
In the first place, that kind of stuffing is pretty much unnecessary, because Pinterest does much of the work for you. But you also want to ensure that the most relevant keywords are used, because that will help Pinterest do its job better in matching your content to searches and queries submitted by users.
Pinterest also makes a practice of displaying related ads to users, all of which are thought to have strong correlation to the Pin itself. This not only gives users a choice when it comes to product searches, but it has a better chance of including the specific product being searched for by a consumer.
From the engineering team
The Pinterest engineering team tried to emphasize a few points in its recent explanation of the classification scheme used by the platform. It specifically referred to the P2I signal as a mechanism which is used to enforce interest targeting by ensuring that the Pin being queried has many of the same interests as those Pins being targeted.
This allows for Pins with the same interest as the query Pin to be displayed to users, while other potential product candidates are filtered out. This filtering out process is intended to remove those Pins which do not share strong interests with the queried pin, and which would therefore have limited interest or no interest to the consumer. A strategy like this allows Pinterest to display far more relevant ads for every query submitted, and to minimize those displays which have low matching value, and probably low interest to the user.
There’s really no secret trick which can be used to make maximum use of Pinterest matching algorithms. However, by having a better understanding of Pinterest’s own matching process, you should be able to improve on your own marketing approach. That in turn, should help to ensure that your Pin content is displayed to the maximum number of interested users, and that your return is worthwhile for the effort you invested.
Heather Hart
Latest posts by Heather Hart (see all)
- 2022 Top Social Media Networks Compared - June 27, 2022
- Social Media Content Trends for 2022 (The Ones You Should Know About!) - February 28, 2022
- 2021 Social Media Image Dimensions Guide for Facebook, Instagram, and More - January 22, 2021
No comments