In some circles, the mere mention of the word “influencer” is enough to elicit groans of protest. True, folks like Kylie Jenner, Kim Kardashian, Felix Kjellberg (i.e., YouTube’s PewDiePie), and others with outsized social media presences might not have spotless reputations, but they’re pretty effective at peddling merchandise to their legions of follows. In a recent survey conducted by Tomoson, the average business makes $6.50 for every $1 spent on influencer marketing.
The trick, of course, is finding the right ambassadors for a given brand. Artificial intelligence (AI) might hold the key — in a newly published preprint paper on Arxiv.org (“Machine Learning Techniques for Brand-Influencer Matchmaking on the Instagram Social Network“), University of British Columbia researchers describe an AI system that can predict the “most fruitful” partnerships between top social media users and companies.
“Unfortunately, it’s not always easy for small brands to find the right influencer: someone who aligns with their corporate image and has not yet grown in popularity to the point of unaffordability,” they wrote. “[So] we sought to develop a system for brand-influencer matchmaking, harnessing the power and flexibility of modern machine learning techniques.”
To keep things simple, the paper’s authors chose to focus on Instagram influencers. (That’s just as well — in a Linqia study, 92 percent of marketers chose it as the most important platform for influencer marketing.) The researchers built a content analysis tool for user and target brand profiles and collected data from the former with an open source Python tool that downloaded all media — including associated captions and hashtags — uploaded to a user’s account.
For the purposes of experimentation, the paper’s authors scraped a sampling of brand profiles and 20 unique user accounts spanning five themes: dogs, cats, mountains, cars, and pizza. To these, they applied an image classification algorithm (Inception-v3) trained on the ImageNet database in Google’s open source TensorFlow framework, which output a list of the five most likely tags for each image.
Next, the three most likely tags were assembled into a string for each user, which were compiled into a grid. Words were then individually assigned numerical values indicating the frequency of their occurrence in a given profile. Finally, both the influencer and target brand data were fed into a model that spit out suggested matches.
In tests, the AI system performed well. For two target companies competing in the same vertical — Domino’s Pizza and Giordano’s Pizza — it recommended several of the same Instagram influencers but in different orders, demonstrating a degree of nuance in its matchmaking. And for every target brand, it managed to highlight the top influencer by category.
The researchers cautioned that there’s more work to be done, noting that the AI system’s performance has yet to be tested on “less distinct” categories and that its predictions rely on accurate tagging by the image classification algorithm. They also say that the quality of the algorithm’s predictions might be improved with custom parameters and additional Instagram profile metadata, such as the number of followers, average number of likes per post, and so on.
However, they believe that it’s a promising first step toward an automated influencer-finding system that could save businesses time, cash, and a whole lot of headaches.
“The results indicated that our algorithm, when presented with a variety of potential influencer profiles, is able to identify those profiles that are most closely aligned with a particular target brand,” they wrote. “If the wealth of data that exists on social media were to be harnessed, it could … be used to facilitate brand-influencer matchmaking. Not only would this help companies find content creators that align with their brand image, it would also provide an opportunity for the small-time creators to monetize their posts, further encouraging the creation of high-quality future content.”