Researchers craft AI framework that explains your social media feeds


Feeds are the cornerstones on which modern music recommenders, news aggregators, and social media platforms are built. If you’re like most people, you spend minutes to hours each day scrolling through songs, clips, articles, questions, public service announcements, and advertisements informed by your interests and preferences. But wouldn’t it be great if the algorithms underpinning feeds (and their recommendations) were a little more transparent?

Researchers at the Max Planck Institute for Informatics thought so, which is why they investigated a framework — Framework for Activity-Item Relationship Discovery, or FAIRY — that systematically discovers, ranks, and explains the connection between users’ actions and what appears in their social media feeds. It’s described in a paper (“FAIRY: A Framework for Understanding Relationships between Users’ Actions and their Social Feeds“) published on the preprint server Arxiv.org.

“Transparency and explainability are key concerns in the modern world of cognitive overload, filter bubbles, user tracking, and privacy risks,” wrote the paper’s coauthors. “Here, a feed results from an intricate combination of one’s interests, friendship network, her actions on the platform, and external trends … Over time, a user accumulates several thousands of actions that together constitute [their] profile (posts, upvotes, likes, comments, etc.), making it impossible for the user to remember all these details.”

FAIRY attempts to solve this dilemma by creating user-specific interaction graphs using information visible to users. It learns models for predicting relevance and surprisal from real-world social media platform data, and then it leverages learning-to-rank techniques to uncover and rank relationships derived from the graphs.

READ  Facebook's DEC AI identified hundreds of millions of fake accounts over 2 years

The features that guide FAIRY are grouped into five sets — user, category, item, path instance, and path pattern — where “path” refers to explanation paths. Users’ influence (e.g., the ratio of followers to followees) is measured as a complement to their activity, as is their engagement with various feed items.

The researchers note that, due to the sheer volume of feed interactions performed by a typical person, graphed entity relationships span from thousands to millions. The aforementioned learn-to-rank approach made them easier to parse by presenting only the top few connections that are either relevant (generally useful as a satisfactory explanation) or surprising (i.e., forgotten or unknown relationships) to users.

In studies, the scientists tasked 20 volunteers to interact with two platforms — Quora and Last.fm — using fresh accounts with five followers each. They spent 20 hours total on at least one of the two services over the course of several sessions, performing a minimum of 12 activities while identifying non-obvious items after scrolling through their entire feed. After each session, the team updated the interaction graphs and selected three non-obvious recommendations per user before mining explanation paths for the feed items.

In a series of tests, the researchers found that FAIRY outperformed three relationship mining baselines on the task of predicting what users considered relevant and surprising explanations. They attribute its success to a “powerful” information network representation of the user’s sphere of influence and the learning-to-rank approach, and they say that the work represents the first step toward a goal of improving transparency with respect to social media feeds.

READ  Kenmore Adds Dash Replenishment to Smart Washers and Dryers

“[I]dentifying explanatory relationships between the users’ online behavior … and the feed items they receive is useful for at least three reasons: (i) they can convince the user of their relevance … (ii) they can point the user toward … a course of action to avoid seeing more of certain kinds of items, and (iii) [they could serve as] a proxy that the users could find plausible,” wrote the coauthors. “For example, if Alice sees a post on making bombs in her feed when she herself is unaware of any explicit connection to such, she might be highly curious as to what she might have done to create such an association.”

In the future, the team plans to implement FAIRY as a browser plugin and to investigate further the effects of users’ activities across multiple platforms.



READ SOURCE

LEAVE A REPLY

Please enter your comment!
Please enter your name here