Abstract
In online social networks, activity feeds frequently showcase friends' actions, leading to the perception that social influence drives user behavior. However, distinguishing between genuine social influence (mimicry) and homophily (the tendency of similar individuals to associate) is a significant challenge in social data science. This paper presents a novel statistical framework, Preference-based Matched Estimation (PME), designed to untangle these factors. By evaluating data from four major social platforms—Last.fm, Goodreads, Flixster, and Flickr—the researchers demonstrate that actual "copy-influence" accounts for less than 1% of total user actions. These findings suggest that personal preferences are far more influential than social feeds in shaping online activity.
Introduction
The "chicken-and-egg" problem of social influence asks: Do friends influence your behavior, or are you friends because you already share similar tastes? In online environments, activity feeds provide constant exposure to friends' likes, ratings, and listens. When a user performs an action similar to a friend's recent activity, it is often labeled as social influence. However, standard observational studies frequently conflate influence with homophily, leading to inflated estimates of how "contagious" behaviors truly are. This research aims to provide a more accurate measure of social influence by isolating it from pre-existing similarities.
Methodology: Preference-based Matched Estimation (PME)
To separate influence from homophily, the authors developed the Preference-based Matched Estimation (PME) procedure. The core logic involves comparing a user's behavior against two distinct feeds:
- The Real Feed: The actual actions performed by the user's friends.
- The "Similar Stranger" Feed: A simulated control feed created from "matched non-friends"—strangers who exhibit nearly identical past activity and preferences to the user's actual friends.
The difference between the overlap in the real feed (Friends-Overlap) and the simulated feed (Strangers-Overlap) provides the true estimate of copy-influence. Since strangers cannot influence the user, any overlap with them must be attributed to shared preferences or external trends.
Key Findings
The study's results across diverse domains—music, books, movies, and photos—reveal several critical insights into the nature of online social influence.
The Overestimation of Influence
Naive estimates that fail to account for homophily overestimate social influence by 4x to 10x. As shown in Figure 2, the gap between simple friend overlap and the actual copy-influence estimate is substantial. Most instances of "following a friend" are actually two individuals with similar tastes acting independently.
The <1% Threshold
Across all four platforms, copy-influence accounted for less than 1% of total user actions. On Last.fm, for instance, only 0.3% of "loves" were attributable to social influence. For more than 75% of users, the copy-influence estimate was zero or negative, indicating that most users are largely unaffected by the routine suggestions in their activity feeds.
Susceptibility and User Activity
The impact of social influence is not uniform across all users. The researchers found that less active users are generally more susceptible to influence. As a user's activity increases and their personal preference profile becomes more established, the relative impact of their friends' feed diminishes.
Platform and Item Variability
Influence levels vary by the type of content. "Quick consumption" items, such as photos on Flickr or songs on Last.fm, show higher relative influence than "high investment" items like books or movies. This suggests that users are more likely to copy low-effort actions than those requiring significant time or financial commitment.
Conclusion
This research highlights that while social feeds are a prominent feature of modern platforms, their actual power to change user behavior is significantly lower than previously thought. By providing a robust method to distinguish influence from homophily, the PME framework offers a path toward more accurate social analytics and better-targeted recommendation engines that can distinguish when to prioritize social versus content-based suggestions.
This summary is based on the paper presented at CSCW 2016 by Amit Sharma and Dan Cosley.
Summary updated