🔬 physics

Direct and Indirect Influence on Likes in Social Media

Using a large-scale VKontakte dataset, this study reveals that a user's likelihood of liking a post is primarily driven by direct and indirect social contagion from active neighbors within two degrees of separation, with the structural diversity of these active connections serving as a significant predictor of engagement.

Original authors: Ivan Kozitsin, Anton V. Proskurnikov

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ivan Kozitsin, Anton V. Proskurnikov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your social media feed as a giant, bustling town square. You are standing in the middle, and you see a new poster (a post) on a wall. The question this study asks is: What makes you decide to put a "Like" sticker on that poster?

Most people assume it's just about your immediate friends. If your best friend likes it, you might too. But this research, conducted on the Russian social network VKontakte (similar to Facebook), digs deeper to see if your "friends' friends" (your second-degree connections) and the structure of your social circle also play a role.

Here is the breakdown of their findings using simple analogies:

1. The "Echo" from Friends-of-Friends

Think of your social network like a room full of people.

Direct Friends (Distance 1): These are the people standing right next to you. If they clap, you hear it immediately.
Friends-of-Friends (Distance 2): These are the people standing next to your friends. They are one step further away.

The study found that even if your immediate friends haven't clapped yet, hearing the clapping from the people standing next to them makes you much more likely to clap yourself.

It's like a rumor spreading. You might not hear it from your best friend, but if you hear three different people in your friend's circle talking about it, you start to think, "Oh, this must be important," and you join in. The researchers found that this "second-hand" influence is very strong, even if your direct friends are silent.

2. The "Crowded Room" vs. The "Diverse Crowd"

The researchers looked at how your friends are connected to each other.

The "Clique" Scenario: Imagine all your friends who liked the post are all best friends with each other. They are in one tight circle. If they all like it, it's like hearing the same message repeated by the same group of people.
The "Diverse" Scenario: Imagine your friends who liked the post come from totally different groups of people (different schools, hobbies, or neighborhoods) who don't know each other.

The study found that diversity matters more than just the number of likes. If you get a "Like" signal from five different, unrelated groups of people, you are much more likely to like the post than if you get five "Likes" from one single, tight-knit group. It's the difference between hearing a song played by one band in a small room versus hearing it played by five different bands in five different cities.

3. The "Super-Connector" Metric (The $m^*$ Index)

The researchers invented a new way to measure influence, which they call the Extended Induced Index.

Old Way: They used to count how many of your active friends had active friends.
New Way (The Study's Discovery): They counted the maximum number of active friends found across all your friends, even the ones who haven't liked the post yet.

Think of it like a fire alarm system.

The old method only checked if the people currently screaming "Fire!" had neighbors who were also screaming.
The new method checks if anyone in your immediate circle has a neighbor screaming "Fire!", regardless of whether the person next to you is screaming or just standing there.

They found this "New Way" was the best predictor of whether you would click "Like." It captures the "buzz" in the entire neighborhood, not just the people currently shouting.

4. The Surprising Twist: Too Many Direct Friends Can Be Bad

This is the most counter-intuitive finding.
When the researchers looked at the raw numbers, having more direct friends who liked the post made you more likely to like it. However, once they accounted for the "Diverse Crowd" and the "Second-Hand Buzz," something strange happened: Having too many direct friends who liked the post actually lowered your chance of liking it.

The Analogy: Imagine you are trying to listen to a conversation in a noisy room.

If your friends are all in different corners of the room (Diverse), you hear the news from everywhere.
If all your friends are huddled in one tiny corner (High Direct Activity, Low Diversity), the platform's algorithm (the "noise machine") might think, "Oh, this person only cares about that one corner," and it stops showing you posts from the rest of the room.
Because you have limited attention, you miss out on the broader signal. The study suggests that when your direct friends are too concentrated, the platform might narrow your view, making you less likely to engage with the post.

5. The "Three Degrees" Rule

There is a famous idea in sociology that influence can travel up to three degrees of separation (you -> friend -> friend's friend -> friend's friend's friend).

The Study's Verdict: In this specific data set, the "magic" stopped at two degrees.
Influence from your friends' friends (Distance 2) was strong and clear.
Influence from your friends' friends' friends (Distance 3) was too weak to detect. It's like trying to hear a whisper from three rooms away; by then, the signal is gone.

Summary

This paper tells us that on social media, who likes a post matters less than where those likers come from.

Second-hand influence is real: Your friends' friends can push you to like something, even if your direct friends haven't.
Diversity wins: Getting a "Like" from many different social circles is a stronger trigger than getting many "Likes" from one single group.
The Algorithm plays a role: The platform's ranking system likely amplifies these effects, showing you content that your wider network is engaging with, but it might hide content if your immediate circle is too "clumped" together.

The study concludes that to understand why we click "Like," we have to look beyond our immediate friends and consider the complex, diverse web of connections surrounding us.

Technical Summary: Direct and Indirect Influence on Likes in Social Media

Problem Statement
This study investigates the mechanisms of social contagion within online social networks, specifically focusing on the "liking" behavior of users on the VKontakte (VK) platform. While existing literature extensively documents the influence of direct neighbors (distance $d=1$ ) on behavioral adoption, the extent to which indirect influence from friends-of-friends (distance $d=2$ ) independently shapes user behavior remains less understood, particularly in large-scale environments with algorithmic content curation. A central challenge in this domain is distinguishing genuine social influence from confounding factors such as homophily (the tendency of similar individuals to connect) and platform-mediated exposure (where algorithms surface content based on network structure rather than direct user interaction). The authors aim to determine if the probability of a user liking a post depends on the activity of second-order neighbors, even when controlling for the activity of direct friends.

Methodology
The authors utilized a large-scale dataset comprising approximately 290,000 users from the VKontakte platform, specifically focusing on the giant connected component of users in Tomsk Oblast, Russia. The dataset includes friendship ties and chronological records of likes on 308 local news posts, totaling 36,286 likes (19,599 from users within the analyzed network component).

The analysis employs a combination of descriptive statistics and logistic regression modeling. Key methodological components include:

Network Metrics: The study defines and analyzes several metrics to quantify neighborhood activity:
- $k_1$ : The number of active direct friends ( $d=1$ ).
- $k_2$ : The number of active friends-of-friends ( $d=2$ ).
- $s$ : The structural diversity index, defined as the number of connected components among active friends.
- $m$ (Induced Index): The maximal number of active neighbors among the active direct neighbors of a focal node (following Xie et al., 2022).
- $m^*$ (Extended Induced Index): A novel metric defined as the maximal number of active friends across all immediate neighbors of a focal node, regardless of whether those neighbors are themselves active.
Regression Analysis: The authors estimated a sequence of logistic regression models to predict the probability of a user liking a post. These models progressively incorporated covariates including $k_1$ , the density of ties among active friends ( $\rho_1$ ), structural diversity ( $s$ ), and various formulations of second-order activity ( $k_2$ , $m$ , $m^*$ ).
Controls and Filtering: To ensure causal inference regarding timing, the analysis applied chronological filtering, measuring the active neighborhood state immediately before a user's like. Robustness checks were performed by stratifying the data by nodal degree and local clustering coefficients.

Key Contributions

Introduction of the Extended Induced Index ( $m^*$ ): The paper proposes $m^*$ as a superior metric for capturing indirect influence compared to the original induced index $m$ and raw counts of $d=2$ nodes. Unlike $m$ , which conditions on active direct neighbors, $m^*$ accounts for second-order activity reachable through inactive immediate neighbors, thereby capturing exposure pathways that are invisible to the original index.
Empirical Validation of Indirect Influence: The study provides empirical evidence that second-order activity ( $d=2$ ) is a significant predictor of liking behavior, persisting even when conditioning on the activity of direct friends.
Re-evaluation of Direct Neighbor Effects: The analysis reveals a counter-intuitive finding: while the raw number of active direct friends ( $k_1$ ) is positively associated with liking in descriptive statistics, its conditional coefficient becomes negative in regression models that include structural diversity and second-order activity metrics.

Results

Second-Order Influence: There is a substantial positive association between second-order activity and the probability of liking. The extended induced index $m^*$ emerged as the strongest single predictor of liking behavior, outperforming both the raw count of active friends-of-friends ( $k_2$ ) and the original induced index ( $m$ ). The relationship between $m^*$ and liking probability was found to be nearly linear.
Threshold Effects: The original induced index $m$ exhibited a stepwise pattern, with sharp increases in liking probability at $m=1$ and $m=3$ , suggesting potential local reinforcement thresholds.
Structural Diversity: The number of connected components among active friends ( $s$ ) was positively associated with liking probability in models using $m^*$ , supporting the structural diversity hypothesis.
Negative Association of Direct Neighbors: In models controlling for structural diversity and second-order activity, the number of active direct neighbors ( $k_1$ ) showed a negative association with the probability of liking. The authors interpret this not as a suppression effect, but as an indication that $k_1$ acts as a proxy for the concentration of signals within a single community, potentially reducing the effective structural diversity perceived by the user due to limited attention and platform ranking algorithms.
Distance Decay: No significant association was detected for nodes at distance $d \ge 3$ once $d=1$ and $d=2$ activities were accounted for.
Demographics: Male gender was generally associated with a lower probability of liking, though this effect varied in subsamples with high clustering or high degree. Nodal degree was positively associated with liking probability.

Significance and Claims
The paper claims to provide empirical support for induced percolation mechanisms in large-scale online social networks, demonstrating that indirect influence pathways (specifically $d=2$ ) play a central role in shaping behavioral diffusion. The authors posit that the superior explanatory power of the extended induced index $m^*$ suggests that platform-mediated visibility mechanisms (such as recommendation and ranking algorithms) expose users to content endorsed by friends-of-friends, even in the absence of direct signals from immediate neighbors.

The study concludes that higher-order neighborhood structure is critical for understanding behavioral diffusion on online platforms. It suggests that the "three degrees of influence" hypothesis may hold, but the strongest detectable nonlocal signal in this dataset is concentrated at distance $d=2$ . The findings imply that social contagion is not merely a function of direct peer pressure but is significantly modulated by the structural diversity of the network and the visibility of second-order connections mediated by the platform's algorithms. The authors note that their interpretation of platform-mediated exposure remains conjectural due to the lack of access to internal ranking data, but the results are consistent with such mechanisms.