Bias Reduction in Social Networks through Agent-Based Simulations

TL;DR

It is shown that a simple greedy algorithm that constructs a feed based on network properties reduces such perception biases comparable to a random feed and offers a tool for mitigating perception biases through algorithmic feed construction.

Abstract

Online social networks use recommender systems to suggest relevant information to their users in the form of personalized timelines. Studying how these systems expose people to information at scale is difficult to do as one cannot assume each user is subject to the same timeline condition and building appropriate evaluation infrastructure is costly. We show that a simple agent-based model where users have fixed preferences affords us the ability to compare different recommender systems (and thus different personalized timelines) in their ability to skew users' perception of their network. Importantly, we show that a simple greedy algorithm that constructs a feed based on network properties reduces such perception biases comparable to a random feed. This underscores the influence network structure has in determining the effectiveness of recommender systems in the social network context and offers a tool for mitigating perception biases through algorithmic feed construction.

Figures (6)

• Figure 1: Agent-based Model Structure Illustration demonstrating how three users are connected to each other on the network, but will only get exposed to other users through the tweets served to them from the "backend" model.
• Figure 2: Local bias ($B_{\text{local}}$) and Gini coefficient G. Graph depicts the difference between the expected local fraction of friends who have $x=1$ and the true global prevalence of the trait $P(X=1)$. Positive implies over-representation, negative implies under-representation. For G, graph depicts the distribution of times each friend (or friend-of-friend) was observed by a core user. 1 implies inequality, 0 implies equality.
• Figure 3: Precision@10 and Precision@30. Graph depicts the cumulative number of tweets liked in the first K positions seen through tick $t$. Connections seen are reset at $t=24$. For Precision @30, graph depicts the total fraction of liked tweets in the first 30 positions in the feed.
• Figure 4: Log number of friends seen and Precision@30. Graph depicts the log total number of unique friends (and friends-of-friends) seen through tick $t$. Connections seen are reset at $t=24$. For Precision @30, graph depicts the total fraction of liked tweets in the first 30 positions in the feed
• Figure 5: Mean number likes generated by core users.