Distributed inference of latent groups for temporal behavior prediction

Understanding drivers of behaviors can help us build better predictive models. Much debate has been dedicated to how social influence permeates through social networks. While on the one hand, social influence is thought to be localized within communities, other work has demonstrated that social influence can diffuse through weak ties within networks. In focusing on how social influence flows through direct ties, however, the literature has missed the broader potential of leveraging social groups, which also exert social influence on their members. Here, we leverage recent psychological findings suggesting that social influence and social groups, rather than depending on direct ties, are the result of the latent communities that one infers by observing others’ preferences and behaviors. We adapt this model for use for large-scale data and introduce a new framework, the Distributed Latent Group Influence Model, wherein we leverage this idea across a large network to predict behavior. We incorporate these latent communities and their heterogeneous influence on members into a model predicting behavior across time. We test this model against other potential accounts on a longitudinal behavioral data set and find that these inferred latent communities (rather than self-reported network structures) can predict significant variation in behaviors. Importantly, this framework allows us to model social influence without needing predetermined, self-reported social network structures—a common feature of current research. It also allows us to integrate networks, individual decisions, and characteristics into an interpretable model from which we can derive insights and accurately predict shifting behaviors over time.

Representation of how future behavior arises for one individual.

The differing color mixture for each individual represents their degree of affiliation to different global hidden communities. These latent groups have different sets of characteristics, exemplified here by the blue latent group typically driving and spending time outdoors, and the orange latent group as biking and going to music events. People identify to these groups with varying strength based on their behaviors and traits. People’s behaviors are governed by a combination of their personal preferences, the preferences of their neighbors, and the preferences of others in their global, hidden communities.

Graphical model of behavior prediction process for an individual i at time t.

The coloring of variables indicates the stage of the process to which they belong. Green indicates the first stage, where we conduct a distributed inference of latent groups on smaller sets of agents. Yellow indicates the second stage, where we aggregate our distributed latent groups to global groups and derive the average and spread of behaviors of group members. Blue indicates the third stage, where behavior prediction is done through elastic net using the predictor variables indicated. Orange indicates our goal—predicting future behavior for a given individual.