Learning Social Influence
with Unobserved Attention

Modeling Complex Systems I at UVM
November 17th, 2026

Some Guy who wandered in

Hi everyone! I’m Phil Chodrow


  • BA in philosophy + theoretical math
  • PhD in operations research(?),
  • Postdoc in applied math,
  • Faculty in computer science(??), Middlebury College(!)

I teach courses in discrete math, data science, machine learning and network science.

I like aikido, tea, my cats, being outside, reading, cooking, and cool math stuff (modeling + data science).

What’s cool to me (lately)


Dynamical systems: deterministic evolution of a system over time. \[ \mathbf{x}_{t+1} = g_{\color{#17868F}{\theta}}(\mathbf{x}_t) \] \(\theta\) is a system parameter that tunes the dynamics.

What is the long-term behavior of a system governed by evolution rules \(g_\theta\)?

What’s cool to me (lately)


Stochastic processes: noisy evolution of a system over time. \[ \begin{aligned} X_{t+1} \sim \mathbb{P}\left(\cdot | X_t = x_t; \theta\right) \end{aligned} \]

Common application: tune the probability structure so that \(X_t\) has a certain specified distribution as \(t \rightarrow \infty\). (Markov chain Monte Carlo, simulated annealing)

What’s cool to me (lately)


Stochastic processes with latent variables: noisy evolution of a partially observed system \[ \begin{aligned} X_{t+1} \sim \mathbb{P}\left(\cdot | X_t = x_t, {\color{#17868F}{U_t = u_t}}; \theta\right) \end{aligned} \]

Here, \(U_t\) is some process in the system which we postulate but do not observe.

Big idea: we don’t need direct data on a phenomenon to include it in our models.

Today

stochastic dynamics with latent variables for modeling social influence




From 0-5, how’s this talk?













Autokinetic effect: a still point of light in a dark room will appear to move if you stare at it long enough.

“So, how far did the light move?”

Sherif (1935)

“So, how far did the light move?”

Sherif (1935)

Social Influence

Any change in an individual’s thoughts, feelings, or behaviors caused by other people, who may be actually present or whose presence is imagined, expected, or only implied.

- American Psychological Association (APA) Dictionary of Psychology, 2026

Some Other Social Influence Experiments


Asch social conformity experiments (1950s)

Face-to-face group gaslighting works, even on obvious problems.

“That intelligent, well-meaning young people are willing to call white black is a matter of concern.”

Crutchfield experiments (1950s)

Face-to-face interaction not required: works if the participants only think they are interacting with the group opinions.

Modern experiments

Many modern experiments study intervening role of internet, social media, other technological media of opinion exposure.

Image adapted from Wikimedia Commons, CC BY-SA 4.0

Image adapted from Wikimedia Commons, CC BY-SA 4.0

+ LLMs, virtual reality environments, etc. etc.

Social Influence

Any change in an individual’s thoughts, feelings, or behaviors caused by other people, who may be actually present or whose presence is imagined, expected, or only implied.

- American Psychological Association (APA) Dictionary of Psychology, 2026

To formalize models of social influence, we need to decide:

  • Which other people?
  • What kind of change?

Towards Formalization: Which Other People?

Towards Formalization: What Kind of Change?


Discrepancy

0 inches 10 inches Opinion x: how much
has the dot moved?

\(x_t\): opinion of the target \(x_s\): opinion of a source at discrepancy \(d\)

Kernel

The kernel \(k(d)\) sets how far the target moves

\(\delta = k(d) + \text{noise}\)

Outcome

The target’s opinion moves by \(\delta\)

Slide h/t Will Thompson

A first model: random neighbor choice and linear influence

DeGroot or linear kernel (DeGroot 1974): \(k(d) = \lambda d\).













Across a broad range of model choices, linear influence with \(\lambda > 0\) tends to produce consensus.

Many Kernels, Many Mechanisms

Linear

Compromise –> consensus

convergence

Empirical justification

  • Vande Kerckhove et al. 2016
  • Takács et al. 2016
  • Clemm von Hohenberg et al. 2017
  • Friedkin & Bullo 2017
  • Friedkin et al. 2021

Repulsion

Compromise with similar opinions, repulsed by distant opinions

polarization

  • Mazen & Leventhal 1972
  • van Knippenberg et al. 1990
  • Krizan & Baron 2007
  • Liu & Srivastava 2015

Bounded confidence

Compromise with similar opinions, ignore distant opinions

clusters

  • Moussaïd et al. 2013

Or do it nonparametrically!

Rahimi + Recht (2007)




Slide h/t Will Thompson

Some common implicit assumptions




At each timestep each agent
always
interacts with
all
of their neighbors.

Core idea for today: selective attention




At each timestep each agent
always sometimes
interacts with
all some
of their neighbors.

Stochastic dynamics with latent variables! We typically don’t observe the interactions: we only observe the opinion changes that result.

The Team

Will Thompson
UVM

Moyi Tian
CU Boulder

Linh Huynh
Texas Tech

Alex Wiedemann
Hamline University

Vicki Modisette
Eastern Kentucky

Phil Chodrow
Middlebury College

Heather Brooks
Harvey Mudd

Work initiated at the AMS Mathematics Research Community on Complex Social Systems (2023)!

Three Innovations

Adams 2022, n = 3,760 updates

Standard: average over all neighbors

Novel: listen to one neighbor, chosen at random

Standard: you always update

Novel: sometimes you don’t

Standard: Gaussian noise (least-squares methods)

Problem: real updates often appear heavy tailed

Novel: select among noise models like Laplace or Gaussian

After Kozitsin (2021)

I can move toward a neighbor but away from the neighborhood average

Moussaïd et al. (2013): far apart, people keep their opinion

→

Chacoma & Zanette (2015): keeping happens at every distance

Ignore retention and the kernel looks like bounded confidence

Slide h/t Will Thompson

Learning the model: maximum likelihood

Our stochastic model defines a likelihood function over observed opinion trajectories, given an observation of a fixed network of possible interactions:

\[ \begin{aligned} \overbrace{\mathbb{P}(\mathcal{D}; \theta)}^{\text{Data set}} &= \underbrace{\mathbb{P}(X_{0} = x_{0})}_{\text{Initial state}}\prod_{t=0}^{T-1} \overbrace{\mathbb{P}(X_{t+1} = x_{t+1} | X_{t} = x_t; \theta)}^{\text{Transition probabilities } q_\theta(x_{t+1}, x_t)} \\ q_\theta(x_{t+1}, x_t) &= \sum_{u \in \mathcal{U}_t}\mathbb{P}(X_{t+1} = x_{t+1} | X_{t} = x_t, \underbrace{U_t = u}_{\text{unobserved\\mechanisms }}; \theta)\mathbb{P}(U_t = u | X_{t} = x_t; \theta) \end{aligned} \]

We’ll try to learn this model by solving an optimization problem: \[ \begin{aligned} \hat{\theta} = \arg \max_{\theta} \mathbb{P}(\mathcal{D}; \theta) \end{aligned} \]

Learning the model: expectation-maximization (Dempster et al., 1977)

We want a sequence \(\theta^{(0)}, \theta^{(1)},\ldots \rightarrow \hat{\theta} = \arg \max_{\theta} \mathbb{P}(\mathcal{D}; \theta)\).

E-step: Update our beliefs about the unobserved mechanisms.


\[ \begin{aligned} z^{(k)}_{t,u} = \frac{\mathbb{P} (X_{t+1} = x_{t+1} | X_{t} = x_t, U_t = u; \theta^{(k)}) \mathbb{P}(U_t = u | X_{t} = x_t; \theta^{(k)})}{\sum_{u' \in \mathcal{U}_t}\mathbb{P}(X_{t+1} = x_{t+1} | X_{t} = x_t, U_t = u'; \theta^{(k)}) \mathbb{P}(U_t = u' | X_{t} = x_t; \theta^{(k)})}\;. \end{aligned} \]

M-step: Update the model parameters to maximize the expected log-likelihood.


\[ \begin{aligned} \theta^{(k+1)} = \arg \max_{\theta} \sum_{t=0}^{T-1} \sum_{u \in \mathcal{U_t}} z^{(k)}_{t,u} \log \left[\mathbb{P}(X_{t+1} = x_{t+1} | X_{t} = x_t, U_t = u; \theta) \mathbb{P}(U_t = u | X_{t} = x_t; \theta)\right]\;. \end{aligned} \]

Learning the model: model selection

We select between competing models (kernels, mechanisms, etc.) by evaluating their predictive likelihoods on held-out validation data.

Our Data

Controlled, single source

Controlled, network

Social media

Prediction markets

Congress

ADAMS
lab, pairs · also TAKACS, BECKER

VK-counting
lab, groups of six · also VK-gauging

CMV
Reddit threads · also SPINOS, VKontakte

MARKETS
Manifold

SENATE
roll calls, cosponsorship · also HOUSE

controlled lab experiments

the social wilds

Slide h/t Will Thompson

Impact of Attentional Mechanisms

Example: Senate political opinions

latent selection

latent retention

heavy tail

baseline

Adding attention changes our conclusion about the shape of the kernel!

Slide h/t Will Thompson

Latent mechanisms make models more predictive

One bar per dataset: its held-out gain over a random-walk null, segmented into mechanisms by order of addition.

Influence kernel Presence of any social influence (rather than random drift)

Heavy tail Selected on 13/15 datasets; on average 28% more predictive power per update.

Latent retention Selected on 13/15 datasets; on average 21% more.

Latent selection Selected on 6/8 networked datasets; on average 8% more.

On average, heavy tails, retention and selection tend to matter more than the actual kernel shape!

Adding attention makes

the model more predictive (on heldout data!) on 14/15 datasets

the inferred kernel less linear on 6/15 (upper left)

the inferred kernel more linear on 2/15 (lower right)

Slide h/t Will Thompson

Some Learnings


Potential exposure \(\neq\) interaction

In experiments and online settings, relationship networks (proximity, following, etc.) define possible rather than actual interactions.

Model latent mechanisms

We can model what we don’t observe directly: we just need to tie it statistically to things that we do observe.

Flexible modeling

Stochastic dynamics + algorithms offer us a rich toolbox for learning about complex systems.

Thanks everyone!



Will Thompson
UVM

Moyi Tian
CU Boulder

Linh Huynh
Texas Tech

Alex Wiedemann
Hamline University

Vicki Modisette
Eastern Kentucky

Heather Brooks
Harvey Mudd