mimee.xyz

Logo
Mimee Xu

PhD Student at New York University

View My GitHub Profile

A cursory, informal outline of a tech policy project I am working on. (*) means nuanced claims that deserve more real estate than what I provide.

Introduction

This project aims to measure the effect of an adversary on an election. The main dynamic of concern is that of a consensus: everyone gets a vote and the winner of the count takes all.

WTS

We aim to bring clarity to a class of causal questions over sparse graphs.

A socially relevant curiosity is to quantify the effect of social media influence, and how much power an adversary would have on election outcomes.

Many subquestions follow this train of thought, in relation to real concerns in technical policy:

  1. Should we ban all sub-communities that encourage self-harm on Twitter as to reduce self-harm in teens?
  2. Are there effective tests to see if a platform or a new social media feature will cause outsized influence?
  3. Does the bidding algorithm in ad-targeting over a social network behave degenerately for some subpopulation over time?
  4. Is setting a proper size limit of Facebook groups useful for limiting adversarial manipulation?
  5. Is it appropriate to punish those who spread "fake news"? If so, how do we quantify harm?

(eventual) Goal

Practical policy recommendations with regards to emerging technology for the integrity of social systems.

Contributions

A novel modeling approach to complement existing techniques in political science on the effect of emerging technology.

Approach

To answer the interdisciplinary problem, we take an eclectic approach, borrowed from economics, social dynamics, data science, with consultation w/ political scientists. This section is broken down into causal models, dynamical systems, network model, belief model, and a summary of its comparative advantage.

Causal Models

Causal modeling is used as an alternative to direct experimentation -- after all, performing trials as adversaries to societal integrity is generally considered unethical (*). Given that, we need some other way to model the effect of potential adversaries. Causality is thus a major part of this picture. For clarity, this section is divided between "zoomed-out" and "zoomed-in" perspectives.

Aggregate Question

Given that information flows through platform amplification over a population, how does the resulting consensus result get affected?

Notably, this question becomes relevant after 2016. There the election result was surprising to many statisticians.

Passive For a sufficiently surprising election result, baselined against Initial Condition e.g. expected, projected, polled result, it is commonly known as a 'swing'.

This model is the base of further interactions.

Election Swing <-{ Network Topology, Initial Condition, New Information}

Amplified Simply add amplification.

Platform Amplification-?-> Election Swing <-{ Network Topology, Initial Condition, New Information}

Note that in causal modeling, this graph "includes" the interactions between causes such as new information X platform amplification.

Adversarial We add adversary to this model.

Platform Amplification -> Election Swing <-{ Network Topology, Initial Condition, New Information, Adversary}

Individual Influences

Only starring at the aggregate graphs, there are some unanswered questions, so we zoom in on an individual at a particular moment. This involves heavy handed "belief modeling", but clarifies time and graph dependencies on the way platform amplification acts on an individual's decision to participate (i.e. Turnout).

(Prev.) Turnout -> Future Turnout <- Received Information

(Prev.) Belief -> Future Belief <- Received Information

{::nomarkdown}

Network Modeling

For a population of N agents and M broadcasters that form graph G of (N+M) nodes, with belief and turnout probability, information flows passively along the edges. This is the setup upon which platform amplification and adversarial activities are added.

Empirically, individuals over the same network may see dramatically differing experiences due to locality in the network.

Platform amplification effects are highly related to network topology, as information flows along the edges of the graph. In addition, altering the underlying recommendation algorithms along the graphs can mediate the effects of platform amplification.

Note Influences along graphs can be calculated in many ways. In social media platforms, however, the ranking among people is often centrality-based.

Dynamical System

The change of belief and turnout under varied influences is characterized by a discrete time simulation. This simulation is a relaxation of continuous-time changes.

Per individual, belief is a unitary vector over various "issues" i.e. the dictionary, while turnout is a probability.

Per time step, platform amplification effect re-normalizes turnout probability along the graph through two forms

  1. Enhancing well-connected agents
  2. Promoting certain broadcasters at random

Belief

Belief is a high dimensional vector over various "issues". The parametrization of belief satisfies some nice properties, such as flexibility to model through a survey with non-orthogonal basis, stability over time, etc.

Turnout vs. Belief vs. New info

Belief is a transformation acting on Turnout, whereas new information is normalized to act on both belief and turnout. Note that belief is considered stable. Empirically, this means regardless how much beliefs on various issues change, where the voter lies on a fundamental (low dimensional) political spectrum does not over a short period of time.

Confirmation bias

Belief itself, though usually hidden, is important to characterize, because people are subjected to confirmation bias i.e. misplaced causal attribution that causes change of belief only when the new information has sufficient alignment with current belief.

Pitfalls of existing techniques

Here are some of the issues I saw that the current model addresses.