Research

Dissertation: Value-Aligned Recommendation Systems

My dissertation addresses the gap between what users say they prefer (stated preferences) and what their behaviors suggest they prefer (revealed preferences). Social media algorithms infer user preferences from engagement behaviors, even when those behaviors don’t reflect what users actually value, often amplifying content that is attention-grabbing but misaligned with users’ true values.

How can we design social media feed algorithms that better align with what users consciously value? And what effects might such value-aligned systems have on individuals and society? To address these questions, I develop a stated preference-based recommendation algorithm that leverages large language models (LLMs) to re-rank content based on users’ explicitly expressed values. I test its effects on both individual- and system-level outcomes through randomized controlled trials and large-scale simulations.

Ongoing Projects

In addition to my dissertation, I have led and collaborated on several projects that explore how digital platforms shape political behavior and discourse:

A digital field experiment on Twitter/X examined how muting low-quality sources influences users’ information consumption and downstream political attitudes.
As a finalist team member in the Prosocial Ranking Challenge , I contributed to designing a recommendation algorithm that fosters constructive, cross-partisan dialogue online.
In a collaborative project, I study how real-time political conversations—embedded with LLMs—can reduce affective polarization.

Future Directions: Human-AI Interaction for Democratic Engagement

Building on these lines of work, I am expanding my research into the domain of human-AI interaction, with a focus on civic applications. I explore how AI can be leveraged to support more prosocial and pro-democratic engagement online—for example, by facilitating constructive political dialogue or helping bridge partisan divides.

Methodologically, I am also interested in integrating AI into social science experiments. This includes AI-augmented survey designs that simulate dynamic conversations or deliberative settings at scale, enabling the study of opinion change, persuasion, and engagement in more interactive environments.

A field experiment to mitigate the harm of inaccurate health information online

Muting low-quality sources vs. Receiving media literacy tips

<!DOCTYPE html> 20240207 20240207 Do Won Kim 2024-02-07 Updated Preliminary Eligibility Check Brief summary of data As of Feb 2 (9:30 AM), I downloaded entries in mercury_user (=832) and following_result (=721) tables from the DB. I first filtered out those participants with the following criteria: (1) those who failed to follow our study account, stored in following_result table; and (2) those without vsid field in the mercury_user table, which denotes that they didn’t finish the Wave 1 survey to the end....

<!DOCTYPE html> Eligibility Check (Preliminary) Eligibility Check (Preliminary) Do Won Kim 2024-02-02 Preliminary Eligibility Check Brief summary of data As of Feb 2 (9:30 AM), I downloaded entries in mercury_user (=832) and following_result (=721) tables from the DB. I first filtered out those participants with the following criteria: (1) those who failed to follow our study account, stored in following_result table; and (2) those without vsid field in the mercury_user table, which denotes that they didn’t finish the Wave 1 survey to the end....

<!DOCTYPE html> 240204 240204 Do Won Kim 2024-02-04 Preliminary Eligibility Check Brief summary of data As of Feb 2 (9:30 AM), I downloaded entries in mercury_user (=832) and following_result (=721) tables from the DB. I first filtered out those participants with the following criteria: (1) those who failed to follow our study account, stored in following_result table; and (2) those without vsid field in the mercury_user table, which denotes that they didn’t finish the Wave 1 survey to the end....

<!DOCTYPE html> Gap between Verasight’s raw data vs. server/DB data Gap between Verasight’s raw data vs. server/DB data Do Won Kim 2024-02-08 Problem The gap between eligible survey sample filtered from raw data (698) vs. what we have in our DB/server (866). Some background information: In the end block of the survey, I embedded codes to send vsid to the DB. Hence, those users with vsid in the DB have finished to the end of the Wave 1 survey....

<!DOCTYPE html> 70% stratified random sampling - Option 2 Code Show All Code Hide All Code 70% stratified random sampling - Option 2 Do Won Kim 2024-02-22 1 Option 2 Randomize within strata by home timeline tweets instead of followers - better captures actual treatment Cutoff at 95% (accounts that take up to 95% of total sum number of followers) = 486 accounts Order by number of times appeared on users’ home timeline (not number of followers) => exposure...

<!DOCTYPE html> Attrition Analysis Code Show All Code Hide All Code Attrition Analysis Do Won Kim 2024-05-16 # Load packages ================================================================ if (!require("pacman")) install.packages("pacman") pacman::p_load( conflicted, # Avoid conflict of functions with same names tidyverse, # Tidyverse umbrella package haven, # read_dta() infer, # chisq_test() sjPlot, # tab_xtab() kableExtra, # kbl() DT # datatable() ) conflict_prefer("select", "dplyr") conflict_prefer("filter", "dplyr") # Import data ================================================================== # Load survey data df_w3 <- readRDS("survey_data/2023-095c_client_wave-3.rds") # wave 3 df_w2 <- read_dta("survey_data/2023-095b_client_wave-2....

<!DOCTYPE html> Eligibility Check Code Show All Code Hide All Code Eligibility Check Do Won Kim 2024-02-11 1 Initial step I compared Verasight’s raw data with our DB data to filter out those (1) who failed to finish the survey to the end, and (2) who took the survey several times with multiple Verasight accounts but authorized the same Twitter account each time. Please refer the this notebook for this initial step: https://do-won....

<!DOCTYPE html> Eligibility Check (Additional Analysis) Code Show All Code Hide All Code Eligibility Check (Additional Analysis) Do Won Kim 2024-02-12 1 Checking Assumption Question: Do people who follow zero qualifying accounts have consistent enough exposure in their home timeline to be worth including? By adopting Following OR Hometimeline as our eligibility criteria, we assume that people are repetitively/consistently exposed to indirect low quality tweets. That is, if we find indirect low-quality tweets from users’ home timelines (i....

<!DOCTYPE html> Eligibility Check Final Code Show All Code Hide All Code Eligibility Check Final Do Won Kim 2024-02-24 1 Description of Data I compared Verasight’s raw data with our DB data to filter out those (1) who failed to finish the survey to the end, and (2) who took the survey several times with multiple Verasight accounts but authorized the same Twitter account each time. Please refer the this notebook for this initial step: https://do-won....

<!DOCTYPE html> Verasight | Checking for potential pop-up blocker issue Code Show All Code Hide All Code Verasight | Checking for potential pop-up blocker issue Do Won Kim 2024-01-20 Data Filtered out suspects (=potential bots) Filtered in progress 100% and 14% cases (Total N=71) 14% (N=34) 100% (N=37) library(readr) library(tidyverse) library(stringr) # Load data dta <- read_csv("raw-responses_2024-01-17.csv") Suspects_Sheet1 <- read_csv("Suspects - Sheet1.csv") names(Suspects_Sheet1) = c("userid", "screename", "vsid", "notes") Suspects_Sheet1 |> mutate(flag = "suspects") -> Suspects_Sheet1 dta |> left_join(Suspects_Sheet1[,c(2,5)], by="screename") |> filter(is....

Progress Dashboard Progress Summary (Oct 23 - Nov 12) Completes Progress: Final target: 13,223 (per budget) ⓘ Overall project target number Expected progress (at 119.1/day x 21 days): 2502 ⓘ Target number we should have reached by now (daily target × days elapsed) Current total: 261 ⓘ Actual cumulative number achieved so far Current achievement rate: 10.4% ⓘ Achievement rate for current period (current total / expected progress) Progress to final target: 2....

Progress Dashboard Progress Summary (Oct 7 - Dec 12) Completes Progress: Final target: 13,223 (per budget) ⓘ Overall project target number Expected progress (at 119.1/day x 52 days): 6195 ⓘ Target number we should have reached by now (daily target × days elapsed) Current total: 743 ⓘ Actual cumulative number achieved so far Current achievement rate: 12.0% ⓘ Achievement rate for current period (current total / expected progress) Progress to final target: 5....

<!DOCTYPE html> Regression Analysis Code Show All Code Hide All Code Regression Analysis Do Won Kim 2025-02-12 # Load packages if (!require("pacman")) install.packages("pacman") pacman::p_load( conflicted, # Avoid conflict of functions with same names tidyverse, # Tidyverse umbrella package readr, # read_csv() arrow, # Load parquet files car, # vif() betareg, # betareg() sjPlot, # tab_xtab() kableExtra, # kbl() broom, # tidy() DT # datatable() ) conflict_prefer("select", "dplyr") conflict_prefer("filter", "dplyr") 1 Data Prep # Set the base directory (When you replicate this, change this to your working directory) base_dir <- "/Users/dowonkim/Desktop/domain_audience_diversity/data" # Define subdirectories for different data categories sub_dirs <- c("age", "gender", "meta_stats", "party", "race") # Generate full paths for each subdirectory dirs <- file....

<!DOCTYPE html> wave1_0207 wave1_0207 Do Won Kim 2024-02-08 [Wave 1] Raw Data A. Comparison of Verasight’s summary vs. raw data The raw data files contain: 4,032 survey starts, including completes and incompletes (anyone who got to the consent page) library(tidyverse) library(DT) library(haven) df <- readRDS("~/Downloads/2023-095a_files/verasight_wave1/2023-095a_client_wave-1.rds") nrow(df) ## [1] 4032 3,517 completes including attention check failures (per Verasight) df |> filter(Progress==100) |> count() ## # A tibble: 1 × 1 ## n ## <int> ## 1 3517 3,064 completes excluding attention check failures (per Verasight)...

<!DOCTYPE html> visualization Code Show All Code Hide All Code visualization Do Won Kim 2024-03-20 1 Scatter Plot Ro’ee: I think it would be useful at some point to create a scatter plot with WTA in round 1 in the x-axis and round 2 in the y-axis (and also just report the correlation). It’s a bit surprising that so many people provided a different answer, we should figure out if this is related to the question wording....

<!DOCTYPE html> visualization2 Code Show All Code Hide All Code visualization2 Do Won Kim 2024-04-22 1 Data Description library(tidyverse) library(ggplot2) library(DT) library(haven) library(ggpubr) library(cowplot) library(ggthemes) library(readr) library(ggExtra) library(scales) library(Rmisc) library(ggbeeswarm) # to make jitter plots library(RColorBrewer) # for the colors library(patchwork) library(ggstatsplot) library(ggsci) # load raw data df_w1 <- readRDS("~/Downloads/2023-095b_files/2023-095a_client_wave-1.rds") df_w1_supp <- readRDS("~/Downloads/2023-095b_files/2023-095a_client_wave-1_supp.rds") df_w2 <- readRDS("~/Downloads/2023-095b_files/2023-095b_client_wave-2.rds") # bind wave 1 data + filter only Progress == 100 df_w1 = rbind(df_w1, df_w1_supp) df_w1 |> select(-user_id) |> filter(Progress==100) -> df_w1 # invited users list (for merge) Users_to_invite_Wave2 <- read_csv("Users_to_invite_Wave2....

<!DOCTYPE html> visualization Code Show All Code Hide All Code visualization Do Won Kim 2024-03-12 1 Scatter Plot Ro’ee: I think it would be useful at some point to create a scatter plot with WTA in round 1 in the x-axis and round 2 in the y-axis (and also just report the correlation). It’s a bit surprising that so many people provided a different answer, we should figure out if this is related to the question wording....

<!DOCTYPE html> WTA_distribution Code Show All Code Hide All Code WTA_distribution Do Won Kim 2024-03-13 1 Data Preparation w1_users: Wave 1 responses of users who took wave 2 survey so far (247 obs) w2_users: These same respondents’ wave 2 responses Based on the above data frames, I made a merged dataset wta_df! library(tidyverse) library(DT) library(haven) library(readr) # load raw data df_w1 <- readRDS("~/Downloads/2023-095b_files/2023-095a_client_wave-1.rds") df_w1_supp <- readRDS("~/Downloads/2023-095b_files/2023-095a_client_wave-1_supp.rds") df_w2 <- readRDS("~/Downloads/2023-095b_files/2023-095b_client_wave-2.rds") # bind wave 1 data + filter only Progress == 100 df_w1 = rbind(df_w1, df_w1_supp) df_w1 |> select(-user_id) |> filter(Progress==100) -> df_w1 # invited users list (for merge) Users_to_invite_Wave2 <- read_csv("Users_to_invite_Wave2....