In the era of digital communication, social media platforms have become more than socialization hubs, evolving into powerful domains for political activism. Images play a significant role in capturing attention and stimulating user enagement. Understanding how different types of images influence user interaction can provide valuable insights for researchers, content creators and platform developers. This project aims to delve into image based engagement on Reddit through image clustering and regression analysis. I hope to shed light on visual factors that drive engagement.
Image Clustering
Regression Analysis
- After observing different college subreddits, some possible category clusters will be: memes, active protests, images featuring cops, school communications, screenshots from other social media.
- For regression analysis, it is hypothesized that images featuring cops will reflect the most positive enagement. This hypothesis stems from the national media coverage and public discourse triggered by the intervention of law enforcement at encampents.
The Mobilizing Power of Visual Media Across Stages of Social-Mediated Protests by Yingdan Lu and Yilang Peng found that crowd-based photos predicted the most engagement when observing the Black Lives Matter, Stop Asian Hate and Women's March protests on Twitter. It is likely that the same will be found when studying Reddit posts, however it will be interesting to see what other types of images produce the most engagement.
- import praw
- import pandas as pd
- from PIL import Image
- import requests
- keywords = ['palestine','protest','israel','encampent']
- subreddits = ['ucla', 'UCSD', 'UCSB', 'UCSC', 'UCI', 'UCDavis', 'ucmerced', 'ucr', 'berkeley','columbia','Harvard','yale','princeton','georgetown','nyu','UIUC','aggies','ucf','Purdue','ASU','rit','uofm','rutgers','VirginiaTech','csuf','CSULB','CSULA','humboldtstate','Drexel','EmersonCollege','Emory','IndianaUniversity','mit','riceuniversity','stanford','PennStateUniversity','Berklee','bostoncollege','BostonU','Bowdoin','BrownU','Caltech','calarts','CalPoly','chapmanuniversity','Cornell','gwu','gatech','LSU','LMU','washu','VirginiaTech','Vanderbilt','UoP','udub','USC','usfca','UPenn','seattleu','scrippscollege','scad','SJSU','Pepperdine','pace','OregonStateUniv','jhu','harveymudd','Fordham','CUNY','claremontcolleges','Grinnell','Gonzaga','cmu','CalPolyPomona','CSUS','AmericanU','amherstcollege','baylor']
- Formed df with Subreddit, Image URL, Upvotes, Date
- Retrieved images from links manually and added them to Google Drive, then uploaded them to Colab
- import numpy as np
- import os
- import cv2
- from keras.applications import ResNet50
- from keras.applications.resnet50 import preprocess_input
- from sklearn.cluster import KMeans
- from sklearn.decomposition import PCA
- import matplotlib.pyplot as plt
- Images:
- resize:(224,224)
- convert from BGR to RGB
- used preprocess_input to preprocess images according to ResNet50 model
- extracted features using model
- normalized features
- Manually added image cluster category to each post in df
- Made new df with only image cluster category and Upvotes
- Applied min-max normalization since there were different amounts of each category
- Tried k=3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20
- Ended up choosing k=4, clusters made the most sense at k=4
- Hard to determine theme from larger amount of clusters
- Used Normalized data set:
- Added upvotes by image type
- Used .idxmax to find which one had the most upvotes
- Misc: 5.668547
- Protest: 11.971039
- School: 3.825000
- Screenshots: 4.557554
- Tried: Random Forest, Linear Regression, Ridge Regression
- Ended up using Gradient Boosting Regression, because it performed the best.
- Coefficients:
- Intercept: 0.23190770384441436
- ID_Misc : 0.03411019786445862
- ID_Protest : 0.029837666143125045
- ID_School : 0.021450598118525465
- ID_Screenshots : 0.014661104828747051
- Intercept: Presents baseline predicted number of Reddit Upvotes. If a Reddit post contained none of the types of images, the model predicts a baseline number of upvotes of approximately 0.232
- ID_Misc: Model expects an increase in number of upvotes by approximately 0.034 units for Miscellaneous Posts
- ID_Protest: An Active Protest image expects and increase in predicted number of upvotes by 0.030 units.
- ID_School: School Communications images expect an increase in predicted number of upvotes by 0.021 units.
- ID_Screenshots: Screenshots type images predict an increase in number of upvotes by 0.015 units.
Main image categories found in posts were Miscellaneous (memes, maps, protest posters), Active Protests, School Communications and Screenshots. In this particular study the most succesful images were Active Protest images, however a regression analysis was done to predict potential enagement. The intercept provides the expected number of upvotes for a post without any specific type of images, while the coefficients quantify the additional impact of including specific type of images on the predicted number of upvotes. In this study Active Protests were the most enagaging similarly to the study done by Lu and Peng, however the regression model predicts that Miscellaneous images (memes, maps, protest posters) will have the greatest positive effect on engagement with Active Protests coming in second.
- The Reddit API only allowed accesss to recent post data
- Only one protest was observed
- The dataset is small
- Retry project with access to more posts/older posts
- Observe more protests
The Mobilizing Power of Visual Media Across Stages of Social-Mediated Protests by Yingdan Lu & Yilang Peng: Analyzed approximately ten million tweets related to three social mediated protests (BLM, Stop Asian Hate, Women's March) to investigate the impact of different image types on message engagement throughout various stage protests. They found that posts containing images and videos generally attracted higher audience engagement compared to text-only posts. Utilizing unsupervised image clustering, the researchers identified four categories: crowd-based protest photos, non-crowd-protest human photos, non-human photos and non-photograph visuals. They observed heterogenous effects on audience engagement across these content categories and protests, with visuals producing more engagement during the ignition phase of the protest than before or after.
- Lu, Y., & Peng, Y. (2024). The Mobilizing Power of Visual Media Across Stages of Social-Mediated Protests. Political Communication, 1–28. https://doi.org/10.1080/10584609.2024.2317951
- PyTorch,"resnet50model", Retrieved from https://pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html, 2024
- scikit-learn.(2024).GradientBoostingRegressor Documentation. Retrieved from https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.GradientBoostingRegressor.html
- PRAW (Python Reddit API Wrapper).(2024). "PRAW Documentation," Retrieved from https://praw.readthedocs.io/en/stable/