Source-linked AI summary
Detection of Cyberbullying Incidents on the Instagram Social Network
Homa Hosseinmardi, Sabrina Arredondo Mattson, Rahat Ibn Rafiq, Richard Han, Qin Lv, Shivakant Mishra
TL;DR
Cyberbullying on Instagram is an important problem involving harmful image and comment interactions, but detecting it requires distinguishing repeated, power-imbalanced bullying from general aggression. The paper labels multimodal Instagram sessions, analyzes associated features, and evaluates automated classifiers. A Linear SVM using text, images, and metadata reaches 87% accuracy, while the classifier remains limited to highly negative media sessions.
Problem
The paper addresses the need to understand and automatically detect cyberbullying in Instagram’s image-and-comment environment, where cyberbullying differs from general cyberaggression.
Method
The authors label Instagram images and associated comments for cyberbullying, cyberaggression, and image content, analyze feature relationships, and evaluate multi-modal classifiers.
Results
87% accuracy was achieved by a Linear SVM incorporating text, images, and metadata for identifying cyberbullying.
Takeaways & Limitations
About 48% of highly negative Instagram media sessions were not deemed cyberbullying by majority vote, showing that substantial negative content does not constitute online cyberbullying.
Takeaways & Limitations
The current classifier is designed only for highly negative media sessions; a general classifier would require substantially more labeled data.
Abstract
from arXiv · showhide
Cyberbullying is a growing problem affecting more than half of all American teens. The main goal of this paper is to investigate fundamentally new approaches to understand and automatically detect incidents of cyberbullying over images in Instagram, a media-based mobile social network. To this end, we have collected a sample Instagram data set consisting of images and their associated comments, and designed a labeling study for cyberbullying as well as image content using human labelers at the crowd-sourced Crowdflower Web site. An analysis of the labeled data is then presented, including a study of correlations between different features and cyberbullying as well as cyberaggression. Using the labeled data, we further design and evaluate the accuracy of a classifier to automatically detect incidents of cyberbullying.
Introduction
The paper addresses cyberbullying on Instagram, where harmful image-based and textual interactions can have serious psychological consequences. It distinguishes cyberbullying from cyberaggression, analyzes labeled multimodal data, and evaluates classifiers for detection.
- Cyberbullying in online social networks is associated with potentially devastating psychological effects, including depression, low self-esteem, suicide ideation, and suicide.
- Instagram enables cyberbullying through humiliating images, hateful comments, aggressive captions or hashtags, and fake profiles.
- Cyberbullying requires repeated intentional aggression and an imbalance of power, whereas cyberaggression broadly concerns intentional digital harm.
- The paper labels Instagram images and associated comments under both the restrictive cyberbullying definition and the broader cyberaggression definition.
- The study analyzes relationships between cyberbullying or cyberaggression and textual, visual, behavioral, and social-network features.
- Multi-modal classifiers are evaluated using text, image, and metadata feature sets to detect cyberbullying.
Data Collection
The study constructs a public-Instagram dataset, then selects media sessions with enough comments and high negativity for tractable human labeling. The selected sessions differ in comment distributions and user-following patterns from the complete set.
- 41K Instagram user IDs yielded about 25K public profiles forming the complete dataset of typical Instagram users.The collected data included posted media, up to 150 associated comments, and follower, following, commenting, and liking relationships.
- Media sessions were selected when they had at least 15 comments and more than 40% negative comments from users other than the profile owner.
- 998 media sessions remained available for labeling after selection, with an average of about 64.3 comments per media.The initial criteria reduced the pool to about 1,203 sessions; unavailable sessions had been deleted or made private or deleted.
- The selected subset contains a higher fraction of media sessions with 15–50 comments, while its distribution resembles the complete set above 50 comments.The authors conclude that relatively higher negativity tends to occur with moderate numbers of comments.
- Selected users have more followed-by users than users in the complete set, while their follows distribution shows the same pattern but ends around 7,500.The selected distributions end around 4.16 ∗10^6 followed-by users, compared with nearly 10^8 for all users.
Cyberbullying Labeling
The study labels Instagram media sessions using both images and associated comments, distinguishing cyberbullying from cyberaggression and incorporating image-content categories. Five contributors evaluate each session after quality-controlled instructions and filtering.
- Cyberaggression is intentional digital harm, whereas cyberbullying additionally requires repeated negativity and an imbalance of power.
- The labeling survey presents each image with its associated comments and asks whether the media session constitutes cyberaggression and cyberbullying.
- Each media session is labeled by five contributors who receive definitions, examples, qualification quizzes, and in-job test questions.
- The study selects image categories to characterize visual content that may help classify cyberbullying in highly negative-comment sessions.
Cyberbullying
The labeling analysis shows that negative comments alone do not reliably indicate cyberbullying, while cyberbullying labels generally remain consistent with cyberaggression labels. Support for cyberbullying is also associated with comment volume and frequent short-interval postings.
- About 48% of highly negative media sessions receive two or fewer cyberbullying votes, so a majority criterion would classify nearly half as not cyberbullying.
- About 40% of sessions receive four or five cyberbullying votes, about 38% receive zero or one, and about 22% receive two or three.
- Cyberbullying votes never exceed cyberaggression votes, consistent with cyberbullying being a subset of cyberaggression.
- Cyberbullying strength has correlation 0.17 with followed-by count but no significant correlation with likes, total shared media, or followings.
- Cyberbullying support correlates about 0.4 with frequent postings within one hour, while correlations weaken as the allowed interarrival duration expands.
Image Labeling Analysis
Among highly negative Instagram media sessions, Person/People, Text, Sports, and Tattoo are common image labels across cyberbullying-support levels. Multi-label analysis shows that Person/People images are often exclusive but sometimes also receive Text labels.
- Person/People, Text, Sports, and possibly Tattoo are the most common image-content labels across most cyberbullying-support levels.
- More than 60% of Person/People images receive no additional category, while about 15% also receive the Text label.
- Figure 10 examines the fraction of other categories assigned to images labeled Person/People.
Classifier Design and Evaluation
The classifier combines comment text, metadata, and image features, with dimensionality reduction before linear SVM classification. Performance improves from the 0.52 baseline to 0.87 accuracy using multimodal features, while simpler metadata reaches 0.71.
- The classifier uses majority voting for labels and incorporates CrowdFlower labeler trust through a weighted confidence level.
- Features include comment unigrams, bigrams, 3-grams, comment counts, short-interval posting counts, image categories, and profile or media metadata.
- 0.71 accuracy is obtained from low-dimensional metadata with Naive Bayes, compared with a 0.52 baseline; adding image categories raises accuracy to 0.72 with recall 0.78.
- 0.87 accuracy with high precision and recall is achieved by combining reduced text, metadata, and image features in a linear SVM.
- The image-category distributions are analyzed for sessions receiving different numbers of cyberbullying votes and cyberaggression votes.
Discussion and Future Work
The paper identifies limits in the current classifier and proposes broader features, wider media-session coverage, expanded labeling, and alternative cyberbullying criteria for future work.
- Classifier scope: The current classifier is limited to highly negative media sessions, so a general classifier covering all media sessions is needed.Expanding coverage would also require substantially enlarging the labeled data set.
- Feature expansion: Future classifiers could add image, commenting-temporal, and mobile-sensor features to improve performance.The authors also plan to automate image-feature extraction with image-recognition algorithms.
- Labeling criteria: The classifier currently uses majority voting, while an alternative definition would classify a session as cyberbullying when at least one labeler identifies it as such.That alternative criterion would require new classifiers.
- Labeling design: Future labeling studies could distinguish cyberbullying roles, including victims, defenders, and bullies.The authors report that streamlining the survey improved response rate, quality, and speed.
Conclusions
The paper distinguishes cyberbullying from cyberaggression in Instagram media sessions and combines image, text, and metadata features to analyze and detect these incidents.
- Definitions and labeling: Cyberbullying is defined using both frequency of negativity and imbalance of power, and is distinguished from broader cyberaggression.The paper applies this distinction in large-scale labeling of Instagram media sessions.
- Labeling findings: About 48% of Instagram media sessions were not labeled cyberbullying by majority vote despite having the highest profanity levels.This finding indicates that substantial negative content did not constitute online cyberbullying under the study’s criterion.
- Labeling findings: Labelers mostly agreed on which Instagram media sessions constituted cyberbullying, and a substantial class exhibited cyberaggression without cyberbullying.The findings support analyzing the two behaviors as related but distinct categories.
- Correlation analysis: Cyberbullying support correlated strongly with the number of text comments and with comments posted within one hour of one another.These correlations involve comment volume and temporal commenting behavior in Instagram media sessions.
- Classification: 87% accuracy was achieved by a Linear SVM using multimodal text, image, and media-session metadata features.The classifier was evaluated for identifying cyberbullying incidents.