Project Title

Efficient Clustering for Users’ Brand Sentiment Analysis on Online Social Media

Presenter Hometown

Richmond

Major

Computer Science

Department

Computer Science

Degree

Undergraduate

Mentor

Dae Wook Kim

Mentor Department

Computer Science

Abstract

Brand-themed user-generated content on online social media impacts public opinion about brands through the way brands are portrayed. Visual sentiment expressed in images, a dominant format for online content, affects how content is received by users. We examine efficient clustering of brand-themed user-generated image content with machine learning techniques to better understand visual sentiment related to brands. We use real data from users’ posts related to 90 different brands on the popular social media site Instagram to study sentiment-based image clustering. Recently, considerable progress has been made in areas such as object detection, image recognition, and visual sentiment analysis through the use of Convolutional Neural Networks (CNNs). We use both object-based and sentiment-based features for the task of clustering images based on sentiment. First, we cluster images by the product they represent for each brand using the object-based features. Then, for each product, we use the sentiment-based features to cluster the images according to positive, negative, and neutral overall sentiment. We examine a few clustering techniques and distance measures to improve the clustering results. We achieve promising results with hierarchical clustering for separating positive, negative, and neutral sentiment images.

Presentation format

Poster

This document is currently not available here.

Share

COinS
 

Efficient Clustering for Users’ Brand Sentiment Analysis on Online Social Media

Brand-themed user-generated content on online social media impacts public opinion about brands through the way brands are portrayed. Visual sentiment expressed in images, a dominant format for online content, affects how content is received by users. We examine efficient clustering of brand-themed user-generated image content with machine learning techniques to better understand visual sentiment related to brands. We use real data from users’ posts related to 90 different brands on the popular social media site Instagram to study sentiment-based image clustering. Recently, considerable progress has been made in areas such as object detection, image recognition, and visual sentiment analysis through the use of Convolutional Neural Networks (CNNs). We use both object-based and sentiment-based features for the task of clustering images based on sentiment. First, we cluster images by the product they represent for each brand using the object-based features. Then, for each product, we use the sentiment-based features to cluster the images according to positive, negative, and neutral overall sentiment. We examine a few clustering techniques and distance measures to improve the clustering results. We achieve promising results with hierarchical clustering for separating positive, negative, and neutral sentiment images.