Research
Ongoing Work
Predicting Self Efficacy using Social Media Content
In this study, we are trying to establish a relationship between the general self-efficacy of a user and their social media interaction by utilizing our novel dataset.
Published Work
Predicting Fans’ FIFA World Cup Team Preference from Tweets
In this paper, we propose a multi-level weighted ensemble model to predict users’ FIFA world cup supporting teams from their tweets.
After collecting supporters of 7 different playing nations such as Argentina, Brazil, Croatia, England, France, Germany, and Portugal, we analyze their tweets and build two different types of classifiers by using Linguistic Inquiry and Word Count (LIWC) and Embeddings from Language Model (ELMo) word embedding based techniques. As our independent models, we use Random Forest on LIWC features and a deep-learning-based LSTM model on the ELMo representation of the tweets. These classifiers predict which team a user prefers from her word usage patterns in tweets.
Finally, we build a multi-level weighted ensemble model by combining the above-mentioned classifiers. Our ensemble classifier achieves better accuracy of 12.68% and 20% than that of independent ELMo and LIWC based model, respectively.
A Context-Based Searching Technique by Extraction and Fusion of Metadata of Digital Photos
This work demonstrates an approach for context-based searching of digital photos using natural language. The study aims to provide an easily accessible platform that is different from other existing works, to preserve and retrieve the memory of a digital photograph.
The key aspects of our system are to collect context and content-based metadata from the photos and facilitate searching and browsing options for them according to the accumulated data.
Here, object, person, location, EXIF data, etc. are content-related metadata, and weather tags, time of the day, actions/activities of the persons present in the photo, etc. are classified as contextual metadata. By using these internally and externally generated metadata, we construct a description of each photo and enable context-based searching option in natural language for them.
Comparison of Machine Learning Algorithms to Recognize Human Activities from Images and Videos Using Pose Estimation and Feature Extraction
In this research, we present the importance of pose estimation and visual feature extraction techniques for classifying human actions from images.
Here, we utilize six different types of machine learning algorithms (i.e., Logistic Regression, Random Forest, KNN, SGD, SVM, and MLP Classifier) for image classification, and show a comparative analysis among them.
