[Paper Review] Yoga-Veganism: Correlation Mining of Twitter Health Data

Tunazzina Islam|arXiv (Cornell University)|Jun 15, 2019

Eating Disorders and Behaviors18 citations

TL;DR

This study applies topic modeling (LSA, NMF, LDA) to 40,000 health-related Twitter tweets to uncover hidden patterns and correlations in lifestyle behaviors. It identifies a significant correlation between yoga and veganism, achieving 66% training accuracy and 51% test accuracy, with model interpretability enhanced via LIME for future analysis of misleading predictions.

ABSTRACT

Nowadays social media is a huge platform of data. People usually share their interest, thoughts via discussions, tweets, status. It is not possible to go through all the data manually. We need to mine the data to explore hidden patterns or unknown correlations, find out the dominant topic in data and understand people's interest through the discussions. In this work, we explore Twitter data related to health. We extract the popular topics under different categories (e.g. diet, exercise) discussed in Twitter via topic modeling, observe model behavior on new tweets, discover interesting correlation (i.e. Yoga-Veganism). We evaluate accuracy by comparing with ground truth using manual annotation both for train and test data.

Motivation & Objective

• To uncover hidden correlations in health-related social media content, particularly between lifestyle practices like yoga and plant-based diets.
• To apply topic modeling techniques (LSA, NMF, LDA) to extract dominant themes from unstructured Twitter data.
• To evaluate model performance using manual annotation and compare with random baseline to validate topic inference accuracy.
• To identify and analyze misleading or unrelated topic predictions to improve model interpretability.
• To explore the feasibility of using streaming data pipelines (Kafka, Spark Streaming) for real-time health trend monitoring on social media.

Proposed method

• Collected 40,000 health-related tweets using Twitter’s streaming API with keywords related to fitness, diet, and wellness.
• Used Apache Kafka as a distributed message queue to stream and process high-volume Twitter data in real time.
• Employed Spark Streaming for parallel and distributed data processing to handle the large-scale tweet stream.
• Applied three topic modeling algorithms—Latent Semantic Analysis (LSA), Non-negative Matrix Factorization (NMF), and Latent Dirichlet Allocation (LDA)—to infer semantic topics from tweets.
• Conducted manual annotation of 100–500 tweets from train and test sets to create ground truth for accuracy evaluation.
• Used LIME for model-agnostic interpretability to analyze and explain misclassified or misleading topic predictions.

Experimental results

Research questions

• RQ1 What are the dominant health-related topics discussed in Twitter data, and how do they cluster across different lifestyle categories?
• RQ2 How accurately can topic modeling algorithms (LSA, NMF, LDA) infer the true thematic content of health-related tweets compared to human-annotated ground truth?
• RQ3 What hidden correlations exist between lifestyle behaviors such as yoga and veganism in public social media discourse?
• RQ4 Why do some topic models assign misleading or unrelated topics to certain tweets, and what factors contribute to such errors?
• RQ5 Can model interpretability techniques like LIME help identify and explain erroneous topic predictions in social media text analysis?

Key findings

• The model achieved 66% accuracy on the training set and 51% accuracy on the test set when evaluated on 500 manually annotated tweets, significantly outperforming the 25% baseline for four topics.
• Topic 1 emerged as the most frequent dominant or second-dominant topic in most tweets, characterized by keywords such as 'diet', 'workout', 'healthy', and 'weightloss'.
• A strong correlation between yoga and veganism was observed in the data, with many tweets mentioning both practices in the same context.
• The model occasionally misclassified tweets due to ambiguous language, such as interpreting 'water in hand' as related to swimming, leading to misleading topic assignments.
• Misleading predictions were identified in 4 out of 10 test cases, highlighting the need for improved interpretability and error analysis using LIME.
• The use of Kafka and Spark Streaming enabled scalable, real-time processing of large-scale Twitter data streams, supporting efficient topic modeling at scale.