01
Video views
One of the strongest indicators of claim-based content.
Understanding TikTok Claims.
Under-
standing
TikTok
claims.
A classification project designed to identify behavioral patterns that distinguish claim-based TikTok videos from opinion-based content.
TikTok sought to better understand how videos containing factual claims differ from opinion-based content.
Using exploratory analysis, statistical testing, and machine learning, I analyzed engagement and content patterns to identify the strongest predictors of claim status.
01
Overview
From raw data
to predictive
insights.
TikTok sought to better understand how videos containing factual claims differ from opinion-based content. Before predictive modeling could begin, the dataset required extensive exploratory analysis, statistical testing, and visualization to uncover meaningful behavioral patterns.
Across multiple stages of the Google Advanced Data Analytics Professional Certificate, this project evolved from exploratory data analysis into regression modeling and ultimately machine learning, demonstrating a complete analytical workflow from business question to predictive model.
01
The Challenge
Which videos
contain claims?
TikTok needed a more scalable way to distinguish videos containing factual claims from videos expressing opinions. The goal was to identify meaningful behavioral patterns, evaluate potential predictors, and build a model that could help prioritize claim-based content for moderation review.
The project used a dataset of 19,382 TikTok videos across 12 variables, including claim status, author characteristics, video duration, and engagement metrics such as views, likes, comments, shares, and downloads.
Business Question
Can engagement, creator, and video characteristics accurately distinguish claims from opinions?
02
Analytical Workflow
One dataset.
Five stages.
The analysis progressed from initial data inspection through exploratory analysis, statistical testing, regression, and machine learning.
Each stage built on the findings from the previous one, creating a complete analytical workflow from data quality review to model recommendation.
01
Inspect
Reviewed structure, data types, missing values, summary statistics, and engagement distributions.
02
Explore
Used Python and Tableau to compare claims and opinions, visualize distributions, and identify outliers.
03
Test
Conducted a Welch two-sample t-test to evaluate differences in views by verification status.
04
Model
Built a logistic regression model and evaluated coefficients, accuracy, and classification results.
05
Optimize
Compared tuned Random Forest and XGBoost models and selected the strongest champion model.
03
Exploratory Findings
Engagement tells
a clear story.
Claim and opinion videos appeared in nearly equal proportions, but their engagement patterns differed. Claim videos generally received higher engagement.
Views and likes showed a strong positive relationship. Engagement variables were highly right-skewed with many high-value outliers, while video duration was distributed more evenly.
01
19,382 Videos
12 variables inspected
02
Claims ≈ Opinions
Nearly balanced target classes
03
Views + Likes
Strong positive relationship
04
Right-Skewed Engagement
Many legitimate viral outliers
04
Statistical Analysis
Verification status
mattered.
A Welch two-sample t-test was used to compare average video views for verified and unverified accounts.
The test found a statistically significant difference. Unverified accounts averaged approximately 265,664 views, compared with approximately 91,439 views for verified accounts.
2.61 × 10−120
Exact p-value from the Welch two-sample t-test.
265K
Mean views for videos posted by unverified accounts.
91K
Mean views for videos posted by verified accounts.
05
Machine Learning
Random Forest
won.
Random Forest and XGBoost classifiers were trained, tuned with GridSearchCV, and evaluated using precision, recall, F1-score, and accuracy.
Random Forest was selected as the champion model because it slightly outperformed XGBoost while producing fewer false negatives—important when the business goal is to avoid missing videos containing claims.
≈1.0
Near-perfect precision, recall, F1-score, and accuracy across validation and test data.
RF
Champion model selected over XGBoost.
5
Engagement variables ranked among the strongest predictors.
06
Key Predictors
Behavior was more
predictive than identity.
The strongest predictors of claim status were behavioral engagement measures rather than creator identity alone.
Video view count, like count, download count, share count, and comment count contributed most strongly to classification.
02
Likes + shares
Helped distinguish engagement patterns between content types.
03
Downloads + comments
Added further predictive signal to the final model.
07
Business Impact
Decision support,
not auto-moderation.
The model could help TikTok prioritize videos likely to contain claims, allowing moderation teams to focus human review where it is most needed.
The recommended implementation maintains human oversight, monitors model fairness and drift, and retrains the model as new content and behavior patterns emerge.
Reflection
This project showed how exploratory analysis, statistics, and machine learning can work together to move from a broad business question to a practical decision-support system.
Project Files
Explore
Interactive Tableau Story
Explore the complete interactive dashboard created in Tableau Public.Python Notebook
View the complete exploratory analysis and machine learning workflow.Executive Summary
Stakeholder-facing report summarizing findings and recommendations.PACE Strategy
Project planning framework and analytical approach.Selected Work
Return to the complete portfolio project index.