Exploratory data analysis and visualization to support automated content moderation.

Business Context / Overview

Background:
TikTok’s mission is to inspire creativity and bring joy. With millions of user-generated videos, the platform receives a high volume of user reports flagging content as “claims” (factual assertions) or “opinions.” Moderators face a growing backlog, slowing response times and affecting user trust.

Why this analysis was required:
To build a predictive model that automatically classifies videos, reducing manual review time and improving moderation efficiency.

Who benefits:

• TikTok’s Content Moderation Team

• Data Science and Operations Teams

• End-users through faster, more accurate content handling

Problem Statement

TikTok’s manual review process for user-reported videos is inefficient and unsustainable due to volume. Without a clear, data-driven understanding of what distinguishes a “claim” from an “opinion,” the team cannot build an accurate classification model, leading to delayed moderation, potential misinformation spread, and user dissatisfaction.

Objectives

• Perform comprehensive EDA to uncover patterns in video metrics between claims and opinions.

• Identify key distinguishing features (e.g., view count, author status, engagement metrics) that could predict video type.

• Deliver clear, accessible visualizations for both technical and non-technical stakeholders.

• Prepare clean, analysis-ready data for future machine learning modeling.

Dataset Description

Source: Internal TikTok dataset (tiktok_dataset.csv)
Records: 19,382 videos
Features: 12 columns including:

• claim_status (claim/opinion)

• video_view_count, video_like_count, video_share_count

• author_ban_status, verified_status

• video_duration_sec, video_transcription_text
  Time Period: Not specified (static snapshot)

Data Preparation & Cleaning

• Handled missing values in claim_status and engagement metrics.

• Verified and corrected data types for numerical vs. categorical fields.

• Assessed outliers using IQR and median-based thresholds.

• Standardized column names and ensured consistent formatting.

Exploratory Data Analysis (EDA)

• Distribution analysis of video duration, views, likes, shares, downloads, and comments.

• Comparison of claims vs. opinions across verified status and ban status.

• Outlier detection for engagement metrics using non-parametric methods.

• Correlation exploration between view count, likes, and claim status.

Analytical Approach

• Statistical summaries using .describe() and .groupby() methods.

• Visual distribution checks via boxplots and histograms.

• Segmentation analysis by author ban status and verification status.

• Threshold-based outlier identification using median + 1.5 * IQR.

Key Insights

• Claim videos receive significantly higher median views (~501k) compared to opinion videos (~4,953).

• Verified users are more likely to post opinions, while unverified users post more claims.

• Authors under review or banned post more claim videos and receive disproportionately high view counts.

• Engagement metrics (likes, shares, downloads) are highly right-skewed, with a small percentage of videos driving most engagement.

Visualizations & Reporting

• Python (Matplotlib/Seaborn):

  • Boxplots & histograms for all key metrics

  • Claim vs. opinion bar charts by author status

  • Scatter plots of views vs. likes colored by claim status

Video Type Distribution:

What percentage of videos are claims vs. opinions?

Engagement Metrics Analysis:

View Count Distribution by Video Type

Author Status Breakdown:

Correlation And Relationship Analysis :

Correlation Heatmap

Views vs Likes Scatter with Trendline:

Outlier Analysis:

Outlier Detection Across Metrics

Summary Dashboard:

Executive Summary Visualization

Accessibility-Friendly Version:

For stakeholders with visual impairments

Tools & Tech Stack

• Language: Python

• Libraries: Pandas, NumPy, Matplotlib, Seaborn

• Environment: Jupyter Notebook

• Visualization Tool: Tableau Public

• Data Storage: CSV

Challenges & Limitations

• Highly skewed engagement data required non-standard outlier treatment.

• Missing values in key columns reduced usable sample size.

• Text data (video_transcription_text) was not analyzed in this phase.

• Domain knowledge required to interpret “ban status” and “verified status” accurately.

Results & Business Impact

• Clear feature understanding now guides model feature selection.

• Stakeholder alignment achieved through accessible Tableau dashboards.

• Moderation team can now prioritize videos based on risk indicators (e.g., high-view claims from unverified authors).

• Data pipeline readiness for building a classification model.

Recommendations

• Build a binary classification model using view_count, author_ban_status, and verified_status as top features.

• Implement automated outlier flagging for viral content to prevent model bias.

• Expand analysis to text data (NLP) for transcription-based classification.

• Create real-time dashboards for ongoing monitoring of claim vs. opinion trends.

Key Learnings

• Business context is critical—understanding TikTok’s moderation needs shaped the analysis direction.

• Right-skewed social media data requires tailored statistical approaches.

• Visualization accessibility ensures insights are actionable for all stakeholders.

• Clean, documented EDA accelerates downstream modeling and decision-making.

Project Status

Status: ✅ Completed