TikTok Nutrition Misinformation Classifier

Overview

TikTok has become one of the most influential sources of nutrition information for young people — and one of the least reliable. Claims spread faster than fact-checks, and the line between well-intentioned simplification and genuinely harmful misinformation is rarely obvious. This project builds an automated classifier to draw that line at scale.

Led as project manager and technical lead across a four-person MSc practicum team, this is a three-stage deep learning system that progresses from a TF-IDF baseline through fine-tuned BERT to a retrieval-augmented transformer grounded in authoritative health guidelines. Dataset: ~39,000 TikTok video transcripts.

Problem Statement

Manual fact-checking of short-form health content is not scalable. Existing content moderation systems flag explicit misinformation but miss the far more common pattern of technically-true-but-misleading claims, dangerous omissions, or pseudoscientific framing presented as evidence-based advice. A classifier that can detect not just factual errors but guideline inconsistency was needed.

My Role

Project manager and technical lead. Responsible for overall architecture design, the RAG grounding layer, knowledge base construction from WHO/USDA/NHS/AHA guidelines, model integration, and team coordination across four members with distinct workstream ownership.

Key Features

Progressive Architecture: Three-model architecture allowing comparison at each stage of complexity
Baseline Floor Validation: Stage 1 — TF-IDF baseline: classical ML classifier establishing performance floor
NLP Alignment: Stage 2 — Fine-tuned BERT: transformer model fine-tuned on labelled nutrition misinformation dataset
Contextual Grounding: Stage 3 — RAG transformer: retrieval-augmented model grounding predictions against WHO/USDA/NHS/AHA guidelines
Dual Classification: Dual classification heads: accuracy head (is this factually correct?) and consistency head (does this align with authoritative guidelines?)
Interpretability: Interpretable outputs: which claims in the transcript triggered flags, and which guideline they conflict with
Robust Taxonomy: Taxonomy of misinformation types developed during EDA: false claims, dangerous omissions, pseudoscience framing, unsupported causation

Technical Architecture

Data pipeline: transcript ingestion, cleaning, tokenisation, and label schema application across 39,000 records
Baseline: TF-IDF vectorisation with logistic regression and SVM classifiers
Fine-tuned BERT: HuggingFace Transformers, PyTorch, trained on labelled misinformation subset
RAG layer: knowledge base from authoritative health guidelines, chunked and embedded for retrieval
Dual heads: separate classification heads on the transformer output for accuracy and consistency scoring
Evaluation framework: F1, precision, recall across misinformation categories, with ablation study across the three stages

Challenges and Learnings

Nutrition misinformation resists simple binary classification. A claim like "carbs make you fat" is not technically false in all contexts, but is dangerously misleading as a general statement. The dual classification head — separating factual accuracy from guideline consistency — was the architectural response to this nuance, and designing the labelling schema to capture this distinction was the hardest part of the project.

Building the RAG grounding layer to work reliably across the diverse, sometimes contradictory guidance from different health bodies (WHO vs USDA, for instance) required careful document structuring and retrieval confidence thresholding.

Outcome

Ongoing MSc practicum supervised by Prof. Theo Lynn (Dean, DCU Business School). The three-stage architecture enables direct ablation comparison of classical ML, fine-tuned transformers, and RAG-grounded classification on the same task — a contribution to applied NLP research in health misinformation detection.