Viewers Comment Model

📌 Model Description

The Viewers Comment Model is a two-stage sentiment classifier designed to predict the overall sentiment of a YouTube video based on its top-liked viewer comments.

Architecture Overview:

1. Stage 1 – Comment-level Sentiment Classifier:

   • Uses TF-IDF vectorization on comment text.
   • Trained using XGBoost on hand-labeled comments (multi-class: 0–4).
   • Balanced using SMOTE and RandomUnderSampler in an imbalanced-learn pipeline.

2. Stage 2 – Aggregator Classifier:

   • Extracts the top 30 most-liked comments per video.
   • Predicts each comment’s sentiment from Stage 1.
   • Counts the number of each sentiment class (0–4).
   • Uses a Random Forest to predict overall sentiment based on this count vector.

---

⚙️ How This Model Was Trained

Stage 1: XGBoost on Hand-Labeled Comments

• Input data:
  allcomments_labled.csv 8000 of the top 50 comments on different movie clips on youtube, each hand labelled with 0 to 4.

  • Columns: text (comment), sentiment (label from 0–4)

• Preprocessing:

  • TF-IDF vectorization with trigrams, min_df=5, max_df=0.7, 30,000 features
  • Class balancing with:
    • SMOTE for minority oversampling
    • RandomUnderSampler for majority class reduction

• Model:
  XGBClassifier with scale_pos_weight set from computed class weights

• Output:
  Multi-class sentiment prediction for each comment:

  • 0: Neutral
  • 1: Pleased
  • 2: Funny
  • 3: Fear
  • 4: Sad

Stage 2: Random Forest on Aggregated Sentiment Counts

• Input data:
  trainingimproved.csv contains 100 data points and is generated from stage 1 model. It correlates a movie sentiment with an array of numbers that show how many comments of a certain emotion label exists within the top 30 most liked comments.

  • Columns: Count_0 to Count_4 (number of each sentiment type from 30 comments)
  • Actual Sentiment: final label for each video

• Model:
  RandomForestClassifier with class_weight='balanced'

• Output:
  Final predicted sentiment of the video

• Evaluation:
  Performance measured manually using 120 different movie clips, out of which 87% were predicted accurately. (https://docs.google.com/spreadsheets/d/17vP7-mdsTEPYhxXXGlTWZKuioEFOPvSmzO1iR6PwhoA/edit?usp=sharing).

🎥 Real-Time Prediction Flow

1. Use YouTube API to fetch top comments (via commentThreads).
2. Filter top 30 comments by like count.
3. Predict each comment's sentiment using the Stage 1 XGBoost model.
4. Count the frequency of each class (0–4).
5. Feed the counts to the Stage 2 Random Forest model.
6. Interpret and return the final sentiment (e.g., "funny", "fear", "pleased").

---

🔧 Sentiment Mapping

Class	Meaning
0	Neutral
1	Pleased
2	Funny
3	Fear
4	Sad

---

📦 Example Output

🎬 Predicted sentiment for video 'Inside Out – Official Trailer': Funny